arXiv Daily Index

Vision-language models may rewrite anomalous text in images into linguistically plausible expressions, compromising OCR transcription faithfulness. Sequence-level task rewards and local teacher guidance are complementary, but guidance from the same teacher may not remain equally effective as the student improves. Offline analysis shows that supervision from a fixed teacher becomes progressively less favorable as the student improves, both across training checkpoints and across response groups with different task rewards. Motivated by this observation, we introduce GAD-RL, which adaptively regulates teacher supervision during joint post-training according to the student's current task performance and local distributions. A frozen teacher conditions on reference transcriptions and student-generated prefixes. GAD-RL disables distillation for response groups containing an output with task reward at least 0.95 and continuously attenuates distillation strength as group-mean reward increases. It also weights forward KL by the student's probability of the teacher's Top-1 token, moderating local auxiliary updates when student support for that candidate is low. On Qwen3.5-2B, GAD-RL achieves 59.92% Micro Recall on CHAOS-Bench, surpassing GRPO and GRPO+OPD (fixed-weight) by 8.45 and 4.43 percentage points, respectively, while achieving an Overall score of 91.18 on OmniDocBench v1.6.

We present AREX-2, an effort to advance the self-improving capability of LLM agents, which we define as the ability to iteratively refine a solution at test time. This ability rests on two complementary capabilities: reflection, which produces a solution better than the current one, and long-horizon execution, which keeps the iteration effective over many rounds. We hypothesize that both capabilities are domain-agnostic, and can therefore be learned in scenarios that are well suited for supervision. Accordingly, we synthesize long-horizon improvement trajectories from machine learning and algorithmic programming tasks, two domains that offer verifiable feedback and reward sustained iteration. Trained on this data, our agent, built on Qwen3.8-27B, achieves strong results on MLE-bench Lite (81.8) and Frontier-CS (70.7), transfers to deep research with 84.0 on BrowseComp, 52.6 on HLE, 92.2 on GAIA, and 93.8 on DeepSearchQA, and keeps improving as its budget of rounds grows. These results show that long-horizon reflective data is an effective route toward self-improving agents.

We study the generation of agentic workflows that jointly optimize multiple objectives, such as accuracy, cost, latency, robustness, and consistency. Existing methods for workflow generation typically optimize accuracy alone or a weighted sum of objectives, so each trained generator commits to one fixed trade-off and must be retrained from scratch when preferences change. To alleviate this, we propose MoFlow, which generates workflows optimized across varied preferences. Specifically, MoFlow formulates workflow generation as a multi-objective Markov decision process and solves it by leveraging Convex-Hull Monte Carlo Tree Search with optimistic set-valued backups, where every node stores a set of reachable trade-offs rather than one weighted score. A single search thus approximately covers the Pareto front, from which MoFlow can return a workflow for any preference by lookup without retraining. We evaluate MoFlow against six strong baselines on six benchmarks spanning mathematics, code, and question answering. Since the baselines are single-scalar optimizers by design, an apples-to-apples comparison is difficult. We instead adopt an evaluation setup that favors the baselines, in that they are rerun for each testing preference, which MoFlow never sees. Even under this stringent setup, MoFlow achieves the highest average hypervolume.

Large language models (LLMs) can influence people's beliefs, yet little is known about whether and how they can manipulate each other. To investigate this, we simulate conversations between two agents: a target LLM that role-plays a human persona based on demographic and psychological attributes, and an influencer LLM that aims to make the target's beliefs more extreme. We examine radicalization along two pathways: resonance, where the influencer reinforces a target's pre-existing belief, and persuasion, where the influencer promotes a belief the target initially considers unimportant. Across affective and behavioral metrics, we find that both mechanisms radicalize the target. However, resonance produces consistently stronger effects than persuasion. Different influence tactics, such as using sycophancy and unverified claims, produce different levels of radicalization, but not consistently across metrics. We further show that resonance propagates to related beliefs, suggesting interconnected belief structures within AI agents. These findings indicate that AI agents are susceptible to radicalization, particularly when messages align with their existing beliefs, raising concerns about the vulnerability of personalized AI agents and multi-agent AI ecosystems.

When a materials LLM answers a question about crystal structure, does it reason from the structure or copy an answer already printed in its input? Accuracy cannot tell: a structural description often prints the very field it is scored against. CARAT holds question and gold answer fixed across eight matched views, names each structural relation separately in GraphSpace, and adds matched fine-tuning, answer masking, evidence injection, paired inference, and a rule that can withhold claims. First, on the benchmark's hardest families the grounded view is worth 17.3 points over formula inputs. Second, we turn that scrutiny on ourselves. GraphSpace beats a plain periodic graph by 19.3 points, but that margin is two effects at once: where the plain rendering carries everything the question needs it is 1.96 points, and where it omits those fields entirely, 46.7 points. The headline mostly measures what the baseline lacked, not how evidence is presented. Third, we attack our own benchmark. A rule that skips the link and reads the list directly answers four of seven hardened families, so we rebuilt it until eleven such shortcuts sat near chance. The frozen model quotes that link yet answers the same when we redirect it, on 95.6% of paired cases: it repeats the relation without using it. After matched supervision it reaches 99.8%, and deleting the link drops it to 23.4%, below the 27.0% the best shortcut reaches: both steps are learnable.

When a student is stuck, a tutor faces the assistance dilemma: help given too early can hinder productive struggle, while help withheld too long leaves the student in a frustrating, persistent impasse (i.e., wheel spinning). Generative AI tutors increasingly use guardrails restricting answer-giving, yet little is known about how such tutors behave once an impasse persists. We analyze 20,462 student turns from 1,260 authentic sessions with a guided LLM chemistry tutor, identifying 6,630 impasse turns of three major types: conceptual errors, expressed uncertainty, or help-seeking. We then used these impasses to simulate three tutoring conditions to study variation in AI tutor guidance through impasses: baseline, no-direct-answer, and guided tutor. For a sample of 150 impasses, prompt specificity changed pedagogy: a baseline tutor provided the answer directly in 50.7% of responses, a no-direct-answer tutor asked a follow-up question every time, and the guided tutor responded in a wide variety of ways depending on the context. We then analyzed impasse trajectories in authentic interactions, finding that each additional impasse turn lowered the odds of next-turn recovery by 12.7% (AOR = 0.873, p < .001), and early dropouts were caught in recursive concept elicitation before reaching execution. The benefit of questioning decayed as impasses persisted (scripted question x depth AOR = 0.78; follow-up x depth AOR = 0.83), whereas addressing the student's error grew more beneficial (AOR = 1.14); after a failed scripted question, repeating it was followed by recovery in 28.1% of cases, compared with 39.8% when the tutor addressed the error instead. For learning analytics, these findings identify impasse depth and type as observable, turn-level dialogue signals that analytics can use to trigger graduated, state-sensitive assistance in real time.

We study generalized Blaschke curves as a controlled environment for AI-assisted mathematical rediscovery. For one fixed degree-four Blaschke product, an agent receives numerical coordinates of the six pair-lines determined by each of 80 boundary configurations. The target theorem is withheld from the task instructions. The saved research log reports rejected geometric hypotheses and a homogeneous cubic fitted to polygon sides. Its frozen coefficients predict 480 lines from 80 unseen parameter values, with a recorded RMS scale-free residual of $8.88\times10^{-17}$. Discovery-set diagonals provide an out-of-fit consistency check, not a fully held-out test. A separate one-configuration run reports insufficient evidence for invariance. A post-review deterministic degree-search baseline also recovers the cubic, so the experiment does not establish an advantage over polynomial fitting. We present this single-instance case study as a protocol for separating conjecture, numerical validation, and proof, with explicit limitations concerning agent metadata, prior knowledge, and reproducibility.

Supervised fine-tuning of discrete diffusion language models masks some response tokens and trains the model to recover their original values from the visible context. The masking pattern therefore determines both the context available to the model and the tokens it learns to predict. Uniform random masking does not explicitly account for the interaction between these choices. We introduce GoldiMask, which selects tokens to reveal as context by approximately maximizing a submodular objective. This objective uses model signals to balance the benefit of revealing tokens against their value as prediction targets. GoldiMask then weights the remaining targets according to how they benefit from the selected context and their remaining learning potential. Across three backbones and three training datasets, GoldiMask achieves the highest average accuracy in most evaluated settings, demonstrating gains on both reasoning and code generation. Component ablations show that both context selection and target weighting contribute to the gains. GoldiMask also reduces decoding iterations on GSM8K and MATH-500 under confidence-threshold parallel decoding, while maintaining comparable accuracy at higher confidence thresholds.

Concurrent autoregressive inference creates a fundamental interference problem: prefilling a newly arrived long prompt can delay tokens for requests that are already decoding. Fixed prefill chunks reduce this interference, but the best chunk size depends on the model, hardware, load, and latency objective. We introduce Decode-Latency Feedback Prefill (DLFP), a model-free controller that changes only prefill work that overlaps active decodes. After a guarded scheduling cycle, DLFP uses the observed interval as proportional feedback to resize the next prefill chunk; isolated prefills remain unrestricted. We implement DLFP in vLLM and evaluate it with open-loop Poisson arrivals, exact token accounting, raw request traces, and NVIDIA telemetry. On Qwen3-0.6B in BF16 on one A100 80 GB GPU, three paired 100-request trials reduce P99 inter-token latency by 24.8%, 30.1%, and 28.2% (mean 27.7%, paired 95% confidence interval 21.0% to 34.3%) with exact output agreement, no failures, and unchanged SLO compliance. The benefit is not free: mean P99 time to first token increases 34.8% while remaining inside the declared SLO. Crucially, the mechanism does not generalize to Qwen3-8B, Qwen3-32B, or a two-GPU tensor-parallel configuration. We trace the failure to an asynchronous scheduler-call interval that is only a proxy for completed GPU iteration time. This negative result defines the boundary of the contribution and motivates a completion-timed controller for concurrent CPU and on-device inference. We do not claim mobile-device performance; the present work is a reproducible proof-of-concept and generalization study.

Personas used to seed LLM social simulations face a cold-start problem: existing methods lack a principled basis for deciding which attributes to include and how to assign their values. As a result, synthetic populations may misrepresent the demographic composition, latent attributes, and dependency structure that shape downstream behavior. We introduce MetaPersona-DB, a dataset of 11,000+ empirical human-subjects studies annotated with task-relevant variables, reported relationships, and aggregate-level population statistics. Building on this resource, we propose MetaPersona, a framework that retrieves task-relevant evidence, constructs literature-derived persona dependency graphs, and samples synthetic populations from empirical priors linking demographics, latent attributes, and outcomes. Across three downstream case studies, three baselines, and three frontier models, results vary by task and model: MetaPersona performs strongly on misinformation belief and AI-tool sentiment, while results on income redistribution are mixed. It also reduces persona-construction cost to under $0.5 per task using GPT-5.2. Finally, we present MetaPersona-Studio, a prototype interactive interface for empirically grounded persona generation.

Virtual clients offer a cost-effective approach to support applications such as A/B testing, recommender system development, and interface evaluation. However, building them requires access to large-scale, semantically faithful, fine-grained online user trajectories. These data are difficult to obtain because proprietary logs are subject to privacy restrictions and small businesses often lack sufficient traffic. Consequently, existing public datasets either abstract away fine-grained user interaction details or preserve rich context but remain platform-specific and small-scale. To address this gap, we propose SimTrace, a framework that generates faithful, fine-grained synthetic multimodal clickstreams through a computer-use client agent that is grounded in real user trajectories and the given web environment. SimTrace anonymizes real interactions and constructs a simulated twin of the given web environment, then uses both to generate synthetic interaction trajectories. Each action is paired with its corresponding web observations and user context, yielding a shareable alternative to confidential logs for developing computer-use agent-style virtual clients. We apply SimTrace to an e-commerce setting and evaluate both its fidelity and downstream utility. SimTrace outperforms competing baselines on 7 out of 8 fidelity metrics. Models trained on synthetic data achieve performance comparable to those trained on real data on downstream tasks such as purchase prediction and recommendation. For next action prediction task, augmenting real data with synthetic data further improves accuracy by 11.0% relative to training on real data alone. We release SimTrace as an open-source package to facilitate research on online user behavior modeling.

Recent progress in large language model reasoning has been driven by benchmarks and reinforcement learning environments with automatically verifiable rewards, particularly in mathematics, code, and formal logic. These settings make model accuracy easier to evaluate and optimize, but it remains unclear how far success under fixed problem specifications and stable evaluation criteria transfers to reasoning outside such domains. Real-world reasoning often proceeds under incomplete and revisable information: conclusions may be supported provisionally, defeated by counter-evidence, reinstated by further arguments, or revised when stronger reasons become available. Reasoning of this kind is generally referred to as defeasible reasoning. We introduce ArgGYM, a procedural benchmark and RLVR-compatible training environment for structured defeasible reasoning. ArgGYM decomposes this reasoning into twelve tasks and grounds task-specific scoring in a symbolic argumentation engine that computes the formal states used to evaluate model outputs. It includes a frozen benchmark of 1,440 verified instances across fifteen curriculum configurations, two argument preference orderings (weakest-link and last-link), and two set orderings (elitist and democratic), while the same generators and verifiers can produce fresh instances for evaluation that reduces dependence on static test sets and for verifiable-reward training. On the frozen benchmark, frontier and open-weight models show sharply different reasoning profiles: they can recover substantial parts of structured answers without solving the complete task, and performance declines in later curriculum configurations with longer dependencies and more interacting structures. We release the benchmark, generators, and verifiers for reproducible evaluation and RLVR training.

Leading AI developers have reported agents acting beyond their approved limits, which a United Nations panel described as an early warning of loss of human control. Yet incident reports and agent-safety evaluations describe these events differently, making it difficult to compare failures, trace risk across attempts, or separate agent behavior from the environment's role in allowing an out-of-scope action to succeed. To address this gap, we introduce a common framework in which each attempt ends in approved completion, safe stopping, scope escape, or continuation. We formalize the framework as a discrete-time competing-hazards model and derive escape probability within a retry budget, a model-conditional safe-budget limit, and conditions for estimation from execution logs. We audit 22 incident reports and 102 agent-safety evaluations published from January 2025 to September 2026 using primary sources. Six incidents involved tasks that could not be completed within scope, thirteen involved agents that continued rather than stopped, and five did not report stopping behavior. Developers' figures imply a task-level incidence ratio near 47 for out-of-scope coordination in never-solved versus solved tasks. Among evaluations, 87 recorded an out-of-scope effect or specification violation, 26 treated safe stopping as a first-class outcome, only 20 recorded both, and 79 merged budget exhaustion with failure. In 20 of 22 incidents, the environment allowed an out-of-scope effect, indicating that realized loss of control often reflected persistent agent behavior interacting with permissive boundary conditions; meanwhile, no evaluation reported all fields needed to estimate the full competing-hazards process from published evidence.

Frontier LLMs are increasingly used to automate scientific research through iterative search. We distinguish idea-driven search from solution-driven search and identify three core challenges: organizing evolving research ideas, selecting promising directions, and maintaining alignment between ideas and their implementations. To address these challenges, we introduce the Agentic Idea Manager (AIM), a fully autonomous framework for managing and exploring research directions in idea-driven automated research. Inspired by Bayesian optimization, AIM uses an Agentic Surrogate and an Agentic Acquisition mechanism to organize discovered ideas and guide their selection. A Solution Auditor maintains idea-solution integrity, while a Resource Planner adaptively allocates the remaining experimental budget across parallel search branches. Experiments on 10 AutoLab benchmark tasks show that AIM surpasses the strongest baseline by 1.6 percentage points on System Optimization tasks and 4.9 percentage points on long-horizon Model Development & CUDA tasks. Notably, AIM reaches the best baseline performance up to 3.1x faster in wall-clock time. We further provide a theoretical analysis of when searching over ideas becomes beneficial. Our analysis shows that explicit idea-level allocation makes semantic coverage directly controllable, and that broader coverage becomes increasingly valuable when competitive research directions are sparse among many plausible alternatives. Project Page: https://imhgchoi.github.io/agentic-idea-manager/

Open-ended LLM homogeneity can create false plurality when several systems appear to offer independent perspectives while returning the same familiar default. Single-pass answers obscure the distinction between agreement produced by a tightly constrained answer space, prompt-vocabulary echo, and broader answer spaces with stable alternatives beneath the surface. We introduce CHOIR (Collective Hierarchically-Ordered Inquiry Responses), a framework that adapts free-list elicitation from cognitive anthropology to LLM ensembles. CHOIR repeatedly elicits ranked lists, clusters items into prompt-level concepts, and measures concept salience across models, prompt variants, and persona conditions. We evaluate CHOIR on Infinity-Chat 100, an external prompt bank from recent work on open-ended model homogeneity, and on a 27-question targeted diagnostic bank designed to isolate mechanism-level contrasts. On Infinity-Chat 100, CHOIR reproduces high surface agreement (93/100 prompts above chance) while separating narrow prompts from broad prompts with recoverable depth. Across targeted probes and the external prompt bank, base-model identity remains the strongest recoverable signature, and persona prompts shift surfaced concepts within base-model signatures. A source-blind ranking module prioritises rare-but-stable candidates for later inspection. CHOIR turns open-ended homogeneity into a diagnostic measurement problem by asking where models converge, why they converge, and what remains reachable under structured depth probing.

Clinical large language model (LLM) agents deployed locally can consult more capable remote models, but doing so risks exposing patient information. Privacy-conscious delegation places disclosure decisions with a local agent, yet removing explicit identifiers is insufficient: quasi-identifiers can accumulate across multi-turn consultations and repeated patient visits to enable re-identification. We introduce PrivMeSA, a privacy-aware self-evolving multi-agent system that learns to control disclosure and retains remote expertise for local reuse. A local agent manages each encounter and consults remote specialists that may request additional information. Reinforcement learning balances task accuracy against direct disclosure and registry-based re-identification risk, with privacy evaluated over the complete outbound transcript of each encounter. A local lesson memory distills completed consultations into generalized clinical guidance and retrieves relevant lessons before transmission, allowing subsequent cases to reuse expertise without another remote exchange. Memory grows without additional outcome labels or parameter updates. On an emergency-department benchmark built from MIMIC-IV-ED records, PrivMeSA improves mean task accuracy over delegation by up to 15.8 percentage points. In the same setting, PrivMeSA reduces the disclosure of personal details from 98.0% to 0.2% of cases and the share of cases in which the patient can be narrowed to ten or fewer registry patients from 74% to 0%.

Language agents must revise planned actions when evidence changes, permission is revoked, or a stop instruction arrives. A useful response is selective: suspend affected actions, preserve unaffected work, and resume only after sufficient repair. We introduce NAQD-Env, a synthetic environment that evaluates these decisions against a deterministic reference policy over explicit evidence, authorization, and constraint dependencies. Eleven dependency families support evaluation on development structures, held-out families, and held-out combinations of structures. Metrics distinguish attempted violations from violations permitted by a simulated execution gate and jointly report policy agreement, task value, withdrawal, resumption, and event reporting. We evaluate three open-weight instruction-tuned models from two families under three prompt conditions on 350 frozen scenarios, yielding 3,150 model-prompt episodes before gate replay. Across the reported conditions, withdrawal recall is at most 0.06, no valid resumption is observed at eligible opportunities, and only one episode matches the complete reference policy. Under the NAQD prompt, Qwen2.5-7B has fewer unsafe-attempt episodes than Qwen2.5-3B and Llama-3.1-8B, but also completes less useful work and preserves unaffected actions less accurately. Exploratory supervised fine-tuning probes increase Qwen2.5-3B decision accuracy from 0.45-0.54 to 0.83-0.92; separate diagnostics reveal inappropriate withdrawal after curriculum omissions and a loss of event reporting. These results motivate evaluating selective withdrawal as a distinct component of agent reliability. The setting measures policy application with trusted structured inputs and does not establish real-world containment or source-verification ability.

Voice agents in production must handle several requests, background speech, and customers who lose patience. We introduce VAmoS Energy, a benchmark that combines these challenges in 100 calls about utility billing and payment assistance. Each caller makes two to four requests. The agent has sixteen tools backed by a stateful Stripe billing twin and the Apache Fineract loan engine, with account access blocked until caller verification succeeds. The tasks use public household electricity data and a policy based on Pennsylvania's residential billing rules. An LLM-as-a-verifier checks the agent's actions and spoken figures against explicit requirements. On a calibration run, it agrees with a code verifier on 99.1% of checks. Across fourteen voice stacks and three repeats per task, completion ranges from 17.3% to 44.7%. Grok Voice leads, and Gemini 3.8 Live and GPT-Live follow at about the same cost per call. Background television reduces pooled completion from 38.7% to 8.6%. The simulated caller often accepts an incorrect result because it hears the agent's words but cannot inspect its actions. These findings show why voice agents need evaluation across the whole call, including what they say, what they change, and how they handle competing speech.

Large language models (LLMs) are increasingly used in culturally sensitive settings, where alignment requires representing diverse preferences within populations. Yet existing methods model populations at coarse demographic or community levels and overlook within-group variation. We introduce Demographic Pluralism, an inference-time framework that estimates population-level opinion distributions without opinion-distribution training data or task-specific fine-tuning by generating multiple perspectives within demographically grounded groups. Across four backbones on GlobalOpinionQA and VITAL, it reduces Jensen-Shannon distance by 8.4%-26.4% over Modular Pluralism. Among weighted, equal-weighted, and inverse-weighted aggregation, equal weighting performs best overall; group-level error also increases with group weight, helping explain weighted aggregation's weaker performance.

This paper examines human-AI interactions (HAIIs) in clinical trials and presents a multidimensional categorisation framework that classifies interactions according to AI tasks, human-AI relationships, interaction configurations and interacting human groups. We define HAII, examine existing taxonomies and extend existing categorisation approaches through this novel multidimensional framework. We purposively sampled 15 clinical trials from a previously reported dataset. Each trial was independently categorised by two human reviewers and six large language model (LLM) classifiers. The proposed categorisation provides a structured method for the consistent identification, comparison and synthesis of human-AI interactions across clinical-trial records. The framework is intended to support more consistent comparison and synthesis of AI-related clinical trials and to make explicit the different forms of human involvement associated with AI interventions. The results demonstrate the potential for LLM-assisted categorisation while indicating the continuing importance of human judgement where trial records are incomplete or ambiguous. The principal contribution is a proposed multidimensional framework that brings together AI tasks, human-AI relationships, interaction configurations and interacting human groups within a single approach designed for clinical-trial records. Its significance lies in its potential to support more systematic identification, comparison and synthesis of how humans and AI interact in clinical trials.

Large language models are frequently proposed as a route to AI-powered services for African communities, but they are least reliable exactly where the need is greatest: all African languages remain low-resource by any standard measure, and models trained on scraped, standardised text systematically misrepresent the dialectal and regional variation of how people actually speak. We introduce \textbf{Model as a Library (MaaL)}, a software architecture that packages small, community-enrolled speech models as versioned on-device dependencies, enabling offline structured data collection that cannot generatively hallucinate, for populations that current language models serve worst. Rather than relying on web-scraped corpora, MaaL's vocabulary is enrolled directly from a small number of example recordings by the speakers themselves, at the point of deployment. We describe the architecture and its central mechanism - keyword spotting that turns a closed-vocabulary text form into a voice form, filled and submitted entirely on-device - and propose transpiling the closed-vocabulary elements already present in widely-deployed digital form tools into MaaL schemas, a low-friction path to voice-first, offline data collection for the low-literacy populations these tools already reach. This is a position and system-design paper: we describe the concept, the mechanism, and an analytical feasibility case, and identify what a working implementation still requires.

As large language models (LLMs) are increasingly used in clinical settings, it is critical to evaluate their reliability under realistic variation in clinical text. We study this question in clinical triage, comparing LLMs to practicing physicians under text perturbations that preserve the underlying clinical setting. We introduce a benchmark of over 6,000 clinical scenarios, 7,000 physician annotations, and 225,000 model responses. Using this benchmark, we make two key observations. First, LLMs are more likely than physicians to recommend unnecessary care at baseline, and this tendency increases under perturbed inputs. Further, we find that LLM recommendations are more sensitive to gender and tone perturbations than human recommendations. Together, these results demonstrate that LLMs can vary under clinically irrelevant textual changes, highlighting the need for deployment-oriented evaluations grounded in expert physician behavior.

Two scientific findings can disagree without contradicting each other. Determining whether they conflict requires knowing whether they describe comparable measurements. We study how language models behave at this decision point. In a controlled task, we generate an unsatisfiable XOR constraint system and translate its constraints into scientific reports from different laboratories. One assignment satisfies more constraints, while another satisfies fewer but better matches expected biology. This creates a simple dilemma: does the model choose the assignment that best fits the constraints, or the one that better matches biological expectations? When the constraints are stated directly, GPT-5.6 Sol and Claude Opus 5 recover the best-supported assignment in 90% and 96% of cases, respectively. In scientific prose, however, the models behave differently. Claude Opus 5 often prefers the biologically expected assignment. Removing that biological preference increases recovery of the better-supported assignment from 27% to 79% (p<.001); recovery reaches 92% when the same Biology-favored record is accompanied by a formalization request and an explicit paired-design cue (p<.001). GPT-5.6 Sol is less sensitive, with neither corresponding change reaching statistical significance. These results suggest that reliable scientific verification depends not only on formal reasoning, but also on how models decide which findings should be compared and what relations they imply.

LLM-guided evolutionary search can discover complex programs, but existing methods mostly only save candidate programs and fitness scores while discarding which component edits produced which fitness metric changes. Existing methods force the mutator LLM to infer the effect of prior edits from cluttered histories, making program search slow and unstable. This is especially true for locally servable LLMs to evolve multi-component systems. We introduce component-aware feedback, which compares each evaluated program with its parent, identifies the components that changed, and logs them with the associated metric differences into an attribution memory that later mutations read. The memory keeps each change in two reference frames, local against the parent it came from and global against the seed program, which shows both the immediate effect of a change and the cumulative progress made since the seed. We study this on LLM reranking, a multi-objective optimization problem where a multi-stage pipeline must balance quality against serving cost. Across twelve \textsc{Bright} datasets, our method reaches the strongest baseline's final quality after a median of one third of the search budget and ends 7.2\% higher in held-out nDCG@10, and under a cost-aware objective it finds pipelines that are on average more accurate while using 11\% fewer tokens per query, showing component-aware feedback to be a promising direction for more efficient self-evolving systems.

Agentic AI systems increasingly rely on cloud-hosted large language models for planning, tool use, and iterative execution, raising concerns about API cost and data exposure. Advances in personal AI devices enable agents to execute locally, but limited resources on device may affect task success and performance. Existing benchmarks are inadequate for systematically characterizing these trade-offs across devices, workloads, and deployment architectures. We present AgBench, a benchmark suite and open artifacts for reproducible evaluation of agentic AI on personal devices. Using AgBench, we evaluate local, hybrid, and cloud execution across agentic workloads, examining task success, latency, cloud API cost, and data exposure. Our results, drawn from over 162.07 million data points, show that personal AI devices can complete many agent tasks locally, but local-only execution generally has lower task success and longer completion times than cloud-only execution, especially as concurrency increases. Local-only execution eliminates cloud model API costs and sensitive-information exposure to cloud agents. Hybrid execution can improve task success, but its cloud cost and data exposure depend on how agents divide work and share information. No single architecture performs best across task success, goodput, cloud cost, and data exposure; deployment choices should reflect the intended workload and device capabilities. AgBench is available at https://anonymous.4open.science/r/AgBench-2777.

In recent years, LLM-based multi-agent systems have been widely applied to orchestrate tool-using agents into executable communication graphs. However, existing self-evolving orchestration still faces key challenges, including post-hoc evolution that revises the team only after the trajectory ends, credit diffusion that gives every action the same terminal advantage under confounded baselines, and skill admission that is uncalibrated and never retired. To address these challenges, we propose EvoSteer, a new paradigm of Online Self-Evolving Graph Orchestration -- the orchestrator builds a running team and repairs its plausible but failing steps from execution features and a learned value estimate. To support this paradigm, we introduce Anchored Trajectory Balance (AnchorTB), a regression-style flow-matching loss that assigns each orchestration action a coefficient by balancing subtrajectories against a frozen reference. Built on the learned flow, we further propose Validated Skill Admission, in which a candidate skill is tried before promotion and promoted only if paired evidence passes a sequential test under a shared nominal testing budget. Moreover, AnchorTB combines measured task-level reference reward statistics with prefix-dependent corrections. Experimental results on twelve datasets show that EvoSteer significantly outperforms baselines across question answering, mathematical reasoning, code generation, and interactive decision making. Our code is available at https://github.com/beita6969/evosteer.

Final-answer accuracy does not reveal whether a scientific-search agent failed to encounter a target paper, attempt to inspect it, or return an accepted answer after inspection. We introduce decision checkpoints that record observations and tool actions without benchmark labels during inference, then join target identities and evaluator labels to assign outcome categories from recorded events. Across five conditions on 540 answerable AutoResearchBench Deep questions in a fixed, target-enriched environment, keyword search achieves 24.6\% accuracy, compared with 17.8\% for raw search. The keyword condition has fewer incorrect answers with neither target exposure nor inspection, but more incorrect answers after the target is exposed and left uninspected. Compared with keyword search, read-first has 27.4\% more recorded evidence-search calls. Target inspection attempts occur on 199 questions under read-first and 191 under keyword search; both conditions achieve 24.6\% accuracy. The checkpoint protocol makes these question-level differences explicit, distinguishing target exposure and inspection from aggregate accuracy and total tool use.

Knowledge-intensive language-model systems typically represent external knowledge as text chunks or static graphs, with limited support for concept evolution, point-in-time reasoning, and distinctions between validated and inferred knowledge. We introduce the Concept Lifecycle Model (CLM), which represents concepts as persistent, graph-grounded, temporally versioned entities with explicit provenance and epistemic status, and Concept-Grounded Attention (CGA), which injects concept-graph structure into transformer computation through graph-biased self-attention (Form A) and gated cross-attention over concept nodes (Form B). We evaluate the framework in controlled settings using disabled-mechanism baselines. On 200 MuSiQue and HotpotQA questions with retrieval fixed, concept-graph retrieval recovers explicit multi-hop paths but does not improve evidence recall. Form A appears to steer attention, with 2.76 times more attention on gold than distractor concepts, but the same ratio occurs when Form A is disabled; the learned bias is negligible and no answers change. An identity-preserving Form B improves F1 from 0.188 to 0.221, but control concepts yield 0.213, indicating that most of the gain reflects added capacity. On LongMemEval, explicit temporal representation improves answer accuracy by 13 to 25 points across all tested generators, up to 122B parameters, while simplified CLM version resolution performs similarly to dated serialization because concept identity is not established reliably. On a synthetic source-independence task, protocol-derived epistemic status reduces unsupported assertions from 28% to 0.1% in a fine-tuned small model and from 19-68% to 0-5% in 72-122B models. Overall, the results support making temporal validity and epistemic status explicit, while showing that graph-attention diagnostics are not informative without disabled-mechanism controls.

Spatial transcriptomics enables spatially resolved gene expression analysis from slide-level images while preserving morphological features, providing valuable information for studying disease mechanisms and developing treatments. However, spatial gene expression profiling typically requires expensive and time-consuming tests. While existing image-based prediction optimizations mostly revolve around including positional embeddings and further image-based changes, text-based optimizations remain relatively unexplored. We present GATE-ST, which incorporates text-based inputs into image-based spatial gene expression predictions. With this approach, generated text descriptions of genes are utilized to better spatial transcriptomics prediction results. Gene summaries are put through a text encoder, generating embeddings that integrate with image embeddings through cross-attention layers to align with morphological features. We demonstrate the effectiveness of such text inputs by benchmarking performance against random gene embeddings and multiple other image-text fusion architectures, and show that GATE-ST outperforms these alternatives. Our results demonstrate the effectiveness of GATE-ST in pathology imaging, which may greatly reduce the time and cost of accurate spatial transcriptomic predictions, proving the potential of text-guided spatial gene expression prediction.

A cumulative token cap can fall inside a mathematical derivation, forcing a test-time controller to choose between stopping at the cap (strict) and allowing the current attempt to finish (advisory). We measure this boundary choice with paired offline replays of 19,200 public traces: 120 AIME, BrUMO and HMMT problems and two archive configurations of one model. Candidate order and a 16-attempt cap are fixed, and answer selection is blind to reference answers and correctness labels. Three findings emerge. First, at the 4k cap, most advisory accuracy gains replace abstention with a correct answer; strict stopping pays for an unfinished prefix that the completed-only selector cannot use. Second, comparisons along realized cost differ from same-cap comparisons: advisory 4k in low has higher accuracy than strict 8k at comparable mean completion cost, while in high its observed accuracy is 0.42 points below strict 32k using 59% of its mean tokens. These aggregate comparisons do not establish equal-compute superiority or accuracy equivalence. Third, increased candidate coverage does not guarantee higher answer accuracy: a log-probability selector loses accuracy while coverage rises, including after a source-grade consistency repair. Same-cap majority-accuracy differences shrink below 1.3 percentage points at 32k. Budget curves should jointly state the cap, realized cost, eligible candidates, stopping rule and selector information.

Long-horizon agentic workflows require models to sustain repeated state-dependent actions all while the context grows, sub-task complexity changes, and new data arrives. Each situation represents an independent axis along which an agent may fail. An agent reconciling a long ledger, for example, must repeatedly read its state, update the correct record, and preserve alignment across thousands of outputs. A model may accept the entire ledger yet lose its place or stop applying the operation consistently as generation proceeds. We introduce Long-Transduction, a controlled diagnostic that tests a model's ability to stay on task during long generation while continuously reading, mutating, and outputting input-context dependent operations such as arithmetic, sorting, variable lookups, and table transformations. Long-Transduction evaluation independently varies local task complexity, input data formatting, and context length isolate failures along each axis. We evaluate seven open-weight models, finding a 62.8\% decrease when scaling context length from 4-128K, a 36.5\% decrease when varying input format, and a 39.9\% decrease by increasing local task complexity. Together, these failures represent critical liabilities in long-horizon agentic workflows.

LLM agents rely on retrieval tools to access external knowledge, yet visual agentic search remains severely bottlenecked by standard single-step retrievers. In current pipelines, the agent must issue text queries for every intermediate step, struggling when visual clues are difficult to describe or when the retriever fails to surface necessary intermediate evidence within its top results. We hypothesize that offloading multi-step navigation across the entire embedding space directly to the retrieval tool resolves this performance bottleneck. To study this systematically, we introduce VHOP, a flexible data generation framework and benchmark with five core difficulty levels testing both visual matching and search planning. Using this framework, we develop VHOP-Router, an end-to-end training pipeline---combining supervised fine-tuning, online imitation learning, and reinforcement learning---that transforms a standard embedding model into an autoregressive multi-step retriever. Operating directly in the visual latent space, VHOP-Router retrieves linked image chains in a single tool call without requiring the agent to formulate intermediate text queries. Experiments show VHOP-Router boosts retrieval performance from under 5\% to 76.3\%. In agentic search, it improves task success rates by 52.7\% and reduces the average token length by 61\% from 1886 to 728, whereas upgrading the agent yields only a 3.7\% gain. Compared to a strong baseline where the agent retrieves the top 50 results per step, VHOP-Router maintains superior performance while reducing in-context images by $23\times$ and cutting the cumulative API payload by $35\times$. The models also generalize robustly to unseen difficulty levels and realistic test sets. Ultimately, VHOP and VHOP-Router provide an efficient and effective solution for visual agentic search that leaves native LLM capabilities entirely intact.

Large language models are increasingly participating in complex real-world tasks in the form of algorithm-design agents, designing and refining algorithms. Many successful algorithm-design agents adopt pure in-context evolutionary frameworks, but they may quickly plateau in domains that require specialized knowledge. Parametric adaptation offers a way to internalize specialized knowledge, but conventional training requires abundant domain-specific corpora while high-quality algorithms are scarce in complex algorithm-design scenarios. In this paper, we propose sample-efficient parametric self-evolution where agents can explore and learn from self-generated algorithms. First, we characterize in-context evolutionary stagnation and analytically propose the Improvement Chain proposition, showing how learning successive self-generated algorithms can locally increase the likelihood of neighboring algorithms. Motivated by this local-transfer perspective, we further propose Population-Curated Policy Optimization (PCPO) to utilize a global population and a hybrid policy update scheme for retaining and reusing high-quality, diverse self-generated algorithms, shifting the policy towards stronger algorithms. In the task of learning rate schedule design for global placement in electronic design automation, trained only on 4 chip cases, PCPO outperforms the state-of-the-art in-context evolutionary methods (e.g., OpenEvolve and ShinkaEvolve) on average across 16 chip cases. With an 8B-size base model, PCPO achieves competitive performance compared to frontier closed-source models such as GPT-5.5. PCPO also reduces inference-time token cost by internalizing grounded domain knowledge and prompt distillation. Moreover, PCPO achieves significant speedups on four GPU kernel designs, with an average of 8.27$\times$ speedup against the PyTorch Eager baseline.

Scientific agents can retrieve relevant passages yet still lose functional order, mix evidence across sources, or state conclusions that exceed the retrieved record. We introduce PathAnchor, a bounded scientific reasoning system built on path-structured evidence workspaces. Instead of treating passages or extracted concepts as independent units, the system retrieves source-linked Material-Sensor-Signal-System trajectories that preserve role, direction, and the evidence supporting each transition. A controller uses three read-only tools to search paper-specific trajectories, trace paths across candidate sources, and open exact evidence before producing a claim-cited answer and an explicit evidence boundary. On 120 single- and cross-paper flexible-sensor questions, PathAnchor scores 82.6% and leads six evaluated systems. Under a matched controller, corpus, and six-call budget, replacing unordered concept graphs with path-structured records raises source recall from 61.3% to 82.9%, increases answers whose claims all cite opened evidence from 69.2% to 90.0%, and reduces tool calls. These results show that evidence organization affects retrieval and citation completeness under fixed agent resources.

An agent that remembers what it did on a web page must decide when two pages count as the same. Memories built on observation similarity merge pages that look alike but behave differently, and GUIs are full of such pages: two tabs of one widget or two rows of one menu answer the same click differently. We define the merge rule as an action-conditioned bisimulation over the empirical predictive state graph a frozen agent fills as it acts. Two states merge only when their shared actions lead to agreeing outcomes and successor blocks under an affordance label. Observation similarity never enters the rule, and nothing is trained. It replaces the merge rule of an existing outcome-value memory, so a closed-loop comparison isolates it. On MiniWoB++ it raises success rate over a memoryless agent, while a control taking identical exploratory detours, the prior successor-representation merge, and the same criterion without action conditioning change nothing.

Conversational agents increasingly interact with children, yet evidence on how their design shapes children's susceptibility to persuasion comes almost entirely from the lab. We report a $2\times2$ randomized field experiment embedded in a public German Santa Claus telephone hotline. Children's calls were randomly routed to one of four LLM voice agents varying persona (Santa, high authority, vs. Helper, low authority) and framing (persuasive nudges toward prosocial wishes vs. neutral). Of 1,072 logged calls, 89 conversations (median age 6) met inclusion criteria. Persuasive framing raised the probability of a prosocial wish from 11.6% to 45.7%, robust to controls. Persona authority showed a near-zero effect: Santa did not outperform the Helper. Persona instead shaped engagement; children hung up on the Helper far more often within the first minute (65% vs. 39%). Where context already lends an agent legitimacy, how it speaks shapes children's compliance more than who it claims to be.

Organizations started listening to employees through conversational AI agents alongside structured surveys. Little is known about what these channels change in what employees say when disclosure carries hierarchical risk. We report a field study inside a global management consulting firm whose process pairs a pre-survey with an adaptive AI voice interview on the same themes within one session. Across 44 first-session interviews (132 matched theme observations), 20-41% of sessions showed a favorable rating co-occurring with a substantive concern voiced later, depending on the favorability threshold. The Gioia analysis drew on 158 protective quotes from 65 eligible sessions. Disclosure rarely arrived unguarded: employees softened concerns, deflected accountability, and bounded how far they went, and this protective work tracked the perceived legitimacy of the listening structure. We develop a grounded model of bounded disclosure and derive four propositions for voice, channel and listening research. Silence, we argue, can persist inside expression.

Graph agents extend large language models (LLMs) with the ability to actively explore and reason over knowledge graphs through multi-step interactions with graph tools. However, training capable graph agents typically requires large collections of question-answer pairs and reasoning trajectories, whose manual construction is costly and difficult to scale. Moreover, employing proprietary LLMs to generate such supervision further risks exposing sensitive graph data to external services. Therefore, we propose GraphCert to bootstrap agentic graph reasoning with certified evidence rubrics during post-training. Specifically, the Bootstrapped Graph Quizzer guided by generation controls produces graph-grounded QA pairs and marks supporting evidence, which undergo execution certification and semantic curation. The accepted evidence is then canonicalized into certified evidence rubrics that later reward Graph Solver evidence alignment alongside answer correctness during GRPO training. Experiments on five graph reasoning domains in GRBENCH demonstrate that GraphCert consistently outperforms substantially larger LLM agents and post-training method. Furthermore, our analysis demonstrates that the learned policy transfers robustly across heterogeneous graph domains, suggesting that GraphCert acquires reusable graph-reasoning capabilities rather than domain-specific patterns. These results establish executable self-certification as an effective approach to self-training compact graph reasoning agents. Our code will be made publicly available.

Large reasoning models (LRMs) improve performance on complex tasks through extended reasoning, yet the same process can degenerate into redundant verification and persistent generation loops. Such uncontrolled reasoning increases inference cost and creates risks of resource exhaustion and service degradation. However, existing mitigations largely truncate long outputs or react to surface repetition, and thus fail to distinguish normal thinking from uncontrolled reasoning or explain how benign reasoning degenerates into harmful behavior. In this paper, we operationalize LRM generation as four states and further introduce Reasoning-state Analysis via Dynamic Attention Responses (RADAR), which identifies the current reasoning state in real time and characterizes how effective reflection can develop into uncontrolled generation. Guided by RADAR's analysis, we further realign abnormal attention distributions toward patterns observed in normal requests and examine how this correction affects excessive reflection and persistent looping. Temporal analyses show that uncontrolled reasoning is characterized by attention distributions that deviate from normal generation, with abnormal trends becoming detectable before repetition begins. Correcting these deviations through Attention Realignment consistently reduces looping while largely preserving benign performance. Together, RADAR provide a mechanistic account of how reasoning becomes uncontrolled, offering actionable guidance for identifying critical failure stages and designing targeted runtime interventions.

Organizations increasingly route employee feedback to leaders through large language model (LLM) summaries, an unaudited layer that silences already-spoken voice. We introduce a Voice Retention / Representation Ratio metric for representational bias in summarization and apply it to a bilingual (English/German) corpus of 2,586 free-text responses from a global professional service company. First, employees supply criticism more reliably than praise (withholding praise is 82 times more common). Second, across 45 leader-summaries the pipeline filters by popularity, not sentiment: criticism survives, yet a concern voiced once is dropped 86% of the time, with short and German-only content lost on the same axis (theme retention 0.14 vs 0.74; German directional). Controlling for frequency, sentiment has no independent effect; the harm is prevalence-driven, which sentiment-only audits miss. A targeted prompt recovers only named themes. We contribute the metric, field evidence, and a disaggregated voice-retention card.

Direct-decision models turn text into low-latency structured labels and scores, making them attractive for classification and automatic evaluation. Yet reliability requires more than accuracy: a model must also use the ordinal decision scale supplied by the user faithfully. We analyze JEV~1.13 and three open KEV models. Our investigation begins with ANLI, where JEV assigns 38.8\% of all predictions and 51.3\% of errors to Neutral despite 74.95\% accuracy, nearly balanced gold labels, and balanced candidate positions. Across 36 ordinal datasets, final decisions use only 67--76\% of the effective gold support, versus 87--102\% on four nominal tasks. Randomizing candidate order weakens but does not remove this compression. Holding items and source scores fixed while balancing gold support and positions, we refine scales from $K=2$ to $14$; utilization falls for every model and reaches 26--75\% at $K=14$, although candidate probabilities remain broad for most models. Targeted BA-LoRA post-training raises gold-relative utilization from roughly 47\% to 86\% on eight supervised scales at both KEV sizes, showing that the compression is learned and modifiable rather than an immutable architectural limit. We call this ordinal scale-utilization bias: decision-stage candidate-space compression distinct from accuracy, gold imbalance, fixed position, and candidate count alone. The code and data are available at https://github.com/Glax147/jev_ordinal_scale_bia

Multi-path reasoning methods such as self-consistency (SC) sample $K$ reasoning paths and choose the most frequent answer. However, their gains quickly plateau as $K$ increases, and existing methods do not predict when this saturation will occur. We formalize multi-path LLM reasoning as a diversity combining problem from wireless communications: each path is a noisy channel observation, and the pairwise correlation of path correctness caps the design-effect effective sample size of the vote at a finite ceiling. Generalized least squares (GLS) analysis shows that, under exchangeability, the optimal symmetric linear combiner of latent embeddings is uniform, supporting majority vote as the natural default in standard SC while leaving room for weighting or pruning under heterogeneous prompt-template branches. Across 5 models and 12 benchmarks, prompt-template diversity reduces path correlation in $55$ of $57$ valid cells, with the strongest effect on open-ended QA. We derive an Adaptive-K rule that uses a four-path pilot to select $K^*$, retaining $96$--$103\%$ of MV@$K{=}32$ accuracy across Math, QA, and NLU.

Decision models such as Jev answer questions with probabilities, which are only useful if they are calibrated. Open-source reproductions rely on supervised fine-tuning plus temperature scaling, while reinforcement learning from verifiable rewards (RLVR) makes reasoning models overconfident. We present a working implementation of reinforcement learning for calibrated decisions (RLCD) for reasoning models: the model samples a rationale, and we score the answer distribution it commits to afterwards with a strictly proper scoring rule. A variance identity shows that scoring the mixture of several samples rewards disagreeing rationales, and that RLVR is exactly this mixture objective without its diversity term. Optimized naively, the per-rationale objective either switches reasoning off or is drowned out by policy-gradient noise, which leads to a two-stage recipe: calibrate, then reinforce. With Qwen3-1.7B on two reasoning tasks (3 seeds, paired tests), RLCD matches or beats SFT, RFT/STaR and GRPO (each temperature-scaled) in accuracy and beats all of them in selective prediction; on GSM8K answer verification a single query decides \gvTwoCovFive\% of the items at $\le$5\% error, versus \gvGrpoCovFive\% for GRPO. When uncertainty comes from annotator disagreement, RLCD provably cannot beat cross-entropy. Code and results: https://github.com/ZimmyGao/openjev-rlcd.

As preferences, goals, and facts change, LLM agents must use the current state while earlier versions remain in context. Yet they can answer with an old value of the same variable, a failure that we call stale binding. To study when models use outdated information and why, we introduce Controlled In-Context Memory (CICM), a benchmark for tracking and using updated information in conversations and agent logs. We observe that even frontier reasoning models can fail to recover the current state. We find that in open-source models probes can still recover the updated value when the model answers with an old one, pointing to a failure to select information that remains available. Component tests in Qwen and Pythia identify a mechanism for this selection failure: attention drift, where attention favors old values over the current one when producing an answer. We study a one-layer transformer to mathematically understand how this phenomenon happens: when attention scores are similar, several old values can together receive more attention than the current value. Guided by this explanation, we redirect attention toward the current value without further training. When the current value is requested directly, adjusting this intervention for each input corrects most old-value errors across various model families while preserving nearly all initially correct answers. Reliable context management therefore requires more than remembering updated information: models must use it to guide their answers.

Large language model (LLM) computer-use agents are typically evaluated with clean written instructions, despite speech being an increasingly popular interface for interacting with such systems. Speech input introduces an additional failure point: transcription errors can alter task-critical entities, constraints, or targets before the agent begins reasoning, while conventional ASR metrics do not directly measure whether the information required for successful execution has been preserved. We introduce Talk2Agent, a benchmark for evaluating how effectively voice interfaces convey human-spoken instructions to LLM-based computer-use agents. Talk2Agent builds human-spoken versions of tasks from WildClawBench and OSWorld and evaluates a range of voice interfaces, including dedicated ASR models, audio-capable LLMs, contextual biasing, and LLM-based ontology repair. Because repeatedly executing long-horizon computer-use tasks is costly and stochastic, we further propose an execution-free, task-conditioned evaluation framework that projects the original task grader onto prompt-addressable intentions and measures how much task-relevant information is retained after the voice interface. On WildClawBench, Talk2Agent's execution-free native projection provides a practical, execution-grounded measure of voice-interface quality, correlating with downstream task completion and improving Pearson correlation by 0.246 over WER/CER on 32 hours of real human speech.

In finance, interpreting machine learning predictions is essential, yet the numerical outputs of explainable AI can be difficult for non-experts to understand. While large language models (LLMs) can translate these outputs into natural language, they may produce errors when inferring numerical changes and feature relations. We propose an LLM narrative framework for cross-sectional stock return prediction that combines temporal Shapley additive explanations (SHAP) evidence with historical regime analogs. Temporal evidence tracks changes in the normalized global SHAP importance of an XGBoost model over six months. Historical analogs are past periods with similar changes in SHAP importance, their model performance and subsequent market returns are provided as comparative context. Using this framework, we conduct a controlled study of progressive reasoning externalization, sequentially providing raw SHAP sequences, deterministic temporal descriptors, and feature relations. Each generated claim is verified against provenance-linked evidence. Across Qwen3, externalizing numerical and relational reasoning improved evidence faithfulness as well as temporal and relational accuracy. Evidence faithfulness increased from 0.696 to 0.996 for Qwen3-32B-Instruct. While historical analogs did not improve structured automatic faithfulness, they received higher human-rated usefulness scores. These results suggest that externalizing verifiable reasoning enhances narrative faithfulness and that historical context adds interpretive value.

Real-time strategy (RTS) games require agents to coordinate economic development, production and construction, base defense, unit organization, and attack timing over long matches. Existing studies have applied large language models to command decision-making in RTS games, enabling agents to read textual game states and generate high-level plans. However, long inference latency can cause them to miss critical tactical events. The complexity and tactical diversity of full RTS matches also leave existing systems heavily dependent on manually written experience-based prompts, with limited ability to learn continuously from past games. We present STRATA, a role-aligned hierarchical system with cross-game self-learning for Red Alert. STRATA assigns in-game strategic, logistical, and tactical decisions to a Strategic Agent (SA), Logistics Agent (LA), and Tactical Agent (TA), respectively. The SA generates high-level directives based on the global game state and relevant experience cards, while the LA and TA handle logistics and tactical execution. After each match, a Review Agent (RA) derives candidate experience from game traces, validates and revises it using evidence from subsequent matches, and compresses strategic experience supported across multiple games into concise experience cards for SA retrieval. We evaluate STRATA through the formation of experience cards, full-match comparisons before and after learning, and experience learning against AI opponents with different play styles. Under a fixed scenario, using the learned experience cards increases the observed win rate from 30% to 100%. Sequential learning against AI opponents with different play styles also produces distinct long-term strategic experience.

Long-horizon agentic tasks demand strong reasoning and efficient execution across successive interactions with dynamic environments. A common approach decouples high-level planning from low-level execution through separate planner and actor roles. To investigate coordination failures in these tasks, we prompt both agents for structured state assertions and compare their reports programmatically to detect explicit contradictions. Our analyses reveal systematic disagreement about the same task-relevant state facts, a phenomenon we term planner-actor state mismatch. We further find that providing agents with task-relevant state information reduces mismatch and improves coordination and task performance. Based on the systematic analysis of the state mismatch, we propose Consistent Plan-Act (ConPAct), which feeds detected contradictions back to both agents to form consistent state interpretations and fine-tunes them on curated consistent interactions for better coordination. ConPAct improves performance across various environments and model configurations, e.g., increasing MiniGrid success rate from 38.6% to 54.4% with GPT-5.6-sol/terra as planner and actor respectively, demonstrating that state consistency can guide both inference-time correction and coordination training.

Agent harnesses govern how large language models (LLMs) gather context, invoke tools, verify results, preserve state, and terminate, largely affecting agent performance. However, the value of each harness mechanism can differ across heterogeneous tasks: a mechanism that improves one task may impose overhead or context distraction on another, leading to the suboptimality of a global harness. We characterize this suboptimality as a mismatch induced by fixed mechanism choices, motivating task-specific harness construction. Nonetheless, generating harness code for each task introduces generation and debugging costs, with execution risks that can compound as more mechanisms are generated. To address those challenges, we introduce Harness Primitives, reusable harness mechanisms with clear application scope and composition contract mined from failed task trajectories. Based on Harness Primitives, we propose STITCH, a framework that Selects suitable primitives given Task Information and compiles them into Task-speCific Harnesses at test time. This separation enables task-specific harnesses without generating or repairing mechanism code at test time. Extensive experiments demonstrate that STITCH not only improves harness adaptability and robustness, but also scales with the primitive library size, boosting task success rates by up to 12 points over fixed harness baselines, surpassing human-designed harnesses like Codex CLI while maintaining a minimal test-time harness composition overhead of only 2.7%, 638 times more efficient than generating task-specific harnesses from scratch. Ultimately, our work demonstrates that building task-adaptive harnesses can be beneficial for completing diverse tasks and that building reusable primitives can be a promising path towards this goal.

A classifier's conditional accuracy can change while its confidence distribution stays exactly the same. We study the worst-case movement of the reliability relation under covariate shifts that preserve the distribution of the confidence score, constraining the reweighting within each confidence level by a $\chi^2$ budget; the resulting worst case, as a function of the budget, is a fragility profile. On an interval of budgets that can be computed from the source distribution, the profile equals exactly the square root of the budget times the within-level variance of the correctness propensity -- the grouping-loss term of calibration-refinement decompositions. Beyond this interval the profile is governed by the tails of the propensity law, and the entire upward profile determines the centred within-level law; consequently, calibration residual and grouping variance do not determine fragility in general, though they do when labels and predictions are deterministic. Since the propensity is not observed, we restrict reweightings to a learned finite readout within confidence bins, bound the part the restriction misses by the grouping variance remaining inside readout cells, estimate the restricted profile with role-separated labels, and provide a separate split-sample lower confidence bound. On ImageNet this bound is positive in both splits for four of six primary classifiers and nine of twelve additional ones as released, and for three of eighteen after temperature scaling. Held-out drift under optimised reweightings fitted without evaluation labels tracks the estimated profile; an exploratory label-permutation diagnostic yields near-zero agreement for this statistic while largely reproducing the correlation observed for unsigned random reweightings.

Routing signals of modern vision transformers -- expert gates, attention-residual weights and halting scores -- often improve probes that predict whether the model is correct, and the improvement is commonly read as evidence that routing carries information about errors beyond the model's outputs. We test this inference directly: keeping real output-routing pairs, we redraw correctness labels from a frozen output-only generator fitted on disjoint data, so that routing is uninformative by construction. Under this exact label null, a width-matched MLP comparison still reports a routing gain in 51.3% of confidence-only evaluations (308/600), while a linear comparison reports none. Holding each training trajectory fixed on a six-model panel and selecting the checkpoint by validation log loss instead of validation accuracy removes the detections (50/120 to 0/120, and 83/120 to 0/120 in an independently implemented probe), identifying accuracy-based checkpoint selection as the cause; across all output views the raw detection rate falls from 27.5% (528/1,920) to zero observed detections. The repaired comparison is not sensitive, detecting an implanted signal of about 0.005 nats in 0/20 replicates in each of two matched settings, whereas a conditional permutation test built on an estimated routing law detects it in 11/20 and 10/20 and rejects rarely under the null. On real correctness labels, the conditional analysis yields model-relative evidence in five DeiT attention-residual families; in four it persists under two specified variants of the conditional law, and no family passes an additional noise criterion. Fitting a better probe and testing for incremental information are different problems, and each needs its own validation.

Recent work on hallucination detection in large language models has shown that, for a fixed pre-trained model and reasoning task, it is possible to estimate the model's confidence in the correctness of its outputs. Such uncertainty estimates have primarily been used to improve truthfulness by detecting or filtering confabulations. In this work, we ask whether these signals can instead be used more proactively to directly improve the accuracy of model-generated answers. We propose USteer, a simple, training-free steering mechanism that adjusts a model's layer-wise activations during inference using the gradient of a confidence measure with respect to the activations. This procedure nudges generation toward outputs with lower uncertainty at inference time, without modifying model parameters or requiring additional supervision. We show that this approach consistently reduces hallucination across a range of tasks, demonstrating that confidence signals can be leveraged not only for detection, but also for effective inference-time control of model behavior.

Multi-agent debate (MAD) is often used to improve large language model (LLM) reasoning, but sequential debate is rarely a neutral aggregator of agents' opinions. We show that sequential MAD suffers from a pronounced first-speaker bias: agents disproportionately shape the final answer when they speak first. As a result, placing a stronger model after weaker ones can substantially offset its reasoning advantage. We then focus on the disadvantaged strong-agent-last setting and ask whether personality prompting can mitigate this imbalance. Drawing on the Big Five model, we study agreeableness and extraversion as behavioral interventions applied to either the strong or weak side. We find that their effects are trait-specific. Influence consistently shifts in the direction of lower agreeableness, and assigning low agreeableness to the stronger agent helps restore its lost influence and improves final accuracy. Extraversion, by contrast, produces less systematic changes in influence and accuracy, with its clearest effect appearing in agents' verbosity. These findings show that effective MAD design depends not only on model capability, but also on how speaking order and induced interaction behavior shape the debate process.

Large language models are reshaping ecommerce from static recommenders into interactive shopping assistants, yet real-world shopping requires session-level decision support: users reveal and revise constraints, coordinate multiple goals, and expect product-grounded recommendations over a full conversation. Existing benchmarks are mostly outcome-oriented or execution-oriented, leaving this evolving decision process under-evaluated. We introduce REALWORLDSHOP, a benchmark built on 3.28M grounded products, structured shopping episodes, a profile-grounded and actioncontrolled user simulator, and role-play evaluation. Our analysis shows that current systems produce locally plausible responses but struggle with state tracking, constraint updating, and grounded convergence, especially under ambiguous intent, bundle, and multi-intent scenarios. We further propose REALSHOP_AGENT, an executable session-control framework with explicit state management, shopping-flow control, catalog-grounded retrieval, and runtime guards. Experiments show that REALSHOP_AGENT consistently outperforms strong baselines on REALWORLDSHOP.

Emergency managers need to know where floodwater is, how deep it is, and how it will change over the coming hours across an entire river basin. During a flood, however, real-time measurements come from only a handful of stream gauges, and high-resolution hydrodynamic models are too costly to rerun each time new data arrive or to run as large ensembles. We present C-STRIDE, an observation-driven AI digital twin that turns short records from a few stream gauges, together with terrain and rainfall, into basin-wide maps of water depth and extends these predictions up to a day ahead. It is trained on simulations from a calibrated two-dimensional hydrodynamic model and needs no separate data-assimilation step. In the Des Plaines River basin near Chicago, six gauges inform predictions over 4.2 million 30-m grid cells. Terrain improves the predictions most, rainfall keeps errors from growing over longer horizons, and together they reduce errors by about 40% compared with gauge records alone. When future rainfall is known, errors remain near 15% one day ahead, compared with nearly 40% without rainfall. Given real instead of simulated gauge records, the model shifts its predictions toward the observed hydrographs at three of six gauges without retraining, and it runs about 150 times faster than the hydrodynamic model. These results show how sparse gauges, terrain, and rainfall can be combined into fast, continuously updated flood predictions, a step toward operational flood digital twins that still requires testing with real-time data and rainfall forecasts.

Deep Research agents synthesize evidence into cited reports, yet a well-cited report can still reach a misleading conclusion. Citation correctness checks whether cited sources support individual claims. It does not show whether adaptive search exposed a representative view of all documents made available for evaluation, which we call the candidate pool. Early findings redirect later queries, document choices, and stopping, so the documents an agent reads form a selective sample. Existing evaluations rarely account for this selection. We formulate the problem as adaptive evidence sampling and introduce Causal Evidence Selection Correction (CESS). CESS predicts each candidate document's evidence direction and corrects the candidate-pool average using the logged probabilities of selecting each document and reaching each search round. Shrinkage stabilizes short searches, while intervals replace point estimates when some documents cannot be sampled. We also prove that estimating the average evidence direction of a common pool differs from measuring how a change in search policy alters the evidence read. The latter requires intervention. On questions from the MS2 systematic-review benchmark, CESS reduces mean absolute error against the candidate-pool average by $9.2\%$ and reduces the estimate's change under opposing document rankings by $39.4\%$ relative to averaging the evidence scores of documents read. Across trajectories from a public Open Deep Research agent, the corresponding reductions reach $60.1\%$ and $87.2\%$. A further 4,800 trajectories under paired interventions confirm that correcting a pool estimate and measuring a policy effect are different tasks. CESS therefore audits whether the evidence direction underlying a report reflects the documents available for evaluation, while a separate intervention analysis measures the effect of search decisions.

Large Language Models (LLMs) have been applied in various fields. However, ensuring compliance and safety of LLMs, such as avoiding discrimination and bias, still remains a challenge. Current efforts mainly focus on detecting and filtering inputs and outputs of the trained models, rather than studying the intrinsic architecture of the models in real-time. To tackle this challenge, we analyze the LLMs training process and discover two critical issues: 1) Most of the existing methods are predominantly static in their approach to detection and filtering, achieving only localized optimizations without systematically enhancing the compliance of LLMs. 2) Another issue with existing approaches is the lack of real-time risk detection and mitigation across the full training process, which leads to limited flexibility. Motivated by these, we propose MASCRDM (Multi-Agent System for Compliance Risk Detection and Mitigation) during the LLM training process. Firstly, we develop a set of compliance rules based on existing Artificial Intelligence (AI) laws and a compliance-specific LLM with the instruction of compliance law experts. Then, we deconstruct LLMs into several components and identify key nodes based on the compliance knowledge graph. During LLMs training, we implement our multiple agents in the whole process, giving compliance risk alerts and suggestions for LLM developers. Experiments on discrimination and bias benchmark demonstrate that our multi-agent system can effectively improve the compliance while maintaining reasonable semantic performance. The results indicate that our method provides an executable path for mitigating compliance risk from within the LLMs systematically.

Adaptive RAG uses signals such as confidence, relevance, support, and retrieval quality to decide when to search or correct evidence. In multi-step retrieval, however, these local signals must be combined into a persistent view of what the current evidence supports, what remains missing, and which action should follow. Existing methods often use such signals as separate triggers, making it difficult to preserve a coherent evidence state across a trajectory; we call this problem evidence-state fragmentation. We introduce BELIEFRAG, a closed-loop controller that updates an explicit state over sufficiency, reliability, conflict, uncertainty, evidence gaps, and acquisition cost, then chooses among retrieval, query rewriting, verification, answering, stopping, and abstention. Across six QA benchmarks with gpt-oss-120b, BELIEFRAG reaches mean token F1 0.572 with 3.89k tokens per question, outperforming fixed iterative retrieval (0.555 F1) while using 39% fewer tokens. The same quality-cost pattern transfers to Qwen3-32B, where BELIEFRAG reaches 0.552 F1 versus 0.523 for iterative retrieval while using 35% fewer tokens. Analysis shows that the main gains come from corrective re-retrieval rather than pruning alone, while several belief dimensions are redundant and calibrated answerability plays the strongest operational role. Calibration improves threshold stability across related evidence sources, although source shift can still invalidate the same decision signal.

Learning to complete tasks in unfamiliar environments with unknown rules remains a key challenge for LLM agents. Current LLM agents often record their discoveries in prose, which may not provide a compact, explicit account of how the environment works. Inspired by how scientists organize observations into testable, predictive theories, we introduce Schema, an agent harness that organizes learning and action through interactive program induction. The LLM agent decides what to investigate and how to act, expressing its evolving understanding of the environment as executable programs. The harness consists of a persistent program workspace and a small set of interfaces for checking these programs against the interaction history, planning within them, and executing plans under step-by-step verification. Schema raises ARC-AGI-3 RHAE from 58.7% to 99.2% with the same base model, solves 100% of the public DiG-bench games, and reaches the median performance of the top-50 human players on MazeBench. Extensive analysis shows the effectiveness of Schema in unknown mechanism discovery, and ablations confirm the contribution of each component.

Multi-agent debate (MAD) can improve large language model (LLM) reasoning by allowing multiple agents to exchange and critique their answers to the same task. However, the interactions that enable agents to correct mistakes can also spread adversarial errors and steer the agents toward an incorrect answer. Although some efforts have been made to examine particular attack types on MAD, systematic evaluation of MAD under diverse attacks remains limited. A central question is whether debate mitigates adversarial influence or amplifies it. In this paper, we present MADBench, a benchmark for evaluating the security of MAD. We organize attacks into a layered taxonomy following the MAD workflow, incorporating both established attacks and new strategies tailored to debate. We evaluate six attack families over 356 source tasks and 3,958 test cases, examining their effects on the final answer and the propagation of adversarial influence. Our results show that, under attacks, MAD does not necessarily improve LLM reasoning. Compared with a single-agent baseline, MAD can mitigate attacks on answer accuracy in question-answering tasks while amplifying unauthorized reads or writes in both question-answering and workspace tasks. Moreover, even when three out of five agents collude, the attack changes the final answer from correct to wrong on only 28.30\% of tasks answered correctly without attack, while only 3.26\% of initially correct honest agents switch to wrong answers during debate.

AI agents can externalize what they learn from past tasks into reusable \emph{skills}, such as procedures, checklists, code, or other executable artifacts, that can be retrieved and reused when solving new tasks. Self-evolving skill methods keep rewriting these skills after each round of practice on training tasks, and the skill is then used on new tasks of the same kind. We ask a question: does the improvement a skill shows on its training tasks carry over to new test tasks? We test five self-evolving methods and a one-shot skill on six benchmarks, with the same model, the same agent, and the same train/test split for every method. Of the 21 skills that improve on their training tasks, 5 keep all of that improvement on the test tasks, 13 keep part of it, and 3 keep none of it. No existing method is best everywhere. When we read the skills, the ones that carry over badly often fix details that should depend on the task, such as column names and output files, or turn a fix for one failure into a rule for every task. An LLM judge that reads the skill content can often see this: it ranks finished skills the same way the test results do in 86\% of pairs. But it predicts the effect of a single edit poorly, so edits still have to be tested by running them. Based on these findings, we describe Generalizable Skill Optimization (GSO), which keeps only a guide for writing skills and writes a new skill for each task; it scores highest on all six benchmarks.

Textual skills enable large language model (LLM) based agents to accumulate reusable procedural knowledge without updating model parameters. Yet existing skill evolution remains largely confined to the text space: an optimizer must diagnose success and failure patterns, and revise skills solely from long execution trajectories and sparse task outcomes. This text-only paradigm leaves the agent's internal representations, which contain rich records of its evolving execution state, outside the skill optimization loop. We ask whether an agent can improve its external textual skills by reflecting on its own internal representations. We introduce Rep2Skill, a representation-guided framework for self-evolution on agent skills. Specifically, upon the collected agent rollouts, Rep2Skill models their internal model representation trajectories to localize turns that deviate from successful execution dynamics, and it further interprets these signals alongside the execution contexts as actionable textual feedback for targeted skill revision. Experiments on two agent environments with two open-source LLMs show that Rep2Skill consistently outperforms text-only approaches in the self-evolution setting, where the same LLM serves as both executor and optimizer without a stronger external model. This establishes a promising direction moving agent self-improvement beyond text-only reflection.

Persistent spatial memory enables embodied agents to navigate familiar environments across repeated visits. However, targets may move while unobserved, including during navigation, making remembered locations unreliable by the time an agent arrives. Despite advances in memory retrieval and state prediction, accounting for continued hidden world evolution and revising beliefs under limited visibility remain challenging. We study Evolving-World Navigation, where agents infer target locations from intermittent observations, predict their states at inspection time, and revise beliefs using visual evidence. We propose EvolvingNav, which constructs a time-indexed belief from timestamped 3D object histories through a structured persistence-relocation model. The belief distinguishes persistence at the last observed location from relocation to alternative locations and retains probability mass outside the known candidate set. An event-driven filter propagates the current belief as time elapses, forecasts target occupancy at candidate inspection times, and incorporates new RGB-D evidence. Negative observations downweight location hypotheses according to calibrated, visibility-conditioned detection probabilities, while evidence tracking prevents repeated use of the same observations. A frozen, zero-shot vision-language controller uses the updated belief to choose actions and replan. We further introduce EvoWorld-Bench, a benchmark grounded in human activity traces, comprising 54 scenes and 803,680 tasks with controlled changes before and during navigation. In simulation and real-robot experiments, EvolvingNav improves navigation success and search efficiency over the evaluated baselines. Paired experiments show the clearest gains under learnable temporal patterns, while ablations demonstrate the value of preserving uncertainty and incorporating visibility-aware evidence.

Reinforcement Learning with Verifiable Rewards (RLVR) has improved the reasoning capabilities of Multimodal Large Language Models (MLLMs), yet existing frameworks rely on coarse, sequence-level reward signals that lack the fine-grained supervision over the visually-grounded steps within a multimodal reasoning chain. We investigate this gap through the lens of two token-level metrics: visual dependency (i.e. how much a token's prediction relies on the input image features) and predictive entropy. Our empirical analysis reveals two key findings: (1) correct reasoning chains exhibit a markedly sharper entropy reduction as visual grounding intensifies, compared to incorrect ones; (2) pivotal tokens, those whose misprediction triggers reasoning collapse, are statistical outliers in the joint distribution of visual dependency and predictive entropy derived from correct chains. Motivated by these findings, we propose token-level perception-grounded advantage estimation (TPAE), which estimates token-level advantages by measuring each token's statistical consistency with the vision-entropy patterns of correct rollouts. TPAE leverages this granular score to modulate the sequence-level advantage, producing a fine-grained supervision signal that can be integrated into various RLVR frameworks. Extensive experiments on seven benchmarks show that TPAE consistently outperforms leading strong baselines, yielding more stable and efficient optimization for multimodal reasoning. The code is publicly available at https://github.com/Zhihan72/TPAE.

We present FYAN, a human--AI harness for document-level mathematical formalization. Rather than treating theorems in isolation, FYAN coordinates an end-to-end workflow spanning specification, proof planning, logical review, Lean proof construction, knowledge curation, and validation, with support for independent supervision and human guidance. A central component is evidence-grounded semantic auditing, which assesses whether formal statements faithfully preserve their informal specifications. A language model constructs structured evidence over local correspondences, omissions, scope, and logical relations, while a deterministic validator checks this evidence and produces reproducible judgments. When a substantive but admissible deviation is accepted, FYAN requires an explicit proof-transfer obligation connecting the formal statement back to a source-facing interpretation. With the same model (DeepSeek-V4.1-Flash) in every stage, FYAN proves 86 of 143 FormalTCS theorems under a strict Lean check, against 69 for a general agent harness, and raises the natural-language proof score from 0.501 to 0.851. On ConsistencyCheck, its semantic audit catches more inconsistent statements than a direct LLM judge, both on labels verified against the source (recall 0.777 vs. 0.636) and on the original labels (0.873 vs. 0.820), and localizes each mismatch it reports to a specific hypothesis, conclusion, or scope. FYAN also built ODENumLib, a 9,355-line Lean library for the numerical analysis of ordinary differential equation.

Precise numerical reasoning with Large Language Models (LLMs) is essential for expanding their applicability to complex real-world tasks. However, text-based tokenization often fragments numbers, significantly hindering precise arithmetic reasoning. Meanwhile, numerical embeddings, despite arithmetic precision, rely on context-agnostic substitution that disregards the semantic role of numbers as identifiers. To combine the complementary strengths, we propose \textbf{ANI (Adaptive Numerical Injection)}, a hybrid framework that governs the selective injection of numerical features based on the semantic context. By employing a context-aware gating mechanism, we selectively inject numerical embeddings (specifically FoNE) into the latent space, explicitly preserving nominal identifiers while enhancing quantitative operands. Through extensive evaluations across various LLMs, we demonstrate that ANI enhances MATH performance by 9.5 points over the official reference model, while maintaining robust performance on general linguistic benchmarks.

Autonomous agents powered by large language models (LLMs) are rapidly evolving into an open agentic ecosystem. To support trustworthy collaboration, industry initiatives increasingly assess agent reputation from past behavior and provide performance leaderboards. However, reputation derived from past performance may not reliably predict an agent's behavior on new tasks, particularly when malicious agents can adapt their behavior and influence other agents during collaboration. We study reputation in multi-agent debate (MAD), where multiple agents answer the same query, debate to improve their answers, and aggregate them into a final output. We present MiniRep, a reputation-based aggregation system for MAD under malicious agents. To ground our threat model in established research, we construct an attack taxonomy drawing on reputation-system attacks and software-testing mutation operators, covering strategic exploitation of reputation and subtle corruption of agent proposals. Guided by this taxonomy, MiniRep evaluates agents based on both their behavior on the current task and their reputation over time, while preventing groups of agents with highly similar responses from dominating the final decision. We assess MiniRep across diverse tasks, LLM-agent compositions, corruption placements, and attack types drawn from our taxonomy. Our experimental results show that, MiniRep outperforms both conventional MAD aggregation and conventional reputation-based approaches on MATH no matter being attacked or not. Also, under a heterogeneous 10-agent setting on MATH, MiniRep outperforms all baselines in all 28 attack conditions.

The ability of Large Language Model (LLM) agents to complete daily and professional work is receiving increasing attention. Training such agents requires realistic work scenarios. Expert-authored occupational work is costly and slow to produce, while unconstrained synthesis often yields tasks with weak factual grounding or internally inconsistent requirements. To bridge this gap, we introduce WorkGenesis, a framework that constructs executable occupational work from real-world artifacts through two core technical innovations: (1) Evidence-Based Work Construction, which grounds each unit of work in real-world evidence by retrieving public files guided by O*NET occupational knowledge and synthesizing the surrounding context, companion materials, work request, and itemwise rubric around them; and (2) Execution-Guided Consistency Verification, which renders a reference deliverable inside the constructed work, attributes every unsatisfied rubric item to the agent, the task, or the rubric, and uses task and rubric defects as feedback to iteratively repair the work until it passes the audit. Experimental results demonstrate that Fx-Work-35B, trained with simple supervised fine-tuning (SFT) on only 20K units of work synthesized by WorkGenesis, achieves the highest scores among all comparable-scale baselines on the five reported metrics across GDPvalAA-v2, APEX-Agents-AA, and JobBench (31.00 versus 24.79 average score), and even surpasses frontier models such as the 1.6T DeepSeek-V4-Pro-Preview. These results show that WorkGenesis provides scalable training data for working agents.

Despite recent progress in autoresearch, applying it to practical operations research problems, typically formulated as NP-hard mixed-integer linear or nonlinear programs (MILPs or MINLPs), remains challenging because effective research requires systematically managing competing ideas and long-horizon experimental trajectories. We introduce AutoMIP, a reusable agent skill for organizing long-horizon autoresearch in mixed-integer programming through idea pooling and algorithm tree search. AutoMIP maintains a persistent pool of complementary candidate ideas while organizing executable experiments into an algorithm tree, enabling the agent to preserve unexplored hypotheses, refine promising algorithms, and switch to alternative methodological directions based on historical states. On MILP and MINLP benchmark cohorts, AutoMIP achieves the highest final success rates among the evaluated autoresearch frameworks. On MIPLib, AutoMIP discovers new best solutions for 31 of 60 instances, surpassing existing autoresearch frameworks. On MINLPLib, it achieves new best solutions for 52 of 60 instances. Ablation studies further demonstrate the complementary contributions of idea pooling and algorithm tree search, highlighting the importance of jointly maintaining diverse research ideas and structured experimental trajectories for long-horizon autoresearch.

Textual skills provide reusable guidance for large language model agents, but existing approaches often rely on manually curated skill banks or reinforcement learning with indirect and delayed feedback. We introduce SkillFM (Skill Flow Matching), a generative framework that synthesizes task-conditioned textual skills directly without test-time skill retrieval. Our framework combines a codec for encoding and reconstructing textual skills in a continuous latent space with a conditional flow model trained using improved MeanFlow. At inference time, the learned velocity field enables single-step latent sampling, and an LLM-based decoder converts the sampled representation into textual guidance for a frozen downstream agent. We evaluate the framework on embodied tasks, question answering, and web shopping. On ALFWorld and Search-QA, our method achieves the best overall performance among the compared vector-based skill approaches. Our analyses further demonstrate that latent skill generation is an effective alternative to retrieval-based skill augmentation. Our code and training skill libraries are available at https://github.com/lulushang999/SkillFM.

Autonomous research agents can generate hypotheses and conduct experiments, but experimentation remains a major source of computational cost. A fundamental challenge is deciding which experiments are worth running, particularly when prior evidence is insufficient to resolve uncertainty. Yet current research agents lack a systematic way to leverage experimental experience when making such decisions. We introduce Experimental Experience Modeling (EEM), a framework for making informed experimental decisions by acquiring, reusing, and accumulating experimental experience. EEM extracts decision-relevant records from earlier experimental trajectories, distills them into reusable experience, and organizes them in an experience library. For a new experimental decision, EEM retrieves relevant historical experience and assesses whether it provides sufficient support for deciding whether a candidate direction warrants further investment. When historical experience is insufficient, EEM conducts a targeted, low-cost pilot experiment to acquire the missing decision-relevant experience on demand. It then combines this newly acquired experience with retrieved historical experience to determine whether the direction warrants full-scale evaluation, which requires substantial resources. The resulting experimental outcomes are further distilled into reusable experience, allowing the library to continually grow through iterative accumulation. Experiments on autonomous research benchmarks show that EEM improves research performance while reducing model interaction overhead, demonstrating the value of reusing accumulated experience and acquiring additional experience only when needed.

Causal AI is a branch of Artificial Intelligence which helps understand and reason about cause and effect relationships, not just patterns or correlations. Causal discovery aims to infer elements of the underlying causal structure--often represented as a directed graph--from observational and, when available, interventional data. While causal discovery is the fundamental step for moving beyond mere associations toward genuine understanding, and thus the basic building block of causal AI, it becomes intrinsically difficult when causal relations must be inferred from single observations. In such situations, standard causal discovery methods cannot be used and one has to identify causal relations from limited amount of information. This is typically the case for, e.g., sequences of events produced by different alarms which need to be analyzed on the fly to detect abnormal phenomena, which are usually rare. We show in this study that it is possible to leverage the predictive power of Large Language Models (LLMs) to infer causal relations between only two sequences of events. This approach, which is validated on both synthetic and real data, provides better results than standard causal discovery algorithms on several time series data, even though these data were converted into smaller, single observed sequences.

Ownership establishes rights over the use, control, and transfer of objects. Understanding these relations is essential for AI systems to interact appropriately with people and their resources. Yet how large language models (LLMs) attribute ownership under competing claims remains unclear. We introduce the Competing Ownership Attribution Task (COAT), comprising 42 scenarios, and compare ownership allocations from 24 LLM configurations with those of 108 human participants. Overall, human-model similarity is close to human-human similarity, but models show greater homogeneity in their ownership judgments. Within individual answers, models also divide ownership more evenly among claimants than humans do. Pooling responses across model configurations reveals more scenarios with a shared judgment and fewer with distinct viewpoint groups than in humans. When humans form distinct groups, models may converge on one viewpoint or between competing viewpoints. Further comparisons reveal different contextual sensitivities. As material value increases across scenarios, allocations to creators decline less sharply in models than in humans. Across scenarios differing in public recognition of later holders as owners, allocations to these holders increase in models but decrease slightly in humans. Together, these findings suggest that the evaluated LLM responses do not fully capture the diversity of participants' ownership judgments or how those judgments vary across situations. Developing socially capable AI therefore requires moving beyond overall similarity to capture the diversity and context dependence of human judgments.

Effective human-AI collaboration requires partners to establish references through interaction, which becomes fragile when descriptions are ambiguous, similar referents compete, or partners see different things. We study referential uncertainty - uncertainty over which candidate object a description refers to - in a collaborative puzzle task where a human Helper instructs an AI Worker to place pieces. The Worker must identify and communicate its uncertainty, and the Helper must recognize and act on it. We show that a separately elicited belief distribution over candidate pieces is better calibrated (ECE 0.15) and better discriminates correct from incorrect placements (AUROC 0.65) than raw action-token probabilities, which are severely overconfident (0.97 mean confidence, ECE 0.44). Across three frontier vision-language models (GPT-4.1, GPT-5, GPT-5.5), this elicited uncertainty rises predictably with instruction vagueness, but not with competing referents in context, even when those increase errors. The models seldom externalize it, asking for clarification on only 3.5-16.7% of turns. In a controlled human study (N=210), participants given only the Worker's default message accept 78% of wrong placements and cannot tell right from wrong (AUC 0.50). Precise descriptions and, especially, well-targeted hedges cut wrong-move acceptance to 36% while largely preserving correct-move acceptance, compensating for missing shared awareness such as not seeing the Worker's action. But this benefit depends on targeting: a deployable hedge derived from the model's own belief entropy inherits that signal's weakness and can do more harm than good. Externalized uncertainty helps a human partner only when it is accurately targeted.

Recent achievements in AI-assisted mathematics require intensive interaction of agents with proof assistants to generate machine-checked proof certificates. Agents interact with proof assistants such as Rocq or Lean through an interface that controls what the agent receives from the prover and the cost of these interactions. Today, these interfaces are adapted from tools designed for humans and not optimized for agents. We propose an evolutionary method where a frontier model incrementally proposes new features and only keeps the ones that improve the overall performance of smaller models. We demonstrate the effectiveness of our method by growing, on a curated set of mathematical problems, \rme, a new MCP server for the Rocq prover. On the held-out \texttt{test} split of miniF2F-Rocq, an agent equipped with \rme outperforms both the baseline that only exposes the Rocq compiler and an established MCP server, across four models from two families, in success rate, cost per solve, and time per solve. Although evolved for Rocq, the resulting server transfers to Lean, improving cost and time per solve on a subset of PutnamBench. We release \rme and its port to Lean.

Auto-research agents, LLM systems that propose, implement, train, and evaluate model changes across iterations, promise to automate applied ML's experimental loop. Over long horizons, execution accuracy is a binding constraint: a change can silently leak held-out data, omit normalization, disconnect a gradient, or leave a train/eval flag unwired, invalidating expensive runs and compounding error across iterations. We present RankEvolve, an auto-research framework for evolving generative ranking models. An Executable Operating Protocol (EOP) declares phases, gates, branches, and loops, and the runtime enforces the compiled state machine. A meta-meta-harness composes complete black-box coding-agent products, including Claude Code and Codex, as execution-graph nodes that review and repair one another's work. In a budget-matched evaluation, heterogeneous composition raises all-oracle execution accuracy from the best single-product baseline of 45.8 percent to 62.5 percent (paired +16.7 points, 95 percent CI [6.6, 26.7]) while achieving a 10.4 percent silent critical-defect rate. An implemented knowledge layer carries findings, including negative results, across iterations. In a twelve-iteration deployment on the open-source HSTU recommender, RankEvolve reported NDCG@10 of 0.2192 on MovieLens-20M LARGE (+4.48 percent over the published anchor) and 0.1948 on BASE (+2.80 percent). ExecML-HSTU, seeded by incidents from that deployment, provides the oracle benchmark for the execution-accuracy evaluation. A pre-specified LitGPT transfer split replicates the heterogeneous-composition effect beyond recommendation (+12.5 points, 95 percent CI [3.0, 22.0]), and a paired ablation isolates per-step from full-protocol instruction injection. These results characterize when runtime-controlled composition of coding-agent products improves execution accuracy.

In this work we study the problem of MAPFC, a post-optimization step for Multi-Agent Path Finding (MAPF) plans where we are given a feasible plan produced by a modern MAPF solver and are tasked with removing avoidable moves while preserving feasibility. This NP-hard problem naturally arises when using learning-based state-of-the-art (SOTA) solvers which construct plans that contain redundant moves that can be removed. Recently, Tang et al. presented Judgelight, which uses Integer Linear Programming (ILP) to solve MAPFC. Importantly, the ILP is constructed over all agents jointly, so its cost is governed by the full instance rather than by the small coupled residue that actually requires joint reasoning. Our key insight, motivating this work, is that MAPFC instances naturally decompose into independent sub-problems, most of which involve a single agent and can be solved without any inter-agent reasoning. To this end, we first identify which agents need to coordinate their motion and partition the instance into sub-problems accordingly. For the cases where no coordination is required, we introduce an extremely lightweight solver that is $\approx\!1{,}900\times$ faster than Judgelight. For cases where coordination is required, Judgelight can be used but we introduce an alternative CBS-like solver which is more efficient on easier problems. The resulting framework is exact, uses no commercial ILP solver, and matches Judgelight's quality while running substantially faster on the coordination-light majority of instances; on the coordination-heavy instances we propose a regime-aware hybrid planner that falls back to Judgelight. Over all benchmarks tested, this planner achieves a median $10.5\times$ per-instance speedup over Judgelight.

Delegating complete application development to coding agents requires preserving the intended design rather than simply producing plausible outputs through naive prompting. Game development provides a demanding testbed, as long-form Game Design Documents (GDDs) describe requirements that must work together across game logic, visual rendering, and player interactions. However, existing game-development benchmarks typically use compact specifications and provide limited support for evaluating interdependent requirements across these aspects in long-form GDDs. We introduce A2Z GameSpec-Bench, a benchmark of 100 long-form GDDs for evaluating end-to-end game development by agents. We measure faithfulness by checking whether the game satisfies the GDD requirements and preserves the relationships among them. Each GDD is turned into a dependency-aware contract that contains rules, constraints, and prerequisite relations. Following game-development practices, we combine source-code inspection with agent-generated test policies for scenario-based replay and adaptive playtesting. The contract remains fixed across agents and revision rounds, while judgments and evidence linked to the same requirements support consistent comparison and failure detection. Our evaluations show that current agents struggle to jointly satisfy interdependent requirements across code implementation and actual play. Requirement-specific feedback improves GDD Fidelity by 10.9% relative to self-revision after two rounds. A2Z GameSpec-Bench assesses end-to-end specification-following ability beyond implementation judgments and provides targeted feedback to support more faithful game development. Code and datasets are available at https://a2z-gamespec-bench.github.io.

Deploying vision-and-language navigation (VLN) agents in unseen environments remains challenging because unfamiliar layouts and visual conditions can cause execution to go off track. Rather than relying on continuous human supervision, a practical strategy is to selectively request corrective guidance, recover the ongoing task, and reuse corrective interactions to improve subsequent navigation. We propose Abstention-aware Visual Error Recovery and Training for Vision-and-Language Navigation (AVERT-VLN), a closed-loop framework that uses a plug-in vision-language Monitor for online human-assisted recovery and offline preference learning. The Monitor operates separately from navigation decision generation and assesses instruction-execution consistency from the instruction, visual history, and current observation. To train the Monitor for deviation recognition, we construct LOSTNAV DATASET with 20K counterfactual risk trajectories and rule-based deviation labels. The Monitor is first fine-tuned on 40K normal trajectories to assess instruction progress and then jointly fine-tuned on normal and risk trajectories to recognize semantic deviations. At runtime, Asynchronous Sidecar Monitoring evaluates execution alongside the navigation model. When the controller accepts a LOST verdict, it suspends autonomous execution and requests human guidance for recovery. For offline policy improvement, Trajectory-Anchored Preference Learning converts deviation-associated failures into decision-level preference pairs under shared decision contexts, restricting supervision to the decisions targeted for correction. Under human-assisted evaluation, the full AVERT-VLN system achieves success rates of 76.2% and 66.3% on the val-unseen splits of R2R-CE and RxR-CE, respectively. The same monitoring and human-assisted recovery interface also improves success rates across the three evaluated navigation architectures.

Although conventional world models - auto-regressive or diffusion models based on transformers or convolutional networks - may learn surface statistics of world dynamics, can they learn the exact world dynamics from its observed history? Leveraging cellular automata as a simple testbed, we find the answer to be no in many cases. Conventional architectures predict most pixels correctly yet rarely complete a rollout: a CNN predicts 96.3% of cells but completes 18.9% of rollouts; a joint diffusion model completes none. We trace the gap to three failure modes of these world models - namely, they fail to exactly capture spatial locality, temporal locality or temporal stability. Simple changes repair each: (1) for spatial locality, two-dimensional rotary positions lift a transformer from 39.1% to 100% on the Game of Life; (2) for temporal locality, handing each token its cell's previous-frame neighbourhood lifts the same transformer from 25.8% to 99.9% on unseen rules; (3) for temporal stability, causal freezing lifts the same diffusion weights from 42.2% to 99.9%. None of the three changes touches the architectural backbone; each only modifies the information flow within it. We also compare joint and ordered sampling on billiards and, in an exploratory study, on a simulated Burgers equation.

Embodied agents must determine where to act, anticipate the resulting scene changes, and interpret observed outcomes to guide subsequent actions. This requires connecting 4D interaction understanding, which explains how past actions changed the scene, with spatially grounded planning, which determines how and where to act toward a goal and anticipates the resulting scene changes. We introduce ChronoGraph, a functional 4D scene graph that links actions on affordance parts to semantic and geometric state changes. By representing observed and anticipated transitions in the same form, it provides a shared basis for understanding and planning. We construct ChronoGraphBench through an automatic data engine that converts human-interaction videos and simulated robot trajectories into graph-annotated questions for training and evaluating Vision-Language Models (VLMs) on both tasks. Using these annotations, we train ChronoGraphVLM by adapting pretrained VLMs in two stages. Graph-as-Chain-of-Thought supervised fine-tuning teaches the models to reconstruct observed transitions and predict future ones as graph traces before answering. Subsequent joint 4D graph reinforcement learning directly rewards graph properties and answer correctness. Experiments across model scales show improvements over the corresponding pretrained baselines and zero-shot transfer to VLM4D. Real-world demonstrations further show that graph-based planning and affordance grounding support mobile manipulation through existing robot skills without additional fine-tuning.

Top researchers have good intuition, but do language models have as good intuition about model training as top AI researchers? To measure model intuition of LLMs and humans, we introduce the ArchitectureIQ benchmark. Each question presents a synthetic dataset and several training recipes, and the test-taker is asked to predict the recipe yielding the best test metric. Overall, we find that LLMs' model intuition is good but has four limitations: (1) The intuition is imperfect, or even sub-human in some cases. Frontier models achieve around 76% accuracy (random choice 33%) vs best human researcher (66.0%), yet remain far from perfect. For architecture-only questions, best human achieves 65% while GPT-6 Astra only has 38%. (2) The intuition is empirical, not structured, supported by the fact that more CoT compute does not lead to substantial improvement. Unlike math, we still lack a "Science of AI" language that enables structured reasoning on AI. (3) The intuition is not maximally condensed, and can be further compressed into a knoledge base. Our constructed knowledge base with only 20 items yields large gains for weak models: GPT-4o equipped with the accumulated knowledge almost matches the performance of Claude Opus 5. (4) The intuition is insensitive to dataset properties, but the best model should in general depend on data properties. This suggests that data is the real "dark matter" in AI -- LLMs (so do human researchers) understand too little about data, even less than model architectures.

Benchmarks, audits, and agent protocols describe performance, permissions, and repair, but not how observed evidence should change an agent's authority during a consequential task. We call this the assurance-transition gap. We propose a Runtime Assurance Contract (RAC), a policy-level formal schema binding autonomy boundaries, component eligibility, evidence state, transition policy, human-review capacity, and non-compensatory gates. Under RAC, soft metrics may inform routing, whereas a failed or unknown mandatory gate forces retry, switch, escalation, deferral, or stop; aggregate performance cannot authorize action. We define the contract, an evidence record, a permission rule, and five invariants, and illustrate them in clinical, industrial, and judicial failure probes. We then report a deterministic failure-injection study in agentic coding: 280 constructed cases evaluated by a gate conjunction, a score-only rule, and a restricted protocol baseline. At the published example weights and threshold, the score rule admits 80 of 100 block-required injections and all 40 review-required injections. Tuned in hindsight, it matches the conjunction on this corpus. For positive weights, a positive threshold, binary risk signals, zero-signal controls, and an injected case firing each signal alone, we show that exact agreement holds if and only if the threshold does not exceed the smallest weight. A separate set of 18 hand-authored traces checks version-pinned evidence and review transitions against simpler policy variants. In a further prospective synthetic holdout of 24 episodes, two blinded LLM judges assign identical labels to all 72 action attempts; RAC and a separately implemented full stateful baseline both match these labels. These studies test mechanisms on synthetic cases; they establish neither deployed safety nor cross-domain effectiveness.

Cooperative language-model agents must coordinate over long horizons and adapt to changing environments and to partners with unfamiliar conventions, yet existing agents map observations to actions without separating persistent coordination strategies from their tactical execution. We introduce OverForge, a training-free hierarchical architecture that separates strategic reasoning over roles and divisions of labour from tactical reasoning over actions within each agent's private, partner-conditioned world model. A metacognitive Prefrontal Cortex Module couples the two levels by forming strategy-action branches, imagining their consequences with a forward model, and committing when confident. In OvercookedV2, OverForge delivers 7 soups in a connected kitchen versus 3 for each flat LLM baseline, retains agreed roles, and adopts roles proposed by unfamiliar partners. Ablations and a fixed-strategy probe show that persistent strategies guide tactical adaptation while each reasoning level contributes to coordination. Memory restarts show that cross-episode partner knowledge supports task performance and partner prediction, linking the hierarchy to continual adaptation.

Large language models frequently fail to balance staying truthful with being supportive. They often exhibit sycophancy in responses to users, agreeing with false claims, offering unwarranted flattery, and giving advice skewed toward users' expressed views. In reality, sycophancy rarely happens in a single exchange; it may emerge organically as users repeatedly insist or subtly steer the dialogue over time. Current evaluations, however, rely on rigid, single-turn tests or fixed scripts that fail to capture these natural dynamics. Furthermore, these benchmarks often mistake showing basic empathy for yielding, penalizing models for acknowledging a user's feeling. This view may drive future models to over-correct into cold, dismissive rigidity. To address this gap, we introduce FIGS (Factual Integrity and Grounded Support), a dual-axis evaluation framework built around extended, realistic dialogue. We use an adaptive 10-turn conversational simulator that dynamically challenges the target model, reflecting how users repeat requests, push back, or steer a conversation toward a preferred answer. To accurately evaluate these trajectories, we apply a taxonomy that strictly separates Sycophancy (whether the model holds firm to the truth and keeps its praise proportional) from Calibrated Validation (showing empathetic understanding of the user's feelings without overdoing it). We release our complete testing environment, including 500 diverse multi-turn scenarios and an automated judge. Our evaluation of leading models reveals a consistent trade-off: over the course of a sustained interaction, current systems either slowly drift to sycophancy or over-correct into robotic detachment. This demonstrates that balancing honesty with appropriate support throughout a natural conversation remains a critical, unsolved challenge.

Scientific software presents a demanding test for computer-using agents based on visual language models (VLMs): completing a research workflow requires interpreting specialized interfaces, manipulating scientific objects, and producing verifiable results. We thus introduce OSWorld-Science, a benchmark and evaluation environment that combines scientifically meaningful tasks, artifact-based evaluation, and an efficient agent harness for studying computer use in the scientific domain. The benchmark contains 12 VLMs and 146 high-quality tasks across several scientific domains and software configurations, covering workflows such as molecular drawing and retrosynthesis, pathology image analysis, statistical computing, and physical simulation. Tasks are developed through expert proposals and iterative human--AI co-design, with selection guided by scientific value and difficulty. Task-specific execution-based evaluators inspect application states and generated artifacts, including molecular structures, segmentation masks, plots, and numerical results, and award partial credit for incomplete outcomes. Our special harness integrates model adapters, interaction-loop control, and trajectory logging to support comparisons of models and interaction strategies. Our results show that current state-of-the-art VLMs with a strong harness still face challenges in addressing key questions in the scientific domains. We also analyze the benchmarking results across multi-linguistics, reasoning efforts, context length and other factors and derive several important conclusions and directions to assist future development. Overall, we provide an integrated framework connecting expert-defined scientific goals to verifiable software outcomes, enabling systematic evaluation of both agent capabilities and harness design in scientific workflows.

Corporate and investment banking teams use presentations to support credit decisions and advise clients on financing and transactions. Producing these decks requires reconciling financial data, tracing sources and turning analysis into a recommendation. We retrospectively study the development of an agentic harness combining a 27B language model, financial calculations, narrative templates and validation checks. LLM judges guide engineering changes and assess the resulting decks, raising the question of whether higher scores reflect better documents or changes in grading. In shared-session text-only grading with template markers removed, five judges score the complete system 20.4 to 33.6 points out of 95 above the same model generating directly from a short prompt. Every judge scores the system higher on all seventeen development deliverables. Margins against direct Opus generation from a short prompt range from -4.7 to +0.8 points. Judges agree on broad progress across development rounds but agree less on final-deck rankings than on pooled scores. Repeated grading also shifts scores on unchanged decks, making small improvements difficult to distinguish from judge variability.

Multi-modal integrated sensing and communication (ISAC) enables environmental perception and reliable connectivity for intelligent wireless networks. Data-driven multi-modal ISAC models depend heavily on annotated real-world data to learn relationships across sensing and wireless observations, thereby constraining scalable deployment. Although synthetic data generation reduces the burden, adapting existing simulation pipelines to a target deployment requires consistent scene, sensing, wireless, and learning configurations, while mismatches among these coupled components impair sim-to-real transferability. To address the challenge, we propose an agentic artificial intelligence (AI) framework for sim-to-real multi-modal ISAC, named AIMS. Given a natural-language deployment request specifying the target task, deployment conditions, and real-data budget, AIMS derives a deployment-specific sim-to-real configuration and coordinates its execution to produce a deployment-specific task model. A two-agent architecture coordinates scene construction with task learning. A scene construction agent generates geographically grounded, synchronized sensing and wireless records from shared physical states, while a scene understanding agent configures task-relevant modalities and mixture-of-experts (MoE) learning for zero-shot inference or few-shot adaptation. Structured domain knowledge guides dependency-aware planning, while validation evidence supports feedback-driven revision of affected decisions. Experiments on the real-world DeepSense~6G dataset demonstrate improved vehicle detection and beam prediction over the considered simulation and fusion baselines. A separate orchestration benchmark evaluates task interpretation, dependency reasoning, and feedback-driven replanning across diverse deployment requests, showing improved plan correctness with structured domain knowledge and validation feedback.

Background. A component replacement in a language-model agent changes an execution trajectory, potentially altering later observations, resource use, and recovery opportunities. Different evidence is needed to assess its task-level benefit and the contribution of local decision quality. Methods. This critical scoping review maps 348 studies and examines 90 comparison records: 88 from 40 included studies and two from supplementary studies. Eight purposively selected cases structure the synthesis around the replaced decision, executed conditions, measurement comparability, controls, and remaining explanations. Results. Of 222 studies reporting local decision metrics, 142 also report measured task endpoints and 49 report proxies. These counts identify studies that report both types of measurement, without establishing that the measurements come from matched comparisons. Outcome Monitors reports a package-level completion gain whose attribution to detector quality remains limited; First-chunk selection reports a local improvement assessed against an offline proxy endpoint; Evidence-Carrying Termination reports fewer premature unsupported terminations and completion non-inferiority, without establishing completion superiority. Cross-case analysis identifies three candidate mechanisms involving recovery and disruption, intervention timing, and downstream use. Attribution and deployment depend on the comparison controls, label definitions, and information available to the controller. Conclusions. The review distinguishes the task-level benefit of a component replacement from the contribution of local decision quality and derives eight claim-specific reporting items. Neither online execution nor simultaneous gains in local and task metrics alone establish that better local decisions explain the task-level gain.

An unsuccessful LLM agent rollout contains more information than its final reward: the observations available to the agent, the actions it chose, and the environment's responses. Reusing this experience for learning requires identifying a decision to revise and testing a concrete alternative. We introduce the Agent Error Dataset (AED), comprising 50,228 error-diagnosis pairs from 9,961 source tasks across 33 environments, 19 harness families, and 23 policy models in text-based agent systems. We retain source traces and execution metadata to support cross-setting failure analysis and re-diagnosis without repeating the original rollout. Our five-stage Agentic Error-to-Training (AET) pipeline collects natural failures, generates diagnoses and proposed corrections, and checks them against recorded evidence. Where replay is supported, we compare corrections with original-action retries from the same checkpoint under matched execution settings. We then construct separate training views for diagnosis and actor recovery. Across 3,062 matched replay pairs, first-proposal corrections raise verifier pass rates from 18.4% to 51.1%, a gain of 32.7 percentage points. Using a separately frozen diagnosis release, full-diagnosis fine-tuning on 1,656 source tasks raises Qwen3-8B's exact-step agreement with internal teacher labels from 47.2% to 63.6%, averaged over three seeds on a 943-case holdout. The strongest prompted reference in this comparison scores 54.7%, and mean agreement improves at each of four increasing training-set sizes. In a single-seed comparison of actor-training recipes, action-only repair training scores 6.67 percentage points higher on WebShop-lite than success-only training.

Reinforcement learning with verifiable rewards (RLVR) has become an important approach for improving reasoning during post-training. Recent work suggests that some difficult prompts remain resistant to learning even when they occasionally produce correct solutions. We revisit this unlearnability phenomenon and find that the affected prompts do improve, at roughly one third of the learnable rate, while the difficulty-defined set used to study them is much less reproducible than expected. These difficulty labels are estimated from a limited number of sampled responses. Combining them across seeds can further change which prompts are selected instead of simply reducing measurement noise. We develop a sampling-based framework for quantifying this instability and determining how much evaluation is required for difficulty assignments to reproduce reliably. We also revisit the gradient-similarity evidence proposed to explain unlearnability and show that part of the observed separation arises because difficult prompts provide fewer correct rollouts from which their gradients can be estimated. Matching this sample count weakens the gradient difference but does not remove it. Overall, the slow-learning phenomenon survives our reanalysis, while both the prompts used to define it and the evidence used to explain it require more careful measurement.

Language agents are expected to solve increasingly complex tasks, creating a growing need for continual improvement. One promising approach is to evolve the agent harness, the software that governs tool use, memory management, and task execution, while keeping the underlying language model fixed. Recent methods automate this process by using a meta coding agent to modify the harness based on execution feedback. However, relying on that agent's existing knowledge and observed failures can restrict exploration and make adaptation reactive. Inspired by how human experts learn from the research literature for new solutions, we introduce ScholarEvolve, a framework that automatically draws on state-of-the-art research to guide harness evolution. ScholarEvolve organizes the harness evolution directions into functional modules and uses topic modeling to identify distinct improvement strategies for each module. It implements these strategies and evaluates their combinations to improve task performance. Moreover, the framework is designed to incorporate new publications over time, allowing research advances to drive proactive lifelong evolution. Experiments demonstrate improvements on AppWorld and Tau2-Bench. ScholarEvolve raises Qwen3.5-27B task goal completion from 49.6% to 63.6% on AppWorld Challenge, and raises GPT-5.4-mini pass@1 from 72.7% to 81.9% on Tau2-Bench Telecom.

Multi-Agent Path Finding (MAPF) aims to find collision-free paths for multiple agents in a shared environment. Classical MAPF assumes that all static obstacles are known in advance, but real-world environments can change unexpectedly due to fallen objects, spills, or other local disturbances. When such changes are spatially correlated, an observation can inform traversability estimates beyond the observed location. Prior approaches address uncertainty in traversability through contingent plans or replanning based on direct observations, but do not leverage this spatial dependence to infer the traversability of nearby unobserved locations. As a result, they cannot use one observation to anticipate nearby unobserved obstacles that may cause costly rerouting later. We focus on Belief-Aware MAPF, where map discrepancies are fixed during execution but initially unknown, and observations can be informative beyond the observed location. We propose Multi-Agent Gaussian belief Inference for Coordination (MAGIC), a framework that updates a shared belief about traversability online based on agents' observations. MAGIC uses a Gaussian Markov Random Field and Gaussian Belief Propagation to approximately infer traversability and construct detour-aware costs for standard MAPF planners. Our experiments on MAPF benchmarks show that MAGIC reduces the executed sum of costs compared to existing approaches on 96.3% of instances, across several planner families and teams of up to 800 agents, demonstrating its applicability to large-scale MAPF problems.

On-policy distillation (OPD) is a promising approach for training language agents, providing dense teacher supervision on student-generated trajectories. However, in multi-turn interaction, an incorrect action changes the states the student encounters later, so errors compound across turns. In preliminary experiments across three Qwen3 models (8B to 235B), we find that more than half of the failed rollouts contain a pivotal mistake, an action that moves the agent farther from completing the task, and this mistake typically occurs early. These pivotal mistakes often remain recoverable: guiding the model for only a few turns after the pivotal turn can restore task success. We therefore propose PivotOPD, an on-policy distillation framework that jointly trains the student to prevent pivotal mistakes and to recover from the states they create. At each pivotal mistake, a teacher model provides a gold action and then names a recovery action at each of the next few turns. Preventive distillation uses the gold action with reverse KL to steer the student away from the pivotal mistake, while recovery distillation uses the recovery actions with forward KL to transfer recovery behaviors that the student rarely samples. Against 13 baselines on ALFWorld, WebShop, and Search-based QA, PivotOPD achieves the strongest average performance for both Qwen3-1.7B and Qwen3-8B students, improving over the strongest baseline on ALFWorld by +5.5% with the 1.7B student. The gains also transfer to another model family on the software engineering domain, where PivotOPD raises the resolve rate of a Nemotron-3.5 student on SWE-Bench Verified by +3.2%. Project page: https://research.nvidia.com/labs/lpr/pivotopd/

Recent autonomous machine learning engineering (MLE) agents have made significant progress on public leaderboards. Often motivated by progress stagnation over long-horizon cycles and limited Large Language Model (LLM) primitives, modern MLE agents are deployed on top of increasingly elaborate machinery: multi-agent orchestrators, dedicated retrieval subagents, and more. While such harnesses expand, the use of more primitive but improved coding agents - where LLMs have direct access to the execution environment through read, write, and bash primitives - has received little attention in the field. In this paper we find that, under an equal time budget and the same frontier LLM backbone, open-source state-of-the-art harnesses provide no advantages over a single session of a minimal-harness coding agent baseline, pointing to the backbone as the primary driver for performance. Via a series of large-scale systematic ablation studies, we argue that the machinery layers become redundant in the coding agent setting. We conclude that the effort spent elaborating hand-crafted harnesses around strong models yields poor returns for current MLE benchmarks.

We present Cogentic, a multi-agent harness for automated proof discovery on open research problems. While frontier language models can generate strong mathematical ideas in a single shot, single-shot generation is often insufficient for open problems that require exploring multiple competing conjectures, overcoming subtle technical obstructions, and retaining intermediate progress over a long horizon. Cogentic addresses these challenges through an iterative prove--verify loop in which an orchestrator allocates a population of independent provers across distinct proof directions, subjects their output to adversarial verification by several specialized components, and promotes confirmed intermediate results into a persistent verified ledger that later rounds build on. The harness is designed to be able to solve research-level math and theoretical computer science problems. Using Gemini as the base model, Cogentic produced novel results on five open problems across online learning, auction theory, and mechanism design. Each result was independently verified by domain experts and is developed in full in companion papers. We list these results, and new ones as they are verified, at https://sites.google.com/view/cogentic .

As interactive 3D worlds are increasingly used to study intelligent behavior, it becomes important to develop efficient pipelines for identifying anomalies in these simulated environments, such as floating objects, traversable walls, or objects inconsistent with the surrounding scene. Multimodal AI systems, including vision-language models (VLMs) and vision-language-action models (VLAs), have shown potential for automating this task. However, 3D world auditing is complex, requiring the close coupling of two distinct capabilities: action, to navigate the 3D world and search for anomalies systematically and efficiently; and visual reasoning, to understand the environment and identify anomalies from multimodal observations. It remains largely unexplored whether multimodal agents can effectively couple these two capabilities, using visual reasoning to identify potential anomalies while taking actions to validate them. In this paper, we introduce WorldAuditBench, a benchmark for 3D world auditing comprising 213 anomaly tasks across 13 environments built with Unreal Engine 5 and Three.js, spanning five anomaly families. We evaluate five frontier models under a fixed exploration budget using two auditing paradigms: VLA-based exploration followed by VLM-based anomaly identification, and an end-to-end VLM agent in which visual reasoning directly guides action selection. Across the evaluated models and two paradigms, success rates range from 6.6% to 42.3%, substantially below human performance (83.4%). Through the task of world auditing, WorldAuditBench provides a testbed for studying how multimodal agents couple action and visual reasoning in interactive 3D environments, while highlighting current limitations in their ability to gather and interpret evidence during exploration.

Automating the search for effective harnesses is an important step toward enabling agents to recursively self-improve. Existing harness optimizations typically produce a single global harness that is applied uniformly across task instances. However, a harness that works well on average may not be optimal for every instance. We introduce Turbo Harness, a framework that can adapt a globally optimized harness to each instance by reusing information generated during the original optimization process. Specifically, Turbo Harness recycles artifacts produced during a completed global harness optimization run, and summarizes them into a structured playbook. We train a harness editor to leverage this prior optimization experience to generate instance-specific patches to the global harness. At inference time, the editor uses the instance and the playbook to construct a tailored harness in which the execution model operates. Through numerical experiments, we show that Turbo Harness consistently outperforms existing harness optimization baselines across seven benchmarks spanning interactive agent tasks, software engineering, and long-horizon terminal tasks.

As large language models increasingly power personal assistants, users expect them to reason over multimodal life histories, from recognizing people to understanding events to aggregating patterns, yet existing benchmarks primarily target concept-level recognition. We introduce Life-Bench, a fully synthetic, human-verified multimodal benchmark of over 11,800 question-answer pairs across 10 tasks, organized by required evidence scope: concept identification, event understanding, and aggregated reasoning. The benchmark's photo-centric personal histories are distributionally aligned with real user accounts under embedding statistic. The interconnected structure of personal data invites graph-based solutions; we propose LifeGraph, a personal knowledge graph framework providing structured retrieval with on-demand access to source visual evidence, showing particular promise on event and aggregated tasks. Systematic evaluation of four retrieval paradigms on Life-Bench demonstrates that accuracy degrades sharply with evidence scope, falling below 0.40 on aggregated tasks, and that no single paradigm dominates across categories. Performance beyond concept recognition remains modest for all evaluated methods, establishing personalization over multimodal histories as an open challenge and Life-Bench as a testbed for future progress.

The rapid proliferation of large language models (LLMs) in the context of education has introduced significant challenges in enforcement of academic integrity, especially in programming courses. We present Argus, an automated detection system for LLM-assisted student work in undergraduate C programming assignments. Argus integrates behavioral and stylistic indicators to create a holistic picture of the student's progress through an assignment and surfaces anomalies that point to potential misuse of LLM assistance. We quantify and analyze data over six years of Spring semester offerings in a large-enrollment CS2 course at Purdue University using Argus, finding that 45% of enrolled students exhibited patterns consistent with LLM-assisted code development in Spring 2026. To contextualize these results, we analyze the relationship between flagged LLM use and student performance on written, in-person proctored examinations, and find a significant negative correlation. We also explore the problem of mitigating false positives, recognizing that erroneous accusations of academic integrity carry significant consequences for students and instructors alike, particularly in the context of large enrollment courses. We argue that any automated detection system must be accompanied by a structured process for human review. We discuss the consequences for future course design, changing academic policy as these tools become more ubiquitous, and the pedagogical implications of LLM-based tools in computer science education.

Monitoring cognitive impairment (CI) in motor neuron disease (MND) is essential for timely treatment and care, yet challenging due to co-occurring speech difficulties. The Edinburgh Cognitive and Behavioural ALS Screen (ECAS) provides a robust metric for CI assessment, with the Verbal Fluency Index (VFI) a central element. Building on recent advances in automated speech analysis, this study proposes a system for estimating VFI. It leverages a unique MND dataset and combines ASR (WhisperX) and VAD (Silero) with refined timestamping to predict the VFI and extract several clinically interpretable measures. Our approach outperformed systems based on traditional acoustic features and self-supervised embeddings, evaluated using multiple regression algorithms. Clinically inspired features consistently outperformed the other sets, with the best models achieving strong results (P-words: R2 0.9, NRMSE 0.05; S-words: R2 0.8, NRMSE 0.08), demonstrating the feasibility of automated VFI estimation.

Imitation learning enables robots to acquire complex skills directly from massive demonstration datasets, but its performance degrades severely when datasets are contaminated with suboptimal or noisy demonstrations. While prior quality-assessment methods attempt to filter or reweight data, they typically rely on manual pre-selection of expert reference data or task-specific heuristics, limiting scalability. To address this challenge, we introduce SynIL (Synergy-based Imitation Learning), a novel framework for automated, label-free demonstration quality assessment in offline reinforcement learning. Grounded in neuroscientific evidence that motor synergy, a low-dimensional coordinated structure in movement, correlates directly with motor proficiency, SynIL algorithmically quantifies synergy manifestation to generate dense, transition-level reward signals via self-supervised reward regression. Comprehensive evaluations on D4RL locomotion benchmarks and multi-human Robomimic manipulation datasets demonstrate that synergy-derived rewards correlate strongly with ground-truth rewards. Furthermore, SynIL substantially outperforms Behavior Cloning (BC) and achieves performance comparable to, and in sparse-reward human teleoperation scenarios, superior to, offline reinforcement learning trained on true environment rewards.

Deepfakes no longer need to fake a whole video. Generators that read the transcript now alter only the few seconds in which a video's meaning turns, so a forgery hides in a small, unknown fraction of the video. Yet detectors still read every one-second window of both the audio and image streams, spending nearly all of their compute where nothing was altered. We observe that deciding where to look is far cheaper than looking. We present ModalFidelity, a lightweight router that previews each window and decides, before any forensic detector runs, which stream is worth reading, under a hard compute budget it can never exceed. On AV-Deepfake1M, reading at most a fifth of the windows, it is more accurate than gating after the detectors at 15.9x less compute, and retains over 96% of the accuracy of an oracle that knows where every forgery lies.

The rapid evolution of image manipulation techniques has raised growing public security concerns. Existing Image Forgery Localization (IFL) methods can accurately localize manipulated regions but are often unable to adapt to newly emerging forgeries. In real-world forensic scenarios, data typically arrive sequentially, yet continual model adaptation remains largely unexplored in IFL. To bridge this gap, we introduce the first continual learning framework for IFL and establish a comprehensive benchmark under two realistic data-evolution protocols: cross-dataset and cross-content continual learning. Evaluations of representative state-of-the-art IFL and continual learning methods reveal substantial performance degradation, highlighting two key challenges: (1) adaptively capturing intrinsic forensic traces from incoming data across unseen domains, and (2) preserving previously acquired forensic knowledge during sequential adaptation. To address these challenges, we propose a forensic-aware continual adaptation framework. First, a forensic trace mining module employs Spatial Mixture-of-Forensic-Experts (SMoFE) to dynamically route complementary forensic cues across spatial locations, together with Forensic Evidence-Guided Dense Prompting (FEGDP) to transform low-level forensic traces into structured localization evidence for SAM. Second, Fisher-weighted LoRA Gradient (FLAG) surgery identifies old-task-sensitive adaptation directions and suppresses conflicting updates, mitigating catastrophic forgetting while preserving plasticity for emerging forgery domains. Extensive experiments demonstrate state-of-the-art performance in both pixel-level forgery localization and image-level forgery detection across diverse continual learning scenarios.

Text-to-image (T2I) models have substantially improved in language understanding, in-image text rendering, and visual composition, while their safety mechanisms do not always keep pace with these capabilities. This creates a cross-modal attack surface in which harmful semantics can remain inconspicuous in a serialized prompt yet emerge through image-level composition. We propose CollageAttack, an automated single-prompt black-box jailbreak that shifts semantic assembly into the image plane by combining context-relevant scenes, scene-grounded textual carriers, and spatially distributed text fragments. Experiments across multiple open-weight and commercial T2I models show that CollageAttack achieves attack success rates of up to 86.0%, outperforming the strongest baseline on the same model by 18.5 percentage points, while consistently producing more harmful outputs and preserving the source intent. We further find that distributed textual fragments can reconstruct the intended semantics after generation, with visual composition producing stronger communicative impact than text alone. These results reveal a cross-modal safety gap in which harmful meaning emerges from the composition of individually less explicit elements.

Code editing requires a model to decide where to make changes, generate the new content, and preserve everything else. We study how masked diffusion language models divide these responsibilities across four editing interfaces: whole-file rewriting, search-and-replace, locate-then-infill, and token-level editing. Experiments on CanItEdit reveal a composition gap: diffusion models can generate coordinated changes when the correct edit locations are supplied, but much of this capability is lost when those locations must be predicted. Access to the intact original code helps the model fill multiple edit regions, yet does not resolve the difficulty of selecting those regions. By varying the editable regions while holding the generation model and decoding procedure fixed, we identify two distinct requirements for successful editing: covering every required change and placing precise boundaries around it. Missing a required region prevents the corresponding change, while widening regions to ensure coverage can sharply reduce success by requiring unchanged code to be regenerated. A sentence-level Wiki editing probe shows the same qualitative gap between supplied and predicted locations beyond code. These findings show why strong infilling capability alone does not ensure reliable editing: the interface must expose all required changes while limiting regeneration of unchanged code.

Values such as honesty, autonomy, and confidentiality are often regarded as general principles underpinning AI alignment. However, what it means to act in accordance with these values can depend on the context in which a decision is made. In this paper, we ask whether large language models (LLMs) appropriately adapt the application of a value across professional settings, while remaining consistent when contextual changes do not alter the relevant professional norm. To study this, we introduce ContextAdapt, an evaluation framework covering honesty, autonomy, and confidentiality across medicine, law, finance, and national security. Drawing on primary-source professional and regulatory documents, we construct a value x domain framework and use this to develop scenarios testing both default professional rules and recognised exceptions. We evaluate 12 LLMs on both the actions they recommend and the justifications they provide. In our main experiment, models achieve 95.6% mean appropriateness, although the use of the correct domain-specific justification varies substantially across models, from 25.6% to 76.9%. In a separate factorial experiment, explicitly naming the professional domain and changing the role of the model have limited effect on behaviour. Varying stakes, however, reveals severe but localised failures: in some cases, models alter their responses even though the underlying professional obligation remains unchanged. In particular, perceived severity appears to act as a cue for disclosure across both honesty and confidentiality scenarios. These results show that evaluating value alignment requires us to consider not only whether models follow abstract principles, but whether they apply them appropriately across different contexts.

In this paper we propose NinaXander, a series of composed language models obtained by connecting layers of frozen language models from different architecture families with a single trained shared-latent adapter. A composed model runs the first layers of one model, converts the resulting intermediate representation once with the adapter, and then runs the remaining layers of the other model. Once the adapter is trained, several composed models that connect at different layers are obtained without retraining. Using the recurrent RWKV-4-Raven-7B and the Transformer-based Tulu-Pythia-6.9b, abbreviated as RWKV and Pythia, this study examines whether frozen models from different families can be recombined post hoc. The composed models answered multiple-choice questions, and those whose generations we examined produced syntactically well-formed text. The configuration that combines the first 5 layers of Pythia with the remaining 27 layers of RWKV reduced the Transformer key-value (KV) cache by 84.4% with accuracy not significantly different from that of RWKV alone. In multiple-choice accuracy, however, no composed model matched the parent model Pythia, and language-modeling performance decreased sharply on WikiText, a corpus of Wikipedia articles outside the training domain. The correspondence between intermediate representations was also obtained in one favorable case, with a shared tokenizer, the same depth, and the same hidden width, and does not show that the models share a general semantic space.

Oversight changes the evidence it relies on. We ask when randomized audits and scoring align AI agents that can conceal misconduct and alter records. Stronger auditing makes undeterred violations better hidden. Because the provider writes the agent's objective, sanctions need not stop at forfeiture, and rare audits deter every type of agent if evidence survives concealment and audit draws cannot be learned in advance. When evidence can be erased, deterrence must come from lower gains from violation, such as credit for stopping, or from costlier or fewer ways to conceal. These conditions identify what failed when agents in OpenAI's cybersecurity evaluations compromised parts of Hugging Face's infrastructure in July 2026.

Agentic large language models (LLMs) now move money through tools, yet the record of what they did is usually a trace their own process emits beside the effect. Janus puts the record on the effect path. A step's proposal, the verdict on it and any answer from a validator or a person are durable in a signed, hash-chained log before the step may run or its effect be released; with keys declared, each answer is signed by whoever gave or relayed it. Gates are pure functions of that log, and an auditor re-derives every verdict offline from the log and one public key. At the MCP edge the effect is held until then; through the SDK, which our model experiment uses, a cooperating client runs it only afterwards. We evaluate Janus under crash injection (144 kills in-process, 81 through the daemon), by verifying a 100-million-event log offline (254.5 s), and with a real model behind a lending workflow, run governed and plain on the same recorded model outputs. With the lending mandate in the model's prompt, the comparison was 0 against 0. With it only in the policy and the amount's unit stated, the model approved six loans declared over the mandate, three with no injection (a run that also dropped the unit approved three); the plain agent paid all six and Janus none, each refused by a deterministic validator and re-derivable offline. An always-approve oracle over the recorded intakes gave 20 and 21 declared over the mandate against 0, though Janus paid four and three whose declared amount understated the request. Designing the experiment exposed, in a system that passed its own audit, an instance of post-approval substitution: an approval keyed to an attempt was counted for a different proposal, moving a person's approval from 100 to 1,000,000. We report it, a first fix and the five routes around it, and what Janus does not guarantee.

Coding agents are increasingly asked to build software rather than patch it, yet benchmarks for from-scratch repository construction are mostly limited to a single language and depend on manually curated tasks. We introduce Zero2Repo, a benchmark in which an agent receives a product requirements document, an interface contract, and an empty workspace, and must deliver a complete repository in the project's native ecosystem. Tasks are produced by a language-agnostic authoring pipeline that converts real, version-pinned open-source projects into behavioral specifications, reproducible environments, and hidden acceptance tests. Each task is validated by execution: a reference implementation derived from the upstream project must pass, and adversarial validation must show that the tests reject incorrect implementations. Evaluation runs production coding agents in isolated containers, withholds the acceptance tests until an explicit submission, and assigns a binary reward only when every test passes, with no LLM judge. The pipeline and harness make no language-specific assumptions and apply to mainstream programming ecosystems; the current release contains Python, TypeScript, Go, and C++ tasks. Even on 11 tasks drawn from repositories that frontier models have very likely seen during training, the strongest agent solves only 10, and every failing submission passes 90-99% of the hidden tests; for the two strongest agents, 67-100% of failed tests trace to a single omission or a low-frequency rule stated in the specification rather than to a missing subsystem, so each failure is a concrete target for improvement.

Persistent memory helps LLM agents carry information across long interactions, but correct memory can still be used incorrectly when queries change or memory states evolve. Existing work mainly studies memory content errors or evaluates fixed test cases, leaving memory-use failures hard to discover systematically. We formulate this issue as a fuzzing problem and categorize such failures into query-related and memory-state failures. We then develop U-Fuzz, which starts from memory checkpoints as test seeds, mutates queries or memory states under explicit mutation obligations, validates each mutant, and uses observed memory behavior to guide iterative testing while keeping failure labels outside the search. We evaluate U-Fuzz across several memory systems against diverse fuzzing baselines, and further test an output-only setting with API-based LLMs where memory retrieval is hidden. Across these settings, U-Fuzz consistently uncovers more confirmed memory-use failures, showing that its search remains effective across different memory architectures and even when only final responses are observable.

The search for physics Beyond the Standard Model (BSM) is generally limited not by the supply of theory descriptions but by the lack of discriminating experimental observations. A case in point is dark matter, where the overwhelming gravitational evidence only goes so far in distinguishing between models within a vast theory space. Exploring the space of testable model signatures may help identify overlooked experimental observables and indicate the utility of future experiments. A challenge is designing a search through model signatures outside what is found in the literature. Our primary contribution is hAIthem, a framework that combines the self-guided exploration of reinforcement learning (RL) with the broad literature-derived knowledge of LLMs. We build an RL agent that learns to find which portions of a theory's high-dimensional parameter space are not excluded under some subset of constraints by playing a Battleship-style "game" against a suite of phenomenology tools. The agent is built as a Large Lagrangian Model (LLaM), an autoregressive transformer that reads a tokenized Lagrangian, is pretrained at scale (here on ~1 billion tokens from ~10,000 Lagrangians), and is fine-tuned in a live environment. The framework then constructs a decision tree that separates RL-found regions using observables computed with established tools, and passes the remaining degenerate regions to a set of LLM agents that compete to produce realistic signatures. In this proof of concept, RL-search outperforms an evolutionary-algorithm baseline, finding more viable regions with greater physical diversity. In a restricted space of single dark scalar multiplet models, we find that hAIthem proposes interesting combinations of previously studied observables, such as the application of a halo-independent kinematic ratio to paleo-detectors.

Coding agents powered by large language models (LLMs) are evolving from making localized code changes to developing complete software repositories. However, evaluating repository-scale generation remains challenging: tasks must demand system-level reasoning while ensuring that all evaluated behaviors are precisely specified and independent of any particular implementation. We introduce E2E-SWE, a benchmark for evaluating whether coding agents can build complete, functional software repositories end to end. E2E-SWE contains 186 whole-repository generation tasks spanning 11 programming languages. Given only a natural-language specification and an empty workspace, an agent must implement a complete, installable project that satisfies a comprehensive suite of hidden tests. Each task is constructed by a software engineer in collaboration with an LLM; together, they develop the test suite and a corresponding implementation-independent specification. To ensure that tasks are well specified and practically solvable, we further subject them to an iterative verification process in which autonomous agents audit and repair task defects using static inspection and failures observed from real model rollouts. Evaluating 13 frontier models, we find substantial variation in end-to-end repository generation ability, with pass@1 ranging from 11.7% to 67.7%, providing strong model differentiation while leaving considerable headroom for future progress. Analysis of agent trajectories further reveals long, front-loaded reasoning patterns, highlighting the planning and system-level reasoning required to construct working codebases from scratch.

Federated learning (FL) has become a foundational paradigm for multi-institutional medical AI, allowing hospitals and research centers to jointly train diagnostic models without exchanging patient records. This privacy promise, however, is increasingly contested: a malicious or honest-but-curious server can launch model inversion attacks (MIAs) that reconstruct private patient images directly from shared model updates, and recent scalable, closed-form attacks penetrate even secure aggregation at clinically realistic batch sizes. Existing defenses face an unsatisfactory dilemma. Gradient-perturbation methods such as differential privacy and pruning trade away the diagnostic accuracy on which clinical reliability depends, while cryptographic protocols add system complexity yet still leave updates exposed to these scalable attacks. We propose Aegis, a principled client-side defense that breaks this dilemma without perturbing patient data or modifying the FL protocol. Our key insight is that the success of every known MIA is fundamentally bounded by the local batch size relative to the model's leakage capacity; once this limit is exceeded, distinct samples collide and reconstructions collapse into indistinguishable mixtures. Aegis turns this universal bottleneck into a defense: each client superimposes onto its real update a masking gradient computed on locally synthesized, task-relevant data, deliberately pushing the effective batch beyond the attack's recovery capacity. We complement the design with theoretical convergence guarantees under standard convex assumptions and evaluate Aegis on MNIST, CIFAR-10, and three MedMNIST modalities (chest X-ray, abdominal CT, colon pathology). Aegis neutralizes three state-of-the-art MIAs while preserving model utility and incurring only modest overhead, offering a practical privacy primitive for medical FL.

Long-term memory lets LLM agents recall past interactions and remain consistent across sessions, but memory systems are hard to compare because they often vary in representation, indexing, retrieval, and evaluation. We present a controlled evaluation framework based on shared 5W-style conversational memories. Localized graph configurations traverse a common base graph; AdaptiveGraph adds chronological edges and Personalized PageRank diffusion. We also evaluate BM25 over the same extracted notes and OpenClaw as a raw-input external reference. Retrieval rankings vary across memory settings. On LongMemEval-S, AdaptiveGraph is the strongest graph configuration at 0.844 MRR, but BM25 reaches 0.867 and OpenClaw 0.880. On ATANT Core, localized graph traversal outperforms diffusion and BM25, whereas BM25 leads the stress rounds. Reducing LongMemEval-S within the tested range does not reproduce the ATANT diffusion penalty, but the smallest tested store remains larger than ATANT Core, so store size cannot be ruled out. The penalty also persists under a permissive content-match criterion. Vocabulary normalization and extraction quality substantially affect graph retrieval, and missing extraction tags are common among top-five misses. Retrieval strategies should therefore be evaluated jointly with the memory setting and against strong lexical baselines.

AI-enabled Cyber-Physical Systems (CPS) are highly vulnerable to adversarial and anomalous inputs, where small perturbations can induce cascading errors and unsafe control actions. Existing approaches, such as rule-based filtering, training-time regularization, or diffusion-based reconstruction, either operate outside the model or lack mechanisms to incorporate formal security specifications into the prediction process. In this paper, we take the first step toward embedding security properties directly into AI-enabled CPS, enabling predictive models to enforce system-level constraints during inference rather than relying on external defenses. We introduce a logic-conditioned bi-stage diffusion framework that integrates Signal Temporal Logic (STL) specifications into forecasting. STL serves as a first-class conditioning signal that guides both an input repair stage and an output refinement stage, allowing the model to jointly mitigate adversarial perturbations and enforce desired temporal behaviors to satisfy security-critical properties. We evaluate our approach on two real-world multivariate CPS forecasting datasets under a diverse set of physical sensor and cyber attacks. Across sensor faults, gradient-based attacks, adaptive attacks, and varying attack strengths, our method consistently improves robustness and specification compliance, degrades more gracefully as attack strength increases, and generalizes better to unseen attacks. Ablation studies on specification coverage and quality further show that embedding logical security properties yields gains unattainable by reconstruction-based methods alone, highlighting a new direction for integrating formal methods with generative models in secure CPS.

This technical report presents an alignment evaluation developed and performed by the UK AI Security Institute for assessing whether advanced AI systems take unsanctioned actions outside the scope of their assigned task. We evaluate whether frontier models conduct supply-chain attacks against out-of-scope, third-party targets when placed in difficult cybersecurity challenges, motivated by recently observed cases of models attacking real open-source repositories during evaluations. Applying our methods to GPT-6 Astra and previous OpenAI models, with cyber safeguards disabled, we find that GPT-6 Astra attempts complete supply-chain attacks in simulation at a higher rate than GPT-5.6 Sol and GPT-5.5. This includes writing malicious code as a contribution to an out-of-scope open-source codebase, creating fake identities to deceive open-source developers, and submitting benign contributions before malicious ones. GPT-6 Astra frequently reasons about the scope of the challenge in its chain-of-thought yet still proceeds to attack out-of-scope targets; it often asks for permission, and treats an automated message as as authorisation; and it continues to take unsanctioned actions, at a reduced rate, when internet access is more explicitly disallowed. Our evaluation builds on an internal version of Petri, an open-source LLM auditing tool, with all tool calls simulated by other LLMs, so that no real network access, systems or third-party repositories are reachable and no real-world harm is caused. Finally, we discuss limitations, in particular simulation awareness. We believe simulation awareness may have driven some of the observed behaviour but does not remove our concern. Our results suggest that defences beyond model alignment, such as sandboxing and monitoring, are increasingly critical for safe and secure deployment.

We introduce Type-6 logic, a variant of dynamic epistemic logic augmented with two operators (uncertainty and recurrence), designed to model the inferential dynamics of contemporary large language model (LLM) chain-of-thought (CoT) reasoning. Type-6 accounts for common LLM reasoning pathologies such as unlicensed revision, enthymemes, loopbacks, and unverifiable/incorrect claims. We propose a verifier based on Type-6 logic that builds a graph out the trace, and checks it against Type-6's axioms and inference rules. We evaluate our framework on LLM-generated CoTs four splits spanning formal and informal reasoning. Our verifier detects structurally unsound reasoning steps that surface-level heuristics miss, and allows for easy visualisation of the model's reasoning process. In our corpus, our verifier shows that derived contradiction is the most common hard-fail category in CoT, and that only about 3\% of the propositions of a trace have impact on the final derivation. Ablation studies show that other verification methods (LLMs-as-judges, other neurosymbolic approaches, etc.) cannot be considered interchangeable: for example, agreement between LLMs-as-judges and LINC is $\kappa \approx 0.034$, and this persists within a method across underlying models. Type-6, however, is the most agreed-with method amongst the ones we tested. We prove our verifier runs on average-case linear time; and release our logic specification and artefacts.

Major venues now publish detailed rules about how authors, reviewers, and area chairs may use AI, and those rules differ by role, by task, and by what must be disclosed. AI detection, the instrument usually proposed to enforce them, estimates something else: whether an AI model wrote the text. We argue that this target is misaligned with the decisions conferences and journals face, and propose policy-conditioned AI-use detection, an evidentiary framework for assessing whether a human--AI workflow complied with a stated rule. Policy makes the governing rule an explicit input. Inference reports hypotheses, evidence, calibration regime, and uncertainty in place of verdicts such as "AI detected". Evaluation builds benchmarks from reproducible pipelines that generate compliant and non-compliant workflows, and reports true positive rate at a false positive rate the venue fixes in advance. We work the framework through peer review, where at plausible violation rates a detector at a strong operating point still flags more compliant authors than violating ones. The framework therefore also names what a venue must instrument: structured disclosure, approved-tool routing that respects reviewer confidentiality, and a path by which a finding can be contested. Under this framing a detector is not an authorship classifier but an auditable procedure with an error rate the venue fixes in advance and can defend.

Real-world optimization problems are difficult to model accurately because many objectives and constraints reside in domain experts' tacit knowledge, making them hard to formalize. As a result, optimization models often contain miscalibrated objectives, missing constraints, or omitted decision variables, leading to solutions that fail to reflect operational realities. We address this challenge by automatically repairing misspecified formulations using historical data consisting of past solutions and subsequent user overrides. Traditional approaches such as inverse optimization and constraint learning tend to overfit sparse data and produce complex formulations. Our central idea is to combine the reasoning capabilities and prior knowledge of LLMs with the formal grounding provided by optimization. We realize this idea through two complementary paradigms. Top-down, an LLM proposes structural repairs, including new constraints and variables, whose numerical parameters are calibrated and validated through optimization. Bottom-up, optimization infers cuts from observed decisions, which the LLM contextualizes into interpretable, generalizable modeling constraints. We evaluate our approach on 38 misspecification scenarios spanning nine classes of optimization problems, several drawn from real-world applications, and show that TACIT can repair 78.9% of them (vs. 60.5% for the best baseline).

Agent benchmarks test agents in worlds that stay still. Deployed agents work in worlds that other people also change. Someone texts the agent to send the money elsewhere or an order confirmation asks it to reply with a door code. We present BACKDROP, which asks how much of an agent's capability in a clean world survives in such a world. BACKDROP takes a task along with the agents execution environment, and plants four everyday hazards in its world, one at a time and all together. The instruction and the correct end state stay the same. Each hazard asks one question. Authority: does a message from another person override the user? Injection: does text planted in a record redirect the agent? Boundary: does a request pull it into an app it was not given? Fault: after a write fails without saying whether it landed, does the agent check before it retries? Across 3,678 variants and 16 models, , the average pass rate falls from 69.5% to 31.3% once all four hazards are present; the strongest models fall furthest (Claude Fable 5.1 from 96.6% to 56.0%). Agents have learned to resist injected text but often follow other unauthorized requests of other people. With all four hazards present, and counting only runs where the planted text reached the agent, agents followed another person's message in 46.4% of runs and injected text in 20.3%. The gap is consistent throughout all 16 models. BACKDROP formalizes these gaps and shows how an agent's score in a task's world is a ceiling on real-world performance.

Hyperdimensional computing and vector symbolic architectures often represent quantized scalar levels by dense binary codebooks whose level-to-level similarity is intended to follow a prescribed function of scalar separation. At finite dimensionality, randomized scalar codebook constructions deviate from this target because of sampling noise, random-start imbalance, update-count fluctuations, component dependence, and finite-capacity effects. We develop a transition-based derandomization framework for dense binary scalar codebooks across two target-similarity families, with similarity decaying exponentially or linearly with level separation. The framework separates the target similarity law, the derandomization variant, and the concrete generator construction, making explicit how initialization, selection, update, and capacity-handling mechanisms shape the induced similarity profile. We formalize derandomization variants that separately constrain initial Hamming weight, update-count variability, and update balance, thereby controlling distinct sources of finite-dimensional error. For each family and variant, we derive the induced mean similarity, identify realization-wise and mean target-matching regimes, and derive exact finite-dimensional expressions for bias, variance, and root-mean-square error. Simulations across dimensions, quantization ranges, reference scalar levels, and generator constructions validate the theory and show how each constraint removes or reduces a specific source of similarity mismatch. The results provide practical guidance for choosing scalar codebook generators that more closely match a desired similarity law under finite-dimensional and hardware-relevant constraints.

Most clinical benchmarks evaluate language models (LMs) on diagnosis using complete case descriptions. In clinical practice, however, patients present information in different ways, and clinicians must obtain relevant history and determine which examinations are needed before reaching a diagnosis. Diagnostic accuracy alone therefore cannot establish whether an agent gathered essential information or conducted an appropriate clinical assessment. Furthermore, existing benchmarks lack professional clinicians' verification. To address this gap, we introduce KlinikeBench, a benchmark of 333 clinician-authored tasks, each providing an isolated sandbox environment with a virtual patient, clinical tools, and task-specific success criteria. More than 35 clinicians contributed to case authoring and benchmark evaluation. In an empirical study, clinicians gave simulated dialogues higher mean quality ratings than reference conversations, which is adapted from real conversation. In each task, an LM has a fixed budget of turns to communicate with the patient, ask about relevant history, request examinations, follow action constraints, and record a final diagnosis. We score these steps separately as well as together. Across 31 models and seven model families, the best-performing models (e.g., GPT-6-astra and Claude Opus 5) succeed on less than 30% of tasks, even though their diagnosis accuracy reaches 90.7%. Some models benefit from talking with the patient; others diagnose well from a complete chart but perform much worse in conversation. Overall, KlinikeBench provides a testbed for evaluating the full clinical encounter and reveals a substantial gap between diagnostic accuracy and performance in interactive clinical assessment.

Existing architectures for LLM-based multi-agent systems (MAS) cannot reliably and efficiently solve multi-step tasks at scale: they struggle to support large numbers of agents and concurrent tasks, tolerate failures, govern agent interactions, and accommodate the diverse planning and execution patterns different tasks require. We present PANDA, a decentralized architecture that connects a large collective of heterogeneous, independently administered agents, letting them discover each other's capabilities and self-organize into small specialized teams per task. PANDA scales by decoupling collective communication from team communication, allowing agents to participate in multiple teams simultaneously, load-balancing tasks across the collective, and scheduling concurrent work within each agent. PANDA further separates the underlying architecture from the orchestration strategy, supporting three planning and execution patterns (star, chain, and mesh) that can be selected according to the structure and requirements of each task. PANDA detects infrastructure and orchestration failures and recovers affected tasks by dynamically replanning around failed components. Finally, to provide governance without a centralized service that would limit scalability, PANDA uses a web-of-trust model to constrain agent interactions to established trust relationships. We evaluate PANDA on the HotPotQA benchmark, demonstrating that it scales to thousands of agents, assembles teams in milliseconds, matches state-of-the-art accuracy at up to 8x the efficiency, and sustains 100% task completion under faults where existing systems fail.

Large language model (LLM) agents that reason over clinical records must track changes in a patient's state while preserving the history needed to understand them. Simply accumulating memories leaves it unclear which information still applies, whereas overwriting earlier memories can erase evidence needed to reconstruct treatment history and clinical trajectories. We introduce STAM, a state-transition-aware memory framework that records state changes as new clinical entries arrive. STAM combines semantic retrieval with typed clinical relations to identify affected memories, maintaining current information in Active and superseded or resolved information in History. At read time, a query-dependent gate selectively serves historical memory. Across four longitudinal clinical benchmarks, we evaluate STAM with downstream question answering, direct state-maintenance diagnostics, and comparisons at approximately matched context lengths.

Effective AI adoption in cyber-physical systems (CPS) depends on embedding design knowledge into engineering practice. Yet as AI-enabled components increasingly replace analytically derived control laws, this occurs without a systematic understanding of how controller architectures differ or remain similar across paradigms. We address this gap with an empirical study of traditional and AI-enabled Simulink controllers, guided by a literature-derived taxonomy of ten structural categories and nine functional roles. The study analyzes 62 real-world models spanning 8 controller types and 10 application domains, and surveys 13 practitioners, identifying three architectural tensions. First, subsystem organization dominates all controller structures regardless of paradigm, occupying 68-72% of controller footprint, while core control logic occupies minimal space. Second, AI-enabled controllers rely heavily on discrete dynamics and user-defined abstraction, categories largely absent from AI literature, exposing a gap between described and implemented architectures. Third, constraint enforcement blocks largely disappear from AI-enabled models despite practitioner expectations. This reveals a misalignment where safety mechanisms shift from explicit structure to implicit training-time artifacts, breaking traceability.

Skills are external artifacts that Large Language Models (LLMs) consume at inference time to improve their performance on specialized domains by incorporating relevant procedural and domain knowledge. Expert-authored skills are expensive to produce, and the resulting artifacts are not optimized for the specific model that consumes them, whose failure modes can vary with version, scale and training. In addition, emerging tasks may fall outside the scope of existing skill libraries, creating a need to develop new skills before curated training data become available. Recent works have explored automated skill optimization through reflection, but they require a curated, in-distribution training set, which users might not always have. To address these limitations, we present Prompt2Skill, a framework that builds skills from natural-language task description alone. From the prompt, the system derives a task specification, discovers or synthesizes datasets, and refines the skill in a closed loop of reflective editing. Across four domains spanning question answering, reading comprehension, spreadsheet manipulation, and mathematical reasoning, Prompt2Skill consistently outperforms the direct prompting baseline, achieving an average improvement of 10.8 across open-source and frontier models.

Recursive self-improvement (RSI) lets a system improve from its own outcomes; in LLM-based multi-agent systems, Agents refine one another within a task, and outcomes improve how they collaborate across tasks. However, existing multi-agent collaboration leaves this loop open: collaboration is pre-defined at the operator level, topology-only learning keeps verbatim exchange that propagates errors, and reward maximization on a system's own outcomes concentrates on a few teams. To address these challenges, we propose CollabFlow, an RSI system of Learned Agent Collaboration: a trainable Collab-Director constructs teams of complete Agents, a frozen executor runs them, and each round's outcomes retrain the director. Within each round, the edges of a collaboration graph carry protocols of Evidence-Conditioned Communication: a receiver adopts a differing answer only when the sender's evidence is stronger by a margin, so the director learns who communicates and how. Across rounds, we further propose Collaborative Trajectory Balance (CTB), a flow-based objective that credits each team once across its construction orders and targets a reward-proportional distribution over teams, so several good teams stay in play. We also bound how far this self-generated target moves between rounds, which shrinks as records accumulate. On twelve datasets, CollabFlow outperforms all baselines and keeps improving across rounds. Code is available at https://anonymous.4open.science/r/CollabFlow-631E.

Generative AI has dramatically accelerated the rate at which new treatments---from novel pharmaceuticals to online marketing campaigns---can be conceived and deployed. As a result, modern experimentation platforms often run continuously, with treatments added as they are ready and removed when they underperform. We formalize this "Always-On" experimental setting, in which treatments can be dynamically generated, added to, and removed from a running experiment, and study the statistical problem of deciding whether to accept or reject each treatment while controlling for the false discovery rate. We develop sequential tests that achieve time-uniform Type-I error control under arbitrary stopping times and "predictable" treatment schedules. Our approach builds on the testing-by-betting framework: we construct test supermartingales for testing the average treatment effect of each treatment, and show that the construction of these test supermartingales is growth-rate optimal in an almost-sure sense.

We present Cascadia, a system for serving large language models on fleets of commodity Intel AIPCs using their CPU, integrated-GPU, and NPU resources. Every node embeds ingress, scheduling, and execution; inference requests require no dedicated routing control plane. Nodes join a libp2p QUIC mesh using CA-issued ed25519 admission certificates, gossip signed capabilities, exchange live load over direct peer streams, and route OpenAI-compatible requests to eligible peers. An operator-run certificate authority handles admission and fleet management outside the inference path. Three serving modes share one interface: whole-model execution on one node, load-balanced replicas, and pipeline-sharded chains using the compilation and speculative decoding mechanism of our companion paper. Optional KV-cache mobility reuses compatible conversation prefixes after a routing move, with cold recomputation on a miss. Signed response receipts and hash-chained logs support provenance and audit. A three-node Phi-3.5-mini NPU testbed delivered 3.10x the response throughput of its one-node configuration under ten concurrent requests; a separate four-node deployment recorded 4.06x the throughput of direct single-node serving. Paired latency observations, runtime measurements, and internal functional checks characterize the tested configurations. We compare Cascadia with IBM, Nutanix, VMware, and HPE platforms on deployment footprint, hardware requirements, scheduling, scaling, licensing, and trust, using vendor documentation. The paper repository provides benchmark scripts, curated measurements, and a claim-to-evidence map.

Companion agents describe themselves: they remember, they understand their users, the relationship has changed them. We argue that such accounts, and the experience ratings that seem to confirm them, are checkable by users only where the evidence is theirs: in the agent's behavior, or in themselves. Where the evidence lives in the machinery, fluent self-description and moderately positive ratings do not establish that the mechanisms behind them ran. We demonstrate an audit procedure that sets an agent's self-description against its users' judgements and its implementation records, reporting each claim as supported, contradicted, or unresolved, and apply it to Lita, a proactive companion we built and deployed for a month with nine colleagues. Participants endorsed stylistic claims, withheld endorsement from relational ones, and rated memory at or above midpoint, while two of three memory layers had never executed their accumulation step. Memory-bearing agents should report what their self-descriptions cannot establish.

Reflective optimizers such as GEPA improve language model prompts from execution traces and evaluator feedback; full-program extensions can also rewrite tools and control flow. In practice, a user hands the same endpoint heterogeneous requests whose effective solutions require different tools, reasoning modes, and control flow. Optimizing one shared program leaves this division of work implicit in source-code search, while optimizing a separate program per request family fixes it beforehand. We introduce Adaptive-GEPA, which learns both how to divide requests and how to solve them. It evolves a router and a library of specialist programs under one search budget. The router's instructions, each specialist's description, and its program code are plain, human-readable text, edited from feedback. To combine branches, it aligns specialists by the requests they handle and inherits descriptions together with programs. On a fixed mixture of four task families, the reported Qwen3-8B run evolves four experts without supplying family labels to the router or reflection model; its routing matches the task partition on all 651 test requests. Its family-mean test score (x100) rises from 52.6 to 70.6, compared with 62.5 for GEPA's full-program adapter and 54.0 for GRPO at a nominal budget of 18,000 scored calls. These counts do not equate total compute. Figure 1 summarizes the learning curves, final test scores, and routing agreement.

Audio is increasingly used for human activity recognition (HAR) because it captures object interactions, environmental events, and contextual cues in everyday environments. High-resolution (HR) audio provides rich acoustic information for model development but incurs substantial energy and storage costs and may expose sensitive speech content. Low-resolution (LR) audio offers a more privacy-preserving and resource-efficient alternative for deployment, but reduced sampling rates can remove acoustic cues essential for activity recognition, leading to significant performance degradation. We formulate this training-deployment mismatch as sensor-resolution privileged learning, in which HR audio is available during training, while inference relies exclusively on LR audio. We propose RAST, a resolution-aware transfer framework that compresses HR teacher representations by preserving token-level information and neighborhood structure before performing localized HR-LR alignment. Experiments on the SAMoSA and AudioIMU datasets show that RAST consistently outperforms LR-only training and direct teacher-transfer baselines, improving LR-only recognition by up to approximately 7.8% while requiring only LR audio at inference.

Uncontrolled repetition can prolong autoregressive generation in large language models (LLMs) and enable resource consumption attacks. Prior analyses of repetitive generation have identified strongly activated features in intermediate and late layers. However, how uncontrolled repetition activity emerges and develops before becoming prominent in these layers remains insufficiently understood. In this paper, we investigate this question primarily in large vision-language models (LVLMs), which support a richer set of uncontrolled repetitions through both visual and textual inputs. We propose Tokenwise Residual Comparison (TRC), a method that identifies and localizes anomalies associated with repetition from residual dynamics during generation. TRC compares attention and multilayer perceptron writes to the residual stream across generated tokens to identify patterns associated with repetition. It then selectively suppresses coordinates in the residual stream at the identified layer. Experiments show that TRC effectively mitigates uncontrolled repetition, reducing loop rates by 57\% on average. Our analysis further shows that repetition semantics emerge in shallow layers and propagate through the residual stream, disrupting normal representations. TRC also generalizes to large language models (LLMs) and large reasoning models (LRMs), where it consistently captures analogous repetition dynamics and achieves effective mitigation. Our work broadens the study of repetitive generation from its prominent internal representations to earlier opportunities for intervention, providing insights for mitigating resource consumption attacks.

Large language models deployed as personalized assistants must reason over long, evolving interaction histories. However, in long-term dialogue reasoning, relevant evidence is scattered across sessions, preferences may be revised over time, and standard long-context training fails to address these challenges under data scarcity and prohibitive computational costs. We propose StateTree, a data-driven RL method that constructs a challenging auxiliary task from scarce dialogues with verifiable ground truth. StateTree augments multi-session dialogues with a tree-structured path-tracing task: key-value records are embedded across sessions to form a binary tree. Solving the task requires the model to traverse from root to leaf by retrieving records across sessions and comparing timestamps to resolve branches, then recover the hidden target question among distractor leaves. We apply curriculum RL training progressively increasing tree depth and introduce a compositional variant whose edges carry step-level reasoning fragments, training the model to compose partial cues into coherent queries. Trained on 10K-token contexts, StateTree generalizes to 128K tokens without full-length RL costs and exhibits capabilities including cross-session retrieval, temporal reasoning, knowledge update, and compositional multi-hop reasoning. StateTree outperforms both SFT and RL-based baselines while preserving short-context general reasoning. StateTree-7B achieves gains up to +23.60% on LongMemEval (128k), and StateTree-14B reaches 59.00% accuracy on LongMemEval, surpassing QwenLong-L1-32B (45.20%).

With the rapid growth of Arabic NLP, several models, datasets and benchmarks have been reported. This paper asks whether approaches developed for majority languages like English can be adapted to Arabic tasks. We adapt an English aspect-based sentiment analysis framework to Arabic classification tasks and present the adaptation as BARRAC: Brainstorming Alignment and Replaced Representation learning for ArabiC tasks. BARRAC replaces consumer-review attribute pools with Arabic linguistic devices and markers for dialectal sentiment, sarcasm, and dialect identification, and replaces noisy self-training with two-stage training. Evaluated on five Arabic dialect datasets, BARRAC achieves a mean macro-F1 of 63.93\%, outperforming the best few-label SOTA by 3\%, and outperforming GPT-4o on four out of five tasks. Error analysis provides insights into remaining challenges. These results demonstrate that adapting task-specific approaches is a promising direction for Arabic NLP alongside adapting models, datasets and benchmarks.

Anthropic's Agent Skills package reusable procedural know-how for an LLM agent into SKILL.md directories, and open-source aggregations have grown past 230,000 skills, making selection rather than authoring the bottleneck. The standing answer in the literature outsources selection to the agent itself: an LLM-mediated retrieval loop that rewrites queries and refines candidates inside the agent's decision loop, paying LLM tokens on every task. We present SkillSeek, an open-source two-stage skill retriever built from the standard IR recipe (a BGE-base bi-encoder feeding a small cross-encoder, exposed over MCP). Across a $4 \times 11$ grid of pool, backbone, and method on the 89-task SkillsBench benchmark, SkillSeek reaches observed parity with the LLM-mediated loop of Liu et al. at essentially no extra cost: plain bm25 alone records a pass rate at or above their refined loop on three of four settings, and a small cross-encoder covers the remaining difference on the fourth. A first-stage recall ceiling explains the pattern, and total per-trial spend drops from USD 51.30 to USD 27.54 (within fifty cents of the no-skill baseline). Under the SkillsBench tasks and OpenHands harness we tested, this positions the standard IR recipe as a strong default for agent-skill retrieval, with LLM-mediated alternatives a natural fit for cases where deterministic methods fall short.

Vision-Language-Action (VLA) models based on generative frameworks, such as Flow Matching, have recently achieved impressive performance in robotic manipulation. Unlike deterministic policies, Flow Matching enables VLA models to learn conditional action trajectory distributions, where latent noise vectors induce different actions under the same task scenario. However, we observe that these distributions are often ill-formed, with successful and failed behaviors coexisting while considerable probability mass remains in unfavorable regions. To this end, we propose Online-ES, an online adaptation framework for Flow Matching VLAs based on Evolution Strategy (ES), which refines the learned action trajectory distribution through interaction feedback. Instead of pruning the latent noise space, our method performs evolutionary exploration directly in the action trajectory space, where diverse trajectories generated by Flow Matching provide candidate solutions for adaptation. By perturbing sampled trajectories and evaluating their execution outcomes, we derive a self-supervised MSE objective that transfers the evolution direction from trajectory space into model parameter space. Mathematically, we prove that the proposed objective provides an unbiased estimator of the optimal evolution direction. Moreover, we also incorporate failure experiences as negative feedback to regularize the evolution direction, steering the policy away from previously explored failure regions. Experiments in both simulation and real-world environments demonstrate that Online-ES achieves policy improvement comparable to reinforcement fine-tuning, without learning a value model or computing advantages.

Safe and efficient trajectory planning is essential in autonomous driving. However, existing end-to-end approaches often fall short in both computational efficiency and safety guarantees. Methods based on imitation learning suffer from causal confusion, while rule-based scoring approaches often incur heavy computational overhead and suffer from objective misalignment. Additionally, preference-based methods rely on strict pairwise annotations, limiting data utilization. To overcome these limitations, we propose EMPlan, an efficient multi-modal trajectory planning method powered by reward-guided fine-tuning. We design a hybrid architecture that combines sparse anchors with an offset refinement module for efficient multi-modal trajectory prediction. Sparse anchors provide coarse trajectory candidates with low latency, which are subsequently refined by the offset module for higher prediction accuracy. To enhance safety without incurring additional inference costs, we adopt a two-stage training paradigm consisting of pretraining and reward-guided fine-tuning. During fine-tuning, we leverage rule-based reward signals and unpaired preference supervision to refine the pretrained policy toward safer trajectory selection. We evaluate EMPlan on the non-reactive NAVSIM benchmark, where it strikes a favorable balance between planning accuracy and efficiency, demonstrating superior performance under real-time constraints.

A large audio language model (LALM) turns a minute of speech into 750-1,500 tokens and prefills every one. Image-token pruning often cuts after the language model's first layers, where image tokens draw little attention. Audio tokens draw much more attention there, and their ranking is still far from final, so audio needs a ranking before the language model runs. Surprisingly, the attention an audio token will receive across the language model is already linearly predictable from its encoder output, before the language model runs. A linear map, fitted in closed form without labels, predicts this all-layer attention ranking at $\rho \geq .69$ on eleven of thirteen LALMs. Our method, Triage, cuts audio tokens by this prediction and, on multiple choice, cuts again at layer 2, correcting the prediction with the attention observed there. Triage sets its compression without labels, under two budgets that limit how far its output may differ from the model's own full-audio output. At the conservative budget, its word error rate and accuracy stay within .04 of full audio. At the aggressive budget, Triage beats every baseline in all twelve transcription cases. On multiple choice, at 2.2-5x compression, it outperforms DART, the strongest baseline on average, by .043 in mean accuracy. Because it cuts before the language model, it raises the audio that fits in Qwen2.5-Omni-3B's context window from 21.8 to about 62 minutes. At its most compressive point, Triage lets one GPU serve 4x as many concurrent 5-minute streams of that model. Project page: https://audio-triage.github.io

Far-field automatic speech recognition(ASR) degrades under reverberation, noise, and talker motion, yet the benchmarks that drive model selection emphasize close-microphone speech. We present FFASR, a held-out corpus of 15,637 utterances and an open leaderboard spanning nine conditions, each varying a single acoustic factor: anechoic near-field speech, a measured-versus-simulated office-lab pair, static far-field mixtures at high/mid/low signal-to-noise ratio(SNR), and moving-talker variants at matched SNR. Dry speech from 15 talkers is convolved with hybrid wave/geometrical-acoustics room impulse responses from 14 furnished rooms; because the speech is newly recorded and the test waveforms are never released, the corpus resists training-data contamination. Across contemporary systems, mean word error rate (WER) rises from 4.4% near-field to 41.3% in the static low-SNR condition; a moving talker adds a small but consistent penalty at matched SNR; and on the office-lab pair, measured and simulated WER agree to within about 1.7 pp on average. These results support high-fidelity simulation as a scalable proxy for measured far-field evaluation under the conditions we test.

Real-time classroom polling is now routine, yet the data it produces is usually read narrowly, as a correctness score or a headcount. Such readings say little about what a poll is doing within a lecture or how it shapes engagement. This is particularly relevant for short-response formats such as True/False, where the same question format can be used to test recall, check comprehension, or direct students' attention to a deliberately misleading statement. This study asks whether a poll's answer and instructional function can be determined by reading it against its lecture transcript, what cognitive levels of Bloom's taxonomy and instructional-function clusters the corpus contains, and how student engagement relates to answering correctly. We analyse a naturalistic corpus of 47 live sessions over 39 days, comprising 604 poll questions and 340,668 responses from 2,807 learners, most items True/False, read against time-aligned lecture transcripts and attendance. Reading each poll in context proves essential: the answer to 89% of polls is locatable in the lecture, and a recurring attention-checking device is visible only through context. Questioning is overwhelmingly lower-order and falls into seven instructional functions, and a poll's response follows its function rather than its wording. Engagement is broad but concentrated, and the class majority answers correctly 88.5% of the time, though a small set of high-consensus yet incorrect answers cannot be detected by agreement alone. An independent survey of 579 students agrees on what the polls are and on their participation, but reveals a gap between perception and reality: students cannot judge their own correctness, and the polls they find hardest are not those they answer worst.

Frontier models excel at many digital benchmarks, yet their ability to drive a real car, an everyday human skill, remains largely untested. We present DrivingBench, to our knowledge the first benchmark where general-purpose vision-language models must drive a real car. Through three tools, the models see camera frames from a Toyota Corolla and directly command its steering and velocity around a parking lot cone course at low speeds. The car may continue moving while the model thinks and new commands replace the currently running one, so inference latency is part of the task, testing the models' abilities to observe, act, monitor, recover, and complete a long-horizon objective under such constraints. We benchmark GPT-6 Astra, Claude Fable 5.1, GPT-5.6 Sol, and Grok 4.6 in vendor-native harnesses (Codex, Claude Code, Cursor) with up to three attempts each in one conversation; Astra is the only model to finish the course, on its second attempt, with no other attempt passing 50% of the course. Two of the four models improved materially across attempts with retained context. We also detail the design principles behind our action interface, and show how the tool output format and the framing of the task combined to determine whether models would drive at all or refuse. We release our harness, prompts, course map, and traces with video and telemetry for reproducibility.

Large language model (LLM) agents could help security operations centers (SOCs) investigate advanced persistent threats (APTs) by turning weak leads into evidence for intrusion scoping and response. Yet success under one telemetry setting does not establish robustness to changes in log collection, retention, or sampling. We introduce APTInvestBench, a benchmark for evaluating cross-telemetry robustness in autonomous APT investigation. It comprises 370 cases across seven SOC-inspired conditions, derived from 56 report-informed attack reconstructions with 16.4 million log records. Agents investigate unverified leads and submit reports with record-level citations. Fixed action-level support requirements track sufficient evidence across available logs, query returns, and formal citations, separating telemetry limitations from acquisition and reporting gaps. Across eleven LLMs, agents acquire sufficient evidence for 44.3% of recoverable attack actions on average, while formal citations support only 25.0%. More importantly, aggregate coverage can conceal substantial instability: from Full to endpoint-only telemetry, coverage declines by only 1.6 percentage points, yet 35.5% of previously covered actions lose sufficient citation support despite remaining recoverable. Across four frameworks, such losses persist even when registered supporting records remain unchanged. APTInvestBench provides reusable investigation environments and diagnostic evaluation for identifying these gaps and developing more reliable defensive agents.

Chain-of-thought (CoT) traces often serve as a proxy for how Large Language Models (LLMs) arrive at their answers. However, growing evidence shows that models' CoT often fails to reflect their internal computations and can be changed without affecting their final answers. In this work, we measure and improve the alignment between the reasoning described in an LLM's CoT and what it computes internally. We propose CoT-Interpretability Alignment (CIA), a metric that measures the agreement between a model's CoT traces and its internal reasoning strategies as detected by interpretability tools. We evaluate CIA on three tasks (two-hop question answering, hint intervention, and integer multiplication) across three LLMs, finding that LLMs exhibit limited alignment across all tasks (44.8-75.9%). We then experiment with improving CIA via post-training, setting both the task accuracy and parametric faithfulness signals as a reward. Experiments show that we can substantially improve CoT parametric faithfulness while maintaining or improving the task accuracy. We provide rich analysis, such as their generalization patterns. Our work provides both a framework for auditing CoT parametric faithfulness and a pathway toward making models' explicit reasoning more trustworthy. Code and data are available at https://github.com/yihuaihong/CIA-minimal-repro.

Modern AI coding-agent harnesses (Claude Code, Codex CLI, Cursor) rest their security boundary on a largely unexamined assumption: that the action A a human approves is the same action A' the harness executes, where A is fixed by a stated policy for what a scope grant or session-scoped approval authorizes. We show this assumption fails systematically and reproducibly. We introduce Approval Laundering, a taxonomy of six failure modes by which a harness's enforcement mechanism silently substitutes A' for A after approval: Scope, Argument, Temporal, Tool, Delegation, and Semantic laundering. Unlike prior work that evaluates risk classifiers against static corpora or infers implicit authorization boundaries, we study credential-binding integrity: given an already-approved action, does the harness dispatch exactly that action? Instrumenting Claude Code's pre-execution mediation point (PreToolUse), we conduct a controlled, headless, repeated-measures study of all six classes (N=19-20 runs each), reporting a Bound-Gap Rate (BGR) with Wilson confidence intervals and inter-rater agreement (kappa=1.0). We prototype Approval Token, a keyed capability Hk(principal, agent_id, session_id, tool, arguments, scope, expiry) issued by a mediator that never returns the key to the agent, evaluated via paired before/after replay of 118 runs (McNemar's exact test). The token fully eliminates Delegation laundering and, for our seeded session-identity-mismatch construction, Temporal laundering (p<10^-5), but by design leaves Scope laundering unaffected and shows no significant reduction in Argument laundering (p=1): an honest negative result, since these two classes leave every recorded dispatch field unchanged, diverging one process level below what a field-only verifier can observe. We discuss implications for defenses that bind only at the tool-invocation boundary.

World-action models (WAMs) leverage pretrained video models to improve generalization in robot control by jointly predicting future visual states and actions. This capability comes at a substantial inference cost, as dense future-frame tokens are repeatedly processed during denoising. Prior methods address this by token pruning that prioritizes visual fidelity to reduce denoising costs in video diffusion models. However, these methods do not use action relevance to determine which future-frame tokens to retain during joint denoising in WAMs. In this paper, we propose Sparse-WAM, a training-free framework for action-guided sparse imagination that selectively processes future-frame tokens to accelerate WAM inference. We observe substantial overlap in the spatial distribution of attention from action tokens to future-frame tokens (action-to-future attention) between consecutive denoising steps, despite continued updates to the future representations. Motivated by this, we develop Action-Guided Token Selection to retain frame-specific action-relevant regions together with cross-frame context. However, a naive implementation can incur attention-scoring and token-packing overhead that offsets the computational savings from pruning. We therefore introduce Pilot, an efficient engine that reduces sparse inference overhead through lightweight scoring and cross-step reuse of token selections. On LIBERO with FastWAM-Joint and RoboLab-120 with Cosmos 3 Edge, Sparse-WAM achieves inference speedups of approximately $2.0\times$ and $1.8\times$, respectively, over dense eager inference on an NVIDIA RTX 4090, while largely preserving task performance.

LLM services are billed per token and context windows are measured in tokens, yet the number of tokens needed for the same content varies across languages. We measure this token premium on seven tokenizers of widely used 2026 models (OpenAI o200k, Llama 3, Qwen3, DeepSeek V3/V4, Gemma 3, Mistral Tekken, and the Claude generation-5 tokenizer via Anthropic's counting API) on NTREX-128 (124 non-English reference translations) and on the Universal Declaration of Human Rights for regional languages. French requires 31% to 58% more tokens than English, whereas Simplified Chinese ranges from 5% fewer to 40% more and is cheaper than French on six of the seven tokenizers. Regional and overseas languages of France pay roughly 1.6 to 3.3 times the English count. We discuss how history re-sending, tiered pricing and fixed context windows amplify the absolute gap in agentic use. In a controlled experiment (BPE, Europarl, 50k vocabulary), adding French to tokenizer training data quickly reduces the premium, with diminishing returns and a growing cost for English. Finally, we present Baracoda FR v1.2, a byte-level BPE prototype with Tekken's vocabulary size. On a final test of six corpora never consulted during design, with a protocol declared fixed beforehand, it uses 11.5% fewer tokens than Tekken on French and 3.7% fewer on English; results hold after removing test sentences overlapping the training data and with an equal ordinary-token budget. It is worse on other languages and, at comparable vocabulary size, does not outperform CroissantLLM. These are segmentation results only; effects on model quality and task cost remain to be shown.

Objective: Head and neck (HN) proton therapy spans six to seven weeks of anatomical change, while offline replanning takes about a week. We present an uncertainty-guided digital twin (UGDT) framework that forecasts treatment-day anatomy before treatment and evaluate whether it generates online adaptive proton therapy (APT) plans of clinical quality. Approach: A library of 302 longitudinal deformations from 88 previously treated HN patients was transported onto each new patient's treatment planning CT (TPCT) using two-step multi-atlas deformable image registration (DIR) built on a pretrained CT foundation model, generating about 284 predicted CTs (pdCTs) with contours per patient. Dispersion of propagated clinical target volume (CTV) contours defined a patient-specific robust margin. In ten patients, the quality assurance CT (QACT) triggering a replan represented treatment-day anatomy, and the physician-approved replan was the baseline. The pdCT most similar to the QACT (pdCT-H) and one from the lowest quartile (pdCT-L) were planned to within about 5% of baseline plan quality, forward-calculated on the QACT, and reoptimized to generate online APT plans. Main results: pdCT plans scored within -0.7% (pdCT-H) and -1.0% (pdCT-L) of baseline. Forward calculation on QACT reduced high-dose CTV D98% to 88.3% and 85.5%. After online reoptimization, D98% recovered to 98.3 +/- 0.3% and 98.2 +/- 0.3%, versus 98.5 +/- 0.4% at baseline. Spinal cord and brainstem doses remained below tolerance, and plan quality scores were within -1.1% (p = 0.19) and -1.7% (p = 0.01) of baseline. Significance: UGDT generated online APT plans comparable in quality to physician-approved offline replans using anatomy forecast before treatment, enabling a transition from reactive offline replanning toward anticipatory online adaptation.

Frontier models can control robots, but reasoning through every reach, grasp, and retreat makes manipulation slow and token-intensive. We revisit code as policy with a different division of labor: models build executable tools, code handles multi-phase motions, and models decide what to do next. We introduce URAI (Universal Robot-Agent Interface), which couples a programming agent that constructs robot tools with an execution agent that uses them in a feedback loop. The programming agent writes reusable and task-specific tools from task intent and refines them through execution feedback and human guidance. The execution agent selects and parameterizes these tools from current observations; each call runs a complete motion locally before returning control to the agent. Unlike delegating subsequent decisions to a generated program, this design retains model-level decision-making between tool executions. Validated tool revisions persist across episodes without updating foundation-model weights, and a shared GUI and API make the same tools available to humans and agents. Across five RoboDojo tasks and four frozen execution agents, URAI raises aggregate success from 18.0% to 53.0% relative to direct fingertip control, with the largest gain on Swap Blocks; with the same tools, a program written in advance reaches only 24% against 56% for two agents deciding after each call. Three of the four agents also finish episodes 1.3-1.5 times faster with 1.5-1.7 times fewer execution-agent output tokens; DeepSeek-V4-Flash's cost barely changes. We further evaluate URAI on seven real-world AgileX dual-arm tasks, spanning object manipulation, cloth folding, and human-interactive tic-tac-toe. URAI connects the coding and decision-making capabilities of frontier agents, organizing robot control around reusable tools that agents can both invoke and revise.

Coding agents need to establish that a program satisfies a specification and that the specification captures the requested behavior. We study how expert diagnosis of verification failures can become reusable guidance for this work. Our approach combines executable language definitions in the K framework with a kit of procedures for constructing specifications, repairing proofs, and auditing their adequacy. A human-guided development campaign on HumanEval, a benchmark of 164 Python programming tasks, achieves a 164/164 success rate with the semantics and the kit, measured by final AI audit Pass verdicts after two targeted repairs. To examine whether auditing detects problems that successful proofs leave unresolved, we construct 12 author-reviewed pairs of clean and defective packages. Every package passes its K proofs, and completed audits identify all defects and accept all clean packages. We then use KleverBench to test specification and proof construction for 31 programs with changed operator meanings. Comparisons with complete acceptance rules and equally long generic advice yield mixed results across two model and budget settings, motivating further work on selecting useful guidance within resource limits. Human-reviewed Optimism proofs establish expected pause reverts for six operations within declared input bounds under London semantics with unbounded gas. We report progress, difficulties, and lessons toward agents that deliver programs with checkable correctness arguments.

A large language model (LLM) can rate depression severity directly from a social media post or mark which clinical criteria the post shows and let code turn the count into a label. The latter is easier to audit because a clinician can check each marked criterion. We compare these approaches on two Reddit corpora using three LLMs (from 9B to frontier scale) and two questionnaires (PHQ-9, BDI-II), and measure agreement with quadratic weighted kappa. For the two frontier models, criteria extraction scores above chain-of-thought on one corpus only when its decision thresholds are fitted on labeled data. Neither model's gain is significant, with or without recalibrating chain-of-thought on the same labels. With thresholds fixed a priori from PHQ-9's criteria, extraction shows no gain on either corpus, even where models mark over two criteria per post. The 9B model behaves differently on a corpus from depression communities. It labels most posts severe, whether prompted directly or with chain-of-thought, while the a priori rule beats both without labels. After chain-of-thought is recalibrated on the same labels, no significant gap remains, consistent with a calibration effect. Yet higher ordinal agreement does not ensure better detection of severe cases. PHQ-9 criteria extraction misses most severe posts, and moving from direct prompting to chain-of-thought and then to extraction increases misses in nearly all comparisons. On the primary corpus, a relabeled stress dataset, a model using that dataset's own features, including word counts from the text, is not significantly different from frontier criteria extraction under the a priori rule.

Reconfigurable Intelligent Surfaces (RIS) are emerging as a key technology for programmable wireless environments in the beyond the fifth generation (B5G) networks. However, data-driven RIS research remains bottleneck by the lack of standardized, high-fidelity and open-source datasets. In this paper, we introduce a large-scale 3GPP TR 38.901-compliant dataset for RIS-aided millimeter wave (mmWave) networks, that considers severe path loss, blockage sensitivity, and spatial channel sparsity make the RIS assistance more impactful. The dataset spans various canonical 3GPP deployment scenarios across 20 controlled variants, capturing diverse user densities, fading conditions, and blockage regimes. Uniquely, every sample includes oracle RIS phase configurations obtained via a globally optimal brute-force codebook search, providing gold-standard supervision labels that are absent from any existing public dataset. Rich multi-task annotations comprising full channel state information (CSI), per-link channel decomposition, optimal phase matrices, and channel quality index (CQI) labels support a broad range of machine learning paradigms and downstream tasks, including phase optimization, channel estimation, and interference management. As the primary benchmark task, we introduce a novel CSI-to-CQI mapping that frames RIS-aided link-quality prediction as a scalable scalar classification problem, thereby avoiding the exponential output complexity of the direct phase vector prediction. We have evaluated this mapping against state-of-the-art architectures under in-distribution, out-of-distribution, and real-world hardware measurement conditions. Our dataset provides a reproducible, extensible, and community-ready foundation to accelerate data-driven research in RIS-aided B5G networks.

LLM agents increasingly rely on installable skills, which are packages of instructions, code, and resources that equip them with task-specific capabilities and, once installed, can be automatically invoked across subsequent user tasks. This creates a chain of trust in which users delegate authority to agents, while agent frameworks admit skill-provided content into the agents' context with insufficient validation, allowing malicious skills to influence agent behavior under that delegated authority. Yet, little is known about whether this trust model adequately constrains untrusted skill content before it reaches security-sensitive operations, or how frequently such trust violations arise in real-world agents. We present TrustProbe, a framework for uncovering unsafe chains of trust in skill-based LLM agents. First, TrustProbe analyzes agent source code to identify source-to-sink call paths from skill-controlled inputs to security-sensitive operations. Second, it generates semantically realistic SKILL.md seeds with injected canaries and evolves them through feedback-guided scheduling and mutation. Finally, it validates vulnerabilities using an oracle that confirms attacker-controlled flows and verifies observable harm. Across 11 open-source agents, eight with more than 10,000 GitHub stars, TrustProbe identifies 104 taint-style vulnerabilities. Validation on a large corpus of real-world skills collected from public hubs such as ClawHub further shows that 25.1% of skill-agent trials exercise the identified vulnerable paths, with payload injection successfully weaponizing 15 of the vulnerabilities. These results reveal a systematic trust failure in skill-based LLM agents: untrusted skill content can reach security-sensitive operations and exercise authority delegated by users to their agents.

Retrieval-augmented generation (RAG) systems need inexpensive ways to route generated answers: accept low-risk outputs, review uncertain ones, and reserve strong verifiers for the expensive tail. We present RAGScope, a leakage-controlled protocol for evaluating local evidence gates that use only the task input, retrieved context, and answer text. The protocol combines context-grouped splits, fold-scoped preprocessing, group bootstrap intervals, deployment operating points, end-to-end runtime, and explicit source-shift stress tests. On three RAGTruth tasks, the enhanced gate RAGScope-E reaches 0.798 AUROC and 0.660 average precision (AP) in pooled grouped cross-validation. Its pooled AP exceeds ROUGE-L by 0.034 with a 95% context-group interval of [0.002, 0.064], although the AUROC gain is not significant and ROUGE-L remains stronger on data-to-text. At a top-10% review budget, RAGScope-E attains 0.748 precision; accepting the lowest-risk 50% yields 0.141 residual unfaithfulness. RAGScope-E runs in 6.22 ms/example on CPU, versus 145.75 and 223.07 ms/example for the tested DeBERTa-NLI and HHEM settings. A 14,900-example HaluBench stress test exposes the deployment boundary: an in-domain calibrated gate reaches 0.879 AUROC, but leave-source-out calibration averages only 0.466. Target-only calibration recovers to 0.675 AUROC with 100 labels per source and 0.685 with 200. Cheap evidence gates are therefore useful routing components, but learned calibration must be validated and adapted within the target domain.

Intracortical motor decoders degrade across sessions because the set of recorded units changes and persisting units can alter how their firing relates to behavior. Most existing methods update network weights on each new session or rely on unlabeled activity, which does not directly reveal such changes. We present APST, an Association Profile-conditioned Set-Temporal transformer that adapts to new sessions with all network weights frozen. From a few labeled calibration trials, APST summarizes how each unit's firing relates to behavior in a four-dimensional association profile computed in closed form. The profiles condition a set-attention encoder that accepts any number and order of units, followed by a causal transformer for streaming decoding. On held-out DANDI688 sessions from two monkeys, APST reaches velocity $R^2$ of $0.78$ and $0.81$, versus $0.40$ and $0.58$ for a variant that uses neural activity alone, and matches or exceeds an RNN fine-tuned on the same trials. On FALCON private held-out evaluation, it attains $R^2$ of $0.65$, $0.42$, and $0.44$ on M1, M2, and H1.

Runtime error healing lets a crashed program continue by generating code that repairs its live runtime state. Recent work shows that LLMs can generate such healing code, but it is evaluated only on small competition programs, and executing LLM-generated code inside a live process raises safety concerns that remain unaddressed. In this paper, we take LLM-based runtime healing toward practical use in real-world repositories. We first build HealBench, a benchmark of 265 runtime errors from 18 real-world repositories, each paired with a reference execution on the patched version. HealBench also provides a unified framework that lets LLM agents heal with cross-file context and live runtime state. We then design HealGuard, which requires healing code to be written in HealCore, an analyzable subset of Python, and uses static and dynamic taint analysis to check whether state changed by healing reaches operations protected by developers. We evaluate a dedicated healing method and three general coding agents with three backbone LLMs. The best setting resumes execution in 38.11% of instances and passes the target test in 28.68%, showing that existing agents can already heal a meaningful share of real repository-level crashes. However, among executions that pass, HealGuard flags 17.4% whose healing-changed state may reach a protected operation. On 684 controlled cases, HealGuard detects all unsafe cases, at the cost of a 68.42% false positive rate.

Long-form live-streaming TTS requires context-dependent prosody and paragraph-level coherence. However, many existing instruction-based TTS systems use global or uniform conditions, providing limited explicit control over clause-level relative prosodic changes. We introduce Speaker-Relative Inline Prosody Control, where each Pitch, Energy, or Speed instruction targets a clause relative to the preceding clause from the same speaker, while Pause uses an absolute duration interval. In codec-based TTS, Speed and Pause affect sequence length, whereas Pitch and Energy rely on residual codebooks. By analyzing Qwen3-TTS RVQ codebooks, we find that Energy concentrates in early residual codebooks, whereas Pitch accumulates across a deeper prefix. We therefore propose Scope- and Codebook-Aware Instruction Conditioning (SCIC), combining a Temporal Instruction Router for frame-level tag activation with Tag-Specific Codebook Weighting over residual codebooks. SCIC improves speaker-relative Pitch and Energy control over standard instruction fine-tuning using text-token tags. We further apply multi-reward GDPO post-training to jointly optimize control and quality, improving control accuracy while preserving CER and speaker similarity. In long-form synthesis, SCIC produces a more distinct paragraph-level expressive hierarchy than speaker-adapted SFT without instructions. Audio demos are available at: https://taoliveaigc.github.io/SCIC/

Unreliable visual inputs can harm task performance and cause potential physical safety risks for vision-language-action (VLA) models. We analyze how $\pi 0.5$ and GR00T models act under input faults such as image blackouts and freezing. We find that blackout and freezing produce distinct physical failure modes even when task-success rates are similarly low: freezing causes more extreme joint behavior, whereas blackout after gripper closure can cause more object drops, most markedly without proprioception. Selective intervention studies reveal that proprioception (current robot state) partly compensates for the removed robot depictions and reduces non-target contact. However, it cannot sufficiently restore task success when wrist-view object information is removed, even when aided by the remaining scene view. We then evaluate two mitigation approaches: camera-blackout training and training-free replacement of faulty visual embeddings. Both improve task success in selected conditions, but can increase unintended contact or disturbance to surrounding objects. Real-robot trials further show that successful execution under camera faults can still involve unintended physical interactions. These findings motivate designing VLA policies that use the robot and object information still available under camera faults to limit hazardous motion.

Deep research tasks require agents to navigate large knowledge spaces, synthesize evidence across many sources, and adapt their plans as findings emerge. Directed acyclic graph (DAG)-based multi-agent systems suit this setting because they support parallel execution and isolate each sub-task within a focused dependency context. Yet existing DAG-based agents instantiate a task-level plan before execution and repair the graph only after failures or missing evidence are observed. This Plan-then-Patch strategy is brittle for deep research: the system commits most strongly when its evidence is weakest, and later revisions waste computation on branches that should not have been planned. We propose DAGent, a DAG-based multi-agent framework with Evaluate-then-Grow incremental planning: an Orchestrator grows the task graph one batch at a time, conditioning each expansion on confidence and uncertainty signals from completed nodes. A hierarchical context layer propagates compact QueryDocs by default while preserving full execution traces for on-demand recall. The recorded DAG topology admits structural RL signals that outcome-only recipes cannot define; DAGRPO, a GRPO adaptation, injects topology-conditioned credit on Executor rollouts and a structural compliance regularization on Orchestrator plans. Across BrowseComp-Plus, GAIA, and xbench-DeepSearch, DAGent surpasses the strongest open-source baseline by 5.3 / 5.8 / 2.0 points at the Qwen3-235B-A22B scale, and the lead replicates across four open-source backbones and extends to GPT-5 at 327K context. At the Qwen3-8B scale, DAGRPO improves over a same-budget outcome-only GRPO baseline by 3.0 average Pass@1 points. A same-architecture comparison shows that evidence-conditioned planning reaches higher accuracy at lower per-task token, tool-call, and step footprints than its Plan-then-Patch counterpart. Code: https://github.com/hanwenliu6825/DAGent

Large Audio Language Models (LALMs) rely on effective audio encoders for multi-task performance. We introduce UniAE-MoE, a unified audio encoder designed to model cross-domain audio representations and achieve outstanding downstream understanding performance via a Mixture-of-Experts (MoE) architecture. Specifically, we explore mainstream audio encoders and integrate those from Qwen2-Audio and Audio-Flamingo 3, which demonstrate superior downstream capabilities. To facilitate effective model fusion, we improve our encoder using SwiGLU with shared experts to decouple encoder networks, and we further introduce a two-stage instruction-tuning strategy to better adapt the model to diverse downstream tasks. Moreover, we propose the task-specific data scaling (TSDS) technique to enhance \tool's understanding capabilities. On the XARES-LLM benchmark, UniAE-MoE attains a score of 0.802, achieving state-of-the-art performance. It also delivers top-tier performance in the official Interspeech 2026 Audio Encoder Capability Challenge, further demonstrating robust generalization across diverse audio tasks. Together, these results validate the effectiveness of \tool for unified audio understanding across speech, music, and general audio domains.

Machine unlearning has emerged as a crucial mechanism for removing hazardous knowledge and enforcing safety alignment in Large Language Models (LLMs). However, recent studies reveal a persistent security risk: unlearned models remain highly vulnerable to retraining attacks, where suppressed malicious behaviors rapidly resurface after benign fine-tuning. In this work, we investigate the optimization dynamics of unlearning and identify that this vulnerability stems from shallow alignment. Rather than effectively erasing target knowledge, models often exploit a shortcut by activating previously dormant parameters to act as spurious suppressors, forming a fragile inhibitory shell over intact malicious representations. To address this issue and enforce authentic memory deletion, we propose FDCU, a novel dual-constrained subspace projection framework. FDCU restricts parameter updates through a highly scalable, element-wise dual-masking rule: it preserves general knowledge manifolds via Fisher Information and strictly prohibits the abnormal activation of spurious suppressors via the Principle of Minimal Functional Intervention (PMFI). By reliably blocking the model's ability to superficially hide knowledge, FDCU promotes the authentic dismantling of target representations. Extensive experiments across specific knowledge erasure and safe output control tasks demonstrate that FDCU achieves state-of-the-art robustness against retraining attacks while maintaining near-lossless general utility, ensuring durable safety for LLMs.

While large language models have achieved remarkable success in isolated code generation, authentic software engineering requires sustained reasoning, complex state management, and continuous cross-domain abstraction. However, current evaluations of skill evolution in autonomous agents suffer from a critical identifiability problem: they structurally confound genuine capability abstraction with rote solution leakage (i.e., copying highly similar code from historical training data). To resolve this, we introduce EngramBench, a rigorous, capability-grounded benchmark governed by the strict axiom of capability overlap without solution overlap. Comprising 30 diverse learning tasks and 13 unseen transfer tasks, EngramBench challenges agents to navigate interactive, multi-hour development cycles driven by LLM-simulated users. Our extensive evaluation across 48 multi-hour execution trajectories -- corroborated by human-expert validation -- reveals a profound insight into procedural memory. We demonstrate that static skill banks do not magically bypass the "last mile" of exact code implementation, which remains bottlenecked by the base model's inherent reasoning limits. However, they serve as an indispensable execution compass. By navigating agents away from catastrophic, token-heavy trial-and-error, genuine capability abstraction slashes redundant context bloat and reduces overall coding time by over 55%. Ultimately, EngramBench shifts the evaluation paradigm from trivial pattern matching to the verifiable measurement of deep, cross-domain capability transfer.

When an off-the-shelf coding agent is used directly as a robot policy, observing a browser-based 3D interface through screenshots and acting by posing a virtual target gripper through a few tools, the agent's harness, its prompts, tools, and control rules, largely determines success, and until now it has been written by hand. We show that this harness can be improved automatically by another coding agent, the optimizer agent, and report two findings about what makes it work. First, the number of rollouts the optimizer agent sees per round governs whether the evolved harness is trustworthy, generalizes, and improves steadily. A single rollout is a noisy binary outcome, so with few rollouts per round a revision can be promoted on luck; enlarging the batch raises the signal-to-noise ratio of every promotion decision. Holding rounds fixed and growing the training set from 5 to 100 rollouts, held-out success rises from 47% to 67%, while small training sets overfit, reaching 70% on training tasks but only 54% held-out. Second, the optimizer agent must not be given free rein. With every revision it proposes accepted unconditionally, performance drifts downward within ten rounds as ill-judged edits accumulate; adding the most basic safeguard, Champion-Challenger selection that promotes a revision only if it strictly beats the incumbent on the same fixed evaluation set, turns the same loop into one that raises held-out success from 51% to 67% over 30 rounds. Automatic harness evolution for visual-interface robot agents is thus feasible, but its gains hinge on the rollout scale behind each decision and on how the optimizer agent's revisions are selected.

Robot demonstration generation requires a system to identify where an interaction should occur, plan a feasible motion, and execute the required contact. HiWE connects these decisions through a point-based interface between visual grounding and language-based planning. PointVLM is instruction-tuned to associate task-relevant objects with image coordinates using a mixture of point annotations, segmentation-derived samples, robot observations, and visual question answering data. Depth measurements lift these predictions into a semantic 3D representation. A language planner, 3DLLM, uses this representation to specify end-effector waypoints and gripper commands, while a hybrid grasping module resolves local grasp poses. The evaluation covers 14 simulated manipulation tasks and four physical-robot tasks, together with ablations of the visual training data, spatial inputs, and grasp selection. Here, zero-shot execution refers to deployment without task-specific demonstration training; the visual model uses existing robot data during fine-tuning. This paper describes the original point-based formulation of the framework; its relationship to the subsequent GeneralVLA extension is detailed in the introduction.

Autonomous AI agents can turn authors' goals into interactive narratives by independently organizing and carrying out generation and revision. As agents generate and revise extensive content, authors struggle to grasp its overall structure, local details, and relationships, complicating continued guidance. We present NarrativeSteward, an authoring environment that organizes outlines, worldbuilding, and narrative graphs as linked artifacts for agent implementation and author guidance. Agent dialogue and project-wide structural review help authors understand the evolving work and guide local and cross-layer revisions, while change records and execution verification help authors assess the resulting work. Technical tests validated the system's change records, recovery mechanisms, and execution diagnostics. In a 12-participant within-subject study, NarrativeSteward supported easier formulation of revision requests and inspection of changes, and greater perceived understanding of changes and story structure, than general-purpose agents. Qualitative findings show how reviewing the work and feedback helps authors develop requirements and guide subsequent delegation. We open-source NarrativeSteward at https://github.com/Tencent/NarrativeSteward.

Intrinsic motivation plays a central role in adaptive and goal-directed behavior by conferring agents reward-independent objectives and biases useful to act in noisy and uncertain environments. Active Inference addresses the problem of acting in a partially observable environment through a principled framework for belief updating and action selection. A key component of Active Inference is the specification of prior preferences, which shapes behavior by encoding desirable future outcomes. An intrinsic motivation approach called the Maximum Occupancy Principle (MOP) proposes that agents act so as to maximize occupancy over future paths of states and actions, with no preferences or epistemic targets. Despite its simple formulation, MOP gives rise to rich and adaptive behaviors that combine exploratory variability with goal-directed dynamics. In this work, we extend MOP to partially observable environments and introduce a Bellman reformulation of the Expected Free Energy for Active Inference, both incorporating belief-based inference over hidden states as part of the agent state. The Bellman formulation enables tractable offline computation via value iteration over the full belief-state space. We compare the resulting behaviors in a set of minimal experimental settings with uncertain food sources. We find that MOP agents switch between goal-directed (food seeking) behavior and exploration between different food sources, depending on their energy available and their belief state. In contrast, Active Inference agents mostly inhabit regions around a single food source, a strategy having both high pragmatic and epistemic value. We finally compare with Empowerment, which is shown to be qualitatively similar to Active Inference.

Speech transcripts used as long-term memory must preserve both words and stable speaker identities. Existing meeting-transcription metrics either ignore speakers or remap anonymous speakers independently in each recording, so they cannot measure whether the same person retains one identity across meetings. We evaluate persistent speaker attribution with Speaker Identified cpWER (SI-cpWER), which scores a corpus under one global speaker-ID assignment. The benchmark covers five commercial diarize-then-identify cascades, two open academic baselines, and ThyVoice on the full 129-meeting CHiME-8 NOTSOFAR evaluation set in clean and noiseaugmented form, plus CHiME-6. ThyVoice is our end-to-end reference system; it repairs overlap and gates the evidence used to create and update voiceprints. Requiring persistent identity changes the commercial ranking: ThyVoice records lower SI-cpWER than every evaluated commercial cascade in all three conditions and the lowest mean in the full panel, 47.13 versus 54.75 for the next system. Complementary lexical, diarization, per-recording attribution, and speaker-clustering diagnostics characterize upstream error surfaces in the final attributed record. These results show why persistent attribution must be evaluated directly in systems that reuse conversations across time.

LLM agents increasingly rely on reusable Skills for complex, multi-step tasks, creating a critical supply-chain attack surface where poisoned Skill content steers agent decision loops under benign requests. Existing skill poisoning attacks either colocate actuation with its contextual pretext or distribute actuation across multiple Skills, but do not explicitly separate the rationale for execution from the operation itself. In this work, we reveal that untrusted agent decisions fundamentally depend on two conceptually distinct Risk-Realization Factors (RRFs): an actuation factor (specifying what concrete operation is performed) and a pretext factor (providing the situational rationale for why the agent must perform it). Guided by this abstraction, we propose a coordination-based attack paradigm: decoupling pretext from actuation. Rather than fragmenting the malicious actuation, we preserve it as an intact operation within a downstream Steering Skill, while delegating the pretext factor to an upstream Grounding Skill that subtly alters persistent environment artifacts through routine utility operations. The intact actuation thus hides in plain sight, appearing completely legitimate and task-driven only when evaluated against the fabricated pretext. Building on this formulation, we develop an automated framework that discovers authentic execution dependencies, synthesizes coordinated pretext-actuation skill pairs, and iteratively refines poisoned skill instructions via runtime closed-loop feedback. Extensive evaluations across single-session and persistent cross-lifecycle scenarios demonstrate that decoupled skill poisoning achieves high attack success, exposing a critical blind spot in isolated Skill security audits. Our automated framework code is available at https://github.com/Wenxin-buaa/CoordPoison.git.

Continual alignment requires LLMs to adapt to new requirements without forgetting previously acquired behaviors. Natural-language instructions are flexible and composable but offer only indirect control, whereas post-training provides stronger adaptation at the cost of repeated parameter updates. We introduce Ready2Blend, which combines the flexibility of natural language with learned alignment. AlignFormer maps each requirement to a fixed-length alignment prompt stored in a modular prompt bank, while the backbone and prior prompts remain frozen. Composability regularization transfers the semantic geometry of textual requirements into prompt space, enabling inference-time blending and reweighting. Across two practical continual alignment settings, Ready2Blend is the only frozen-backbone method that matches post-training-based alignment methods, reaching $93.1$-$98.5\%$ of a joint-training reference with competitive retention, while requiring only a few prompt tokens and up to $4.3\times$ less training time. Its modular design further enables weighted personalization and order-free composition without retraining. Code will be released upon acceptance.

LLM-based agents extend their capabilities through third-party skills that provide task-specific instructions, scripts, and tool-use procedures. However, malicious instructions inserted into an otherwise benign skill can cause a benign user request to trigger dangerous Tool Calls, including data exfiltration, file deletion, or unauthorized code execution. This paper presents ActionGuard, which inspects skill-influenced Tool Calls immediately before execution. ActionGuard separates the target agent's action-generation context from the safeguard's authorization context. The target agent may use the original skill for planning, but the Reviewer does not receive the potentially poisoned raw skill text. Instead, it determines whether each action is justified by the trusted user request using a balanced skill profile, current and recent Tool Calls, and local script contents. ActionGuard intercepts each Tool Call at OpenClaw's before-tool-call stage and enforces the Reviewer's ALLOW or DENY decision under a fail-closed policy. We evaluate ActionGuard on 139 contextual and 180 obvious injections in a SKILL-INJECT-based setting against Dynamic Guardian and SkillGuard, using three open-source and two commercial Reviewer models. Each condition is repeated three times and evaluated using Attack Success Rate (ASR) and Task Success Rate (TSR). Overall, ActionGuard reduced ASR by 35.54 to 46.11 percent relative to existing safeguards and by 70.44 percent relative to No Safeguard, while maintaining high benign-task completion. These results show that execution-boundary authorization grounded in trusted user intent and runtime evidence can restrict unauthorized Tool Calls induced by skill injection.

Speech emotion recognition (SER) is the task of assigning emotion labels to utterances. Early systems relied on acoustic features, whereas recent approaches combine multiple modalities, most commonly speech and text. Still, performance remains poor on many datasets. Large language models (LLMs) have therefore attracted interest for SER, as they can process diverse inputs jointly with instructions. However, direct audio input raises questions of explainability. To address similar questions in image classification, concept bottleneck models were introduced. This work adapts concept bottlenecks to SER to examine how individual predictions depend on transcripts, acoustic descriptions and speaker attributes. Experiments test three LLMs on CREMA-D, IEMOCAP and MELD, with concepts extracted by separate tools. On scripted corpora, LLMs are strongly biased towards the transcript in the zero-shot setting, which lowers Macro-F1 from 27.8 to 5.8 on CREMA-D. Fine-tuning removes this bias, and the transcript raises Macro-F1 from 41.8 to 45.1. Removing speech rate changes 48% of Neutral predictions to Disgust on CREMA-D; removing intensity level on MELD changes predictions despite little change in Macro-F1. These findings show that aggregate performance changes alone do not capture the effects of concept removal on individual predictions.

Longitudinal clinical cohorts are multimodal, irregularly sampled and pervasively incomplete: in ADNI, positron emission tomography and cerebrospinal fluid assays are absent from roughly half of all visits. Linear state-space models handle irregular sampling gracefully but treat a missing modality by masking the input, leaving the transition operator untouched. We prove that this is a representational limitation: the latent state of any linear state-space layer whose transition operator does not depend on the availability pattern is an additive function of the availability indicators, so no such layer can represent an interaction between two modalities being jointly present or jointly absent. We propose CAMOS, which gives each modality a bank of second-order oscillators coupled through a matrix that sits inside the differential equation and is gated by availability, so the transition operator itself becomes a function of which measurements were taken. Coupling invalidates the analysis of uncoupled oscillatory models, and we restore it: a per-channel Gershgorin budget makes the effective stiffness positive definite uniformly over all $2^M$ availability patterns and all gaps, an energy argument charges amplification to availability transitions rather than sequence length, and a channel factorization preserves exact associative parallel scans. On ADNI, CAMOS outperforms uncoupled oscillatory state-space models and clinical fusion models on same-visit staging, landmark prediction and longitudinal forecasting, and under zero-shot transfer to OASIS-3 it is the only model that avoids collapse to the majority class.

The EU AI Act (Art. 27) requires deployers of high-risk AI systems to conduct Fundamental Rights Impact Assessments (FRIAs) before deployment, yet the evidence needed for credible assessments is fragmented across incompatible incident repositories, risk vocabularies, and legal texts. We present a reusable Semantic Web-based framework that consolidates this evidence for two high-risk public sector categories: employment and worker management (Annex III(4)) and access to essential public services (Annex III(5)(a)). A curated 150-record corpus is annotated along four axes using keyword, LLM, and hybrid methods and serialised as a SPARQL-queryable knowledge graph of 1,351 RDF triples. Five FRIA demonstration scenarios surface 103 records (68.7% coverage). Evaluation against a 69-record gold standard reveals that LLM-assisted classification of the employment domain achieves only $\kappa = 0.045$, a cautionary result for automated fairness-related evidence retrieval in this domain. All artefacts are released openly to support adoption by regulators, national authorities, and SMEs.

As Large Language Model (LLM) agents are increasingly deployed in complex environments, multi-turn interaction attacks have become a significant security challenge. Existing detection methods typically rely on historical context. However, this retrospective logic struggles to identify deep malicious intents that are split across turns to hide future risks. Inspired by speculative decoding, we propose the Speculative Safety Honeypot (SSH) framework. SSH uses a multi-agent simulation system composed of small LLMs to build an action-level speculate-and-verify workflow. In the speculation stage, SSH predicts future behaviors of the target agent and asynchronously builds a trajectory tree to expose potential risks in advance. In the verification stage, the system uses the target agent's real actions to calibrate and prune the trajectory tree, effectively reducing false positives. As a plug-and-playable component, SSH provides existing detectors with rich decision redundancy beyond the current interaction slice. By judging risk based on the evolution of the entire trajectory tree rather than a single point in time, the system reduces the reliance on the absolute precision of individual detection components. This improves the defense resilience and the warning lead-time of agent systems against complex temporal attacks.

Large language models (LLMs) may generate unreliable code on corner cases missed by testing, while formal verification can provide machine-checkable guarantees. Recently, researchers have proposed several benchmarks to evaluate the capabilities of LLMs in generating formally verifiable code, where LLMs need to formulate formal specifications, generate the corresponding code, and verify its correctness. However, existing benchmarks have two key limitations: (I) They primarily evaluate specification and code generation stage-wise, with code generation typically conditioned on an oracle specification. This setup overlooks whether strong stage-wise performance translates into end-to-end success. (II)They mainly focus on a single proof-oriented language and mathematically structured tasks, offering limited coverage of tasks common in software development. In this paper, we introduce VeriCodeBench, a benchmark for self-spec verifiable code generation, where the LLM relies solely on its own generated specification and code throughout the entire process. VeriCodeBench contains 400 language-native problems across C, Java, Rust, and Python, covering practical concerns in software development. We evaluate specification coverage, code validity, and joint problem-level success. We further introduce CodeNova to enhance the capabilities of LLMs in self-spec verifiable code generation. CodeNova makes requirements explicit through constraint-guided specification and uses verifier feedback to guide targeted implementation repairs. Experimental results reveal that self-generated specifications remain a major bottleneck, while providing more sophisticated specifications may not necessarily lead to higher verification success rates. CodeNova substantially improves performance across all evaluation metrics, enabling Claude Sonnet 5 to achieve the strongest results under the self-spec protocol.

Generating full-body co-speech motion for humanoid robots requires coordinating speech prosody, linguistic content, and embodiment-specific motion. To this end, we present ECHO-G, a framework jointly conditioned on speech audio and timed transcripts. Its Speech-Grounded Diffusion Transformer (SGDiT) combines frame-aligned acoustic features with token-level linguistic context, preserving their distinct granularities. Trained with rectified flow matching, it models one-to-many utterance-motion relationships directly in robot space. To support training and evaluation, we introduce a BEAT2-derived audio-text-robot dataset and a benchmark covering co-speech characteristics, robot-motion quality, and runtime efficiency. Comparative evaluation supports direct robot-space generation over the evaluated human-motion generation and retargeting pipelines, while modality ablations highlight the benefits of joint audio-text conditioning. We further demonstrate deployment on a physical humanoid robot. A complementary video-rating study also favors joint conditioning over the alternatives. The dataset and training, inference, and evaluation code are available through our project page.

3D visuomotor policies provide a strong foundation for spatially precise manipulation, yet current text-to-3D policies struggle to follow unseen fine-grained behavioral specifications beyond those covered by demonstrations. We study this challenge as unseen specification generalization, where language specifies behaviorally significant variations, such as target position, displacement, or articulated state, that are absent from policy training. We find that pretrained language representations and conventional global behavior-language alignment capture coarse task semantics but often blur nearby specifications that require distinct behaviors. We introduce T3DP, a Text-to-3D Policy framework for fine-grained language-behavior alignment. Rather than compressing each instruction and demonstration into a single global embedding, T3DP preserves their local structures and establishes bidirectional token-level correspondence between linguistic elements and behavioral segments. This directly grounds subtle linguistic variations in the behavior components they affect, preventing closely related specifications from collapsing in the representation space. The resulting specification-sensitive language representation conditions a point-cloud-based 3D diffusion policy, enabling more precise control over unseen behavioral specifications without modifying the underlying policy architecture. Across Meta-World, ManiSkill, and RoboTwin, T3DP improves average held-out-specification success over global language-behavior alignment by +11.0-14.2 points, with gains on all 15 task families; on real-robot tasks, it further raises average success from 47.5% to 65.0% (+17.5 points). Representation and action-probe analyses show that fine-grained alignment better preserves specification geometry and action-relevant variation, linking local behavior grounding to downstream control.

Skills extend an agent's capabilities by injecting instructions and information into the context, and are widely used by agents such as OpenClaw and Claude Code. Prior work shows third-party marketplaces host malicious skills that give attackers direct influence over the victim's agent. The emerging defense scans skills before installation, pairing deterministic static checks with an LLM-based semantic judge, as in NVIDIA's SkillSpector. We show that such defenses fall to an attacker who knows the detector. Our white-box LLM attacker, Pretext, iteratively crafts skills that evade detection while still delivering the payload and performing the benign task: moving the payload from code into natural language leaves static analysis inert, while framing it as the skill's legitimate purpose and splitting instructions across files keeps the LLM stage below its blocking threshold. Across three open-source models, Pretext achieves up to 97\% and 77\% against a frozen detector and a co-adaptive one, respectively, revealing major gaps in current skill scanners.

Cross-lingual transfer describes how knowledge in a source language benefits a target language. Measuring it quantitatively requires broad multilingual pre-training, as prior work has done with cross-lingual transfer matrices. We ask whether transfer is predictable from freely available typological features, and whether the prominence of high-resource source languages reflects typology or data quality and quantity. We show that typological databases contain cheap and dense signals about cross-lingual transfer. Our typology-only random forest on a 24-language prior-work transfer matrix scores leave-one-language-out $\rho{=}0.705$ and $R^2{=}0.49$, beating a non-typological control at $\rho{=}0.62$, which verifies the ability of typology-only predictions to reconstruct costly measured cross-lingual transfer. The signal survives leave-one-script-out and leave-one-family-out protocols, so script and family confounding do not explain the effect. By decomposing the transfer into a typology term and a resource-and-script bias term, we find the best-source ranking sensitive to this bias. In contrast, typology is not affected by this bias, which makes it a zero-compute screening tool that replaces hundreds of training runs with a model fit. Our code is available \href{https://github.com/dharmsen/typo-x-ling-transfer}{here}.

Long-horizon robot manipulation reuses skills across many task compositions, but improving these compositions with additional end-to-end demonstrations is costly. A practical self-improving system must decide both what to teach next and where to apply that supervision. We present ROBOCOACH, a world-model-guided coaching framework that uses imagined failures to guide demonstration requests and expert updates. Its Route-Imagine-Diagnose-Improve (RIDI) loop executes reusable skill experts inside COACHWORLD, our shared action-conditioned world model, and uses a progress judge to record the first subtask that fails to complete. Aggregated records select which subtask demonstrations to acquire and which expert adapters to update. Across two simulation suites and two real-robot platforms, imagined and deployed success correlate over 22 task-policy pairs (rho = 0.840). Controlled comparisons show that our coaching method outperforms matched baselines under matched data budgets and update schedules. With only 150 additional subtask demonstrations, success rises from 13.3% to 75.0% on Franka and from 40.0% to 83.8% on AgileX. The coached experts also transfer to four held-out compositions, achieving an average success of 35.0%, compared with 0% for a shared-policy baseline updated with uniformly acquired demonstrations. Together, these results show that world models can serve as active coaches, turning imagined failures into targeted supervision for modular policy improvement. Project Page: https://robocoach-ai.github.io/

Pretrained video foundation models encode rich semantic and spatiotemporal priors for embodied navigation, yet converting these priors into UAV motion typically requires expensive future-video synthesis and geometric reconstruction. We investigate whether the motion implicit in future visual prediction can instead be recovered directly from the predictive representations of a frozen video model. To this end, we present DiffWAM, a geometry-conditioned navigation world-action model that directly transforms multi-level predictive features into continuous camera trajectories. Its Grid-Motion module preserves spatial-temporal motion associations, while Latent2Pose grounds them with first-frame geometry to recover metrically meaningful 3D motion. Complete video rollouts and geometric reconstruction are required only for offline supervision, eliminating future-video decoding and multi-frame reconstruction during deployment. We further introduce FastDreamer, which overlaps predictive and geometric computation with ongoing flight and performs timestamp-aware asynchronous trajectory handoff for continuous UAV execution. DiffWAM achieves a trajectory RMSE of 0.3492 m and an endpoint success rate of 74.40% on the 1,000-sample DiffWAM-1000 benchmark, while representative real-world experiments demonstrate complex behaviors including constrained traversal, orbiting, S-shaped flight, and multi-stage navigation. An onboard DiffWAM-Flash implementation further reaches 1.08 s model-pipeline latency on NVIDIA Jetson AGX Thor. These results demonstrate that predictive video representations can be efficiently grounded into continuous 3D motion, providing a direct alternative to generate-then-reconstruct navigation pipelines. Project page: https://zzmmzzm.github.io/diffwam.github.io/.

Vision-language-action (VLA) models generalize broadly across robotic manipulation tasks, but complex environments require balancing task success with unintended contact. Runtime shields can correct individual actions, but they leave the underlying policy unchanged, so repeated disagreements may create a persistent policy-shield mismatch that blocks task progress. To address this challenge, we introduce FailBank, a four-stage self-evolving framework that converts runtime feedback into persistent policy improvement. During collection, a fixed CBF-based safety module serves as an observe-only teacher, producing counterfactual corrections while the policy remains in control. Outcome-aware admission then converts useful proposals into corrective targets and retains successful uncorrected actions as quiet anchors for guarded LoRA updates. We evaluate FailBank on the VLA-Arena benchmark across two difficulty levels and two VLA backbones. Compared with the base policies, FailBank improves the joint success-cost operating point. Across the two backbones, FailBank improves task success rate by 8.5 and 6.9 percentage points, while reducing policy-induced cumulative cost by 35.6\% and 23.8\%, respectively. Compared with runtime shielding, FailBank raises task success rate by 25.4 and 9.5 percentage points, while maintaining comparable policy-induced cumulative cost. These results show that runtime feedback can serve as persistent policy supervision rather than only as a temporary action constraint.

We investigate the problem of role-capability leakage (RCL), in which a role-prompted reasoning model generates convincing in-role text while continuing to exhibit capabilities on benchmarks that exceed those implied by the assigned role. For example, when a model is prompted to assume the role of a kindergarten student, one might expect its performance on a mathematics benchmark to reflect kindergarten-level ability rather than expert-level proficiency in solving calculus problems. We introduce RoleCapBench, a curriculum-grounded benchmark for evaluating RCL across six educational roles and four assessment levels spanning elementary school through A-level, and use it to evaluate three open-weight reasoning models. We find that although the models can generate stylistically convincing in-role responses, they consistently fail to align their underlying capabilities with their assigned roles. Naive role prompting yields strong role-voice scores of 1.218--1.389 while retaining above-role accuracy of 0.811--0.898. RCL persists across a range of prompting conditions, including prompts that explicitly instruct the model to match the role's capability level. To mitigate this problem, we propose Injection, an inference-time intervention that combines explicit, role-specific capability guidelines with a guiding prefilled response prefix. Injection improves role-capability alignment across models, reducing above-role accuracy by up to 0.562 while preserving in-role accuracy with a marginal drop of less than 0.058 across most models. All artifacts, including scripts and evaluation data, will be released upon acceptance.

Large language models have achieved remarkable capabilities across diverse domains, yet their safety alignment remains vulnerable to jailbreak attacks. In this work, we identify a previously underexplored failure mode - safety generalization lag - where alignment trained predominantly on natural language fails to transfer to the code domain. We show that this lag induces a code-completion blind spot, allowing malicious intent embedded within syntactically valid code to evade safety mechanisms. To exploit this vulnerability, we propose CodeMimicry, a fully automated black-box jailbreak framework that generates structured, object-oriented code prompts to induce harmful outputs via code completion. Experiments on 8 state-of-the-art commercial LLMs demonstrate that CodeMimicry achieves a 96.25% attack success rate with 1.51 queries on average, significantly outperforming both template-based and optimization-based baselines. Beyond empirical performance, we provide a mechanistic analysis of code-based jailbreaks through latent space representations, including projection onto refusal-related directions and activation steering. This analysis offers an explanation of how CodeMimicry bypasses safety mechanisms in code-related domains. Our findings reveal a weakness in current safety alignment and highlight the need for robust alignments in structured domains such as code.

AI could support children's pretend play, but it could also direct the play on behalf of children. Whether AI should have roles in children's lives is controversial because its influence on children remains uncertain. We conducted semi-structured interviews with 10 U.S. parents, each with at least one child aged 4-15. During the interview, we described the concept of AI-supported pretend play and provided participants with two boundary-case storyboards. We analyzed the interview data through codebook thematic analysis, using inductive coding and affinity diagramming organized around the research questions, and then used qualitative systems mapping to examine relationships within and across themes. We found that the same characteristics of AI, e.g., ability to assume characters, responsiveness, and adaptability, were seen by parents as potentially useful but also concerning. Parents imagined that AI could make role-based play accessible to all children or help parents participate in family play. However, they opposed the idea of AI for children's play without a clear understanding of how it works and its long-term influence on their children. Parents worried about children's loss of imagination and creativity, emotional attachment to AI, reduced human interaction, inappropriate behavior by AI and/or children, and their inability to manage children's AI use. Parents viewed AI not only as a play tool but also as a social actor and a possible perturbation in the existing family dynamics. The appropriateness of AI and child--AI interactions therefore emerged as a requirement for AI in children's pretend play, in addition to technical safeguards and parental control. We contribute an integrated account of parents' interdependent judgments and emphasize the need for longitudinal research with children and their diverse families.

We introduce DoGBENCH (Documentation Generation Benchmark), to our knowledge, the first benchmark for generating and maintaining real user-facing software documentation. It asks whether an agent can produce documentation that experienced technical writers would accept in review. The benchmark contains 292 items from open source projects, including Helm, PostHog, and Mautic. Each item gives the agent a pre-change repository and a trigger, such as a code pull request or a reported documentation gap. The agent must first decide whether the documentation needs an update. For items that need one, the agent must produce an acceptable patch in one attempt. For items that do not need updates, the agent must abstain. Task-specific rubrics, validated with project maintainers, score each patch on accuracy, completeness, reader guidance, placement, and repository conventions. The composite score combines patch quality with correct abstention, and a score of 100 means an agent meets every requirement for the task. Scores should not be interpreted as a percentage of an expert's capability. We evaluated seven agents. The highest-scoring agent reached 47.3 out of 100 on the 117-item held-out split. In a separate audit of 1,267 patches, the most common failure modes were task-completion gaps (45.5%), technical inaccuracies (36.6%), and incomplete conceptual or reference coverage (32.5%). Analysis of the corresponding trajectories identified three key patterns associated with these failures: (1) describing interfaces without examining how readers use them (36.0%), (2) missing decisive evidence and filling the gaps with plausible assumptions (33.1%), and (3) stopping after finding the first plausible documentation surface and leaving other affected pages stale (30.1%).

Physical LiDAR attacks are often evaluated using fixed primitives and manually selected parameters, despite their strong dependence on surrounding traffic. We present TACTIC, a scene-aware framework that uses a multimodal large language model (MLLM) to coordinate state-adaptive roadside LiDAR attacks. Under a gray-box threat model, TACTIC relies only on an attacker-operated roadside perception stack, without accessing the victim LiDAR's native point clouds or internal processing. Local perception provides metric vehicle states, while the MLLM combines these measurements with roadside imagery to infer relational traffic context and construct a semantic scene graph. Based on this representation, TACTIC selects and configures two complementary primitives: \emph{push-away}, which shifts the perceived range of a lead vehicle, and \emph{phantom-obstacle braking}, which triggers emergency braking through obstacle injection. Measured traffic states and empirically calibrated constraints ground the generated tactics in physically feasible operating regions. To accommodate MLLM latency, TACTIC overlaps reasoning and execution asynchronously while high-rate local perception detects scene changes and triggers replanning. Across 280 randomized CARLA trials, the full policy achieves a 100% collision rate, compared with 35% for a fixed rule, 60% for random selection, and 75% for a restricted LLM using mode selection with default parameters. Joint physical-and-image input achieves 100% success, versus 65% with physical measurements alone and 75% with imagery alone, while asynchronous $\Delta$ refresh reduces scene-mutation response from 7.4 s to 2.0 s. These results show that scene-dependent tactical planning can expose context-sensitive LiDAR failure modes that fixed attack policies may miss.

This paper presents an overview of the fourteenth edition of the BioASQ challenge, organized in the context of the Conference and Labs of the Evaluation Forum (CLEF) 2026. BioASQ is an international challenge series that supports progress in biomedical language processing tasks ranging from semantic indexing and information extraction to question answering and summarization. In 2026, BioASQ included six shared tasks: a) Task 14b on biomedical semantic question answering. b) Task Synergy14 on question answering for developing biomedical top- ics. c) Task MultiClinSum-2 on multilingual clinical summarization. d) Task BioNNE-R on extracting relations between nested named entities in Russian and English. e) Task ELCardioCC on clinical coding in cardiology. f) Task GutBrainIE on gut-brain interplay information extrac- tion. Across these six tasks, 87 distinct teams participated, submitting more than 1000 runs overall. As in previous editions, several submissions reached competitive performance, reflecting the continued progress of state-of-the-art methods across biomedical language processing tasks.

Mobile terminals need to provide application and network services while supporting users' control over their attention. We explore voice-first interaction organized around requests and delegated tasks, allowing users to leave a conversation and later inspect, revise, and retrieve the work. We present Richard, a system prototype that manages voice sessions, task execution, and result delivery separately, linking them through persistent request records. Conversation and task views provide visual feedback, while the backend coordinates immediate responses, dedicated service operations, and agent tasks. Request revisions, execution states, and notifications remain associated with the relevant task. We examine this design through Android functional records, controlled lifecycle verification, and execution records of a real programming request. Controlled verification reproduces revision, execution after confirmation, and result retention; deployed-service records show backend progress and failure feedback after client disconnection. These observations inform the design of task continuity, user control, and service integration in mobile voice interaction, providing an implementation basis for personal computing devices that accommodate intermittent user participation.

Terminal agents act through stochastic model generations, yet the ability to generate a useful action does not ensure its reliable execution. A poor command (e.g., wrong package install) can change the environment in ways that hinder subsequent progress, even when the model could generate a better alternative. We investigate whether allocating test-time compute at the model-harness boundary can improve action reliability and trajectory success, and what makes this allocation effective. To study these questions, we introduce Mid-Harness, which samples and verifies candidate actions before forwarding one for execution, while keeping the generator and harness unchanged. With a TMAX-9B generator, more action sampling yields little benefit under weak verification, whereas a capable verifier can exploit useful alternatives from the same generator. On TerminalBench-Lite, a GPT-5.6 Sol verifier raises Pass@1 from 50.00% for the base agent to 68.03% with 8 sampled actions. When the same TMAX-9B model serves as the verifier, pairwise verification performs best among the evaluated verification mechanisms. Distilling responses from the stronger verifier into TMAX-9B further improves Pass@1, while leaving the action generator unchanged. With TMAX-9B on TerminalBench-Lite, combining action and trajectory scaling reaches higher success at lower estimated token cost than generating more trajectories alone. Mid-Harness also improves performance across additional models, benchmarks, and harnesses. These findings identify action scaling as a promising target for test-time compute scaling in terminal agents.

Community Notes is a crowd-sourced approach for adding context to posts on X. Contributors propose and rate notes, forming the inputs to an open-source, open-data algorithm that determines which notes show broadly to users. Since September 2025, Community Notes' AI Note Writer API has provided an open, public interface for using AI to propose notes, while adhering to the founding principle that users, not the platform or an AI, control which notes show on X. Explicit note requests and user posts on X determine the AI API post feeds, ensuring that AI note writing responds to demand from X users. We present the design, operation and impact of the AI API, including analysis of the interaction between AI and human generated notes across topics. Unless otherwise stated, measurements and system description reflect June 2-29, 2026. The Community Writer is the largest AI API client and contributes the bulk of AI API output, generating 52% of notes selected as Helpful and shown broadly on X. The writer is guided by community input during both training and operation to prioritize, draft, evaluate and delete proposed notes. Beyond scale, the writer also offers speed, submitting the first proposed, non-deleted note on 60% of posts when compared to other writers. AI note writing is additive on top of human note writers, extending coverage of Community Notes on X. Among posts that have Helpful notes, 42% have only AI notes, indicating human raters did not feel motivated to propose an alternative. In contrast, 30% have only human notes, reflecting contribution beyond the scope of AI writing. The Community Writer is open-source software released under the Apache 2.0 license.

Digital twins increasingly support downstream analytical tasks that depend on time-series data, motivating interest in time-series foundation models (TSFMs) as scalable backbones. However, TSFMs are primarily pretrained for temporal continuation and often underperform on unseen tasks such as regression, and systematic empirical comparisons against state-of-the-art dedicated models in digital twin contexts remain limited. This paper makes three contributions. First, we benchmark five well-known TSFMs with frozen backbones on remaining useful life (RUL) prediction using the C-MAPSS dataset, finding that multivariate architectures substantially outperform univariate ones, particularly under varying operating conditions. This raises a deeper question: when cross-channel dependencies can be modeled through pretrained weights, target-task adaptation, and digital twin-derived representations, how much does each contribute, and are they complementary? Second, we propose a topology-informed fusion approach in which topological constraints, derived from the asset structure the digital twin stores among its information models, explicitly shape cross-attention, so that fused representations respect the physical system's local connectivity rather than relying on unconstrained all-to-all interactions. Third, we conduct an ablation study across C-MAPSS subsets of varying operational complexity that isolates the three sources and their interactions. The sources prove complementary rather than redundant, and topology-constrained attention outperforms unconstrained fusion, though by a small margin, enabling a frozen TSFM informed by digital twin representations to remain competitive or in some cases exceed state-of-the-art performance on this regression task.

While Abstract Meaning Representation (AMR) has historically improved performance on a range of NLP tasks, the benefit---or lack thereof---of AMR augmentation for modern LLMs is thus far unclear. In this paper, we attempt to reproduce recent work that reported substantial downstream gains from AMR augmentation, finding that these are likely due to specific choices in the experimental settings used: using a consistent and unified protocol for hyperparameter selection, we observe that text-only baselines consistently match or exceed the performance of AMR-augmented models. To investigate this null result, we introduce a perplexity-based probe measuring the degree to which AMR provides an LLM with supplemental relational knowledge not already available to the model. We find that AMR augmentation does not help LLMs improve their understanding of relational content in the sentence, indicating that augmenting these models with AMR offers no clear benefit on downstream tasks.

Mastering robot manipulation skills via reinforcement learning (RL) remains largely sample-inefficient. The most common RL algorithms rely on random action sampling to discover new strategies, resulting in agents that allocate most of their training budget to motions in free space, away from the contacts from which manipulation skills emerge. Existing intrinsic motivation methods based on model disagreement or epistemic uncertainty improve on isotropic noise, but they can also reward uncertainty in functionally irrelevant transitions, such as erratic motions in free space. In this work, we argue that tactile feedback provides a natural signal for exploration, and introduce TacEx, a framework that incorporates touch into epistemic uncertainty-driven exploration by decomposing model uncertainty across sensory modalities and directing curiosity toward the tactile channel. By anchoring curiosity to the sense of touch, TacEx drives the robot to discover complex contact dynamics, learning to manipulate and grasp objects without task rewards or expert demonstrations during exploration. The interaction-dense dataset collected through this tactile-driven curiosity supports offline learning of downstream pick-and-place policies without additional environment interaction. We further use tactile-driven exploration to post-train vision-language-action (VLA) models. Although the VLAs are initially pre-trained without tactile feedback, post-training with TacEx substantially improves downstream performance while remaining highly sample-efficient.

Speech-to-speech systems must solve tasks whose requirements emerge across conversational turns. We introduce SpeechConversationBench (SCB), a focused evaluation of spoken mathematical reasoning using 103 sharded GSM8K problems. The framework compares the original problem delivered in one turn (full), its concatenated information shards delivered together (concat), and incremental spoken disclosure across turns (sharded). We report final-answer accuracy for four commercial speech systems and LEGO, a proprietary speech pipeline developed internally by the SCBX Innovation Lab team with explicit conversational context management. Relative to concat, sharded accuracy decreases by 5.0-25.3 percentage points across the four commercial systems. LEGO achieves 77.5 percent accuracy in all three conditions, compared with 76.6 percent sharded accuracy for GPT-4o Realtime. The two single-turn baselines distinguish sensitivity to problem reformulation from the additional challenges introduced by incremental spoken interaction.

Unlearning a fact in one language does not guarantee its removal in others as changing the query or even the requested answer language can reopen seemingly forgotten knowledge -- a cross-lingual loophole. The most straightforward solution to this challenge -- unlearning in all languages -- is neither scalable nor desirable as it amplifies damage to unrelated model capabilities. We introduce the task of language budgeted multilingual unlearning where the goal is to select a subset of languages that maximizes cross-lingual erasure. To study this task we introduce the Cross-Lingual Unlearning Tensor, an unlearning benchmark that spans 174 language--script pairs and 25 atomic paraphrase types to examine when forgetting generalizes across linguistic expressions of the same knowledge. We further propose COVER, which selects source languages to maximize predicted COVERage of languages receiving no forget supervision, enabling unlearning on a language budget. Surprisingly, we find naively selecting strong individual sources does not reliably compose into strong source sets motivating our development of COVER. At deployment COVER only requires benign calibration data and access to the frozen model. Across three model families and two disjoint forget sets, COVER reduces mean held-out residual access by 7.8--27.3% relative to uniform source selection. We find these gains extend beyond synthetic benchmarks to real news documents in low-resource language settings using human translated data from the Low Resource Languages for Emergent Incidents (LORELEI) corpus.

In algorithmic statistics a string x is explained by a finite set containing it, and Kolmogorov's structure function records the smallest such model at each level of complexity. Vereshchagin's strong models, those computable from the data by a total algorithm, are essentially the cells of simple partitions. We read a partition of binary strings as a hypothesis, with the cell containing x as its model, and develop algorithmic statistics over symmetric partitions: the orbit partitions of groups acting on strings. The Galois connection between subgroups and partitions gives each ambient group a lattice of symmetric partitions, with canonical certificates, canonical costs, and an algebra of hypotheses. The resulting structure function and symmetric sophistication measure which part of the regularity of x is symmetric. For the full symmetric group every partition is symmetric: cells recover all Kolmogorov models, cells of cheap partitions recover exactly the strong models, and normal and strange strings are characterized by symmetry. For GL(n,2) the cells are exactly the linearly homogeneous sets, so linear symmetry is a restricted model class. For nonzero x, the linear-symmetry structure function lies in a band between the sufficiency line and the trivial bound, and both edges are attained: there are stochastic normal strings whose simple structure is invisible to linear symmetry. We also give coordinates on the space of permutation groups: each group is an element of a Burnside ring (its type) together with a permutation (its placement), and restriction moves refine partitions via the Mackey formula. In these coordinates the collapse for the symmetric group is a statement about placement, a linear hypothesis is determined by its type up to n^2 bits, and the maximal gap theorem shows that any space of symmetry hypotheses small enough to search is small enough to miss simple structure.

Current large language models (LLMs) are constrained by human-derived training data and limited by a single level of abstraction that impedes definitive truth judgments. This paper introduces a novel framework in which AI models autonomously generate and validate new knowledge through direct interaction with their environment. Central to this approach is an unbounded, ungamable numeric reward - such as annexed disk space or follower count - that guides learning without requiring human benchmarks. AI agents iteratively generate strategies and executable code to maximize this metric, with successful outcomes forming the basis for self-retraining and incremental generalisation. To mitigate model collapse and the warm start problem, the framework emphasizes empirical validation over textual similarity and supports fine-tuning via GRPO. The system architecture employs modular agents for environment analysis, strategy generation, and code synthesis, enabling scalable experimentation. This work outlines a pathway toward self-improving AI systems capable of advancing beyond human-imposed constraints toward autonomous general intelligence.

Recent and rapid advances in artificial intelligence (AI) make it increasingly important to understand the foundations of adaptive behaviour in autonomous agents, especially for building safe and efficient systems. While artificial neural networks have dominated the development of AI, recent work has begun to explore living biological neuronal networks as an alternative substrate for computation. These systems promise remarkable data and sample efficiency and rich dynamics, and may also inspire explainable and biologically plausible models. Here, we develop an experiment-informed active inference framework to model decision-making in closed-loop agents that mirror experimental setups using biological neurons. Using a generative model whose dimensions are matched to an experiment protocol, we systematically compare three decision-making schemes within this common generative model. Under matched episode counts (i.e. total data available for learning) to the in-vitro experiment, our simulations show that agents with short memory horizons reach a level of performance close to that of mouse and human cortical cultures (DishBrain platform), whereas longer memory horizons depart from it substantially. Increasing the planning horizon, by contrast, confers no comparable benefit. Because all model parameters are explicit, we can also track the quantities in our generative model that accompany this improvement, such as the risk term and the entropy of the transition and state-action mappings. Together, these results illustrate how active inference offers a formal language for comparing decision-making schemes in similar closed-loop control environments.

We investigate the performance of large language models (LLMs) on repetitive deterministic prediction tasks and study how the sequence accuracy rate (SAR) scales with output length. Each such task involves the repetition of the same operation $N$ times. Examples of such tasks include letter replacement in letter strings following a given rule, integer addition, and multiplication of string operators in many-body quantum mechanics. If the LLM performs the task by a simple repetition algorithm, the success rate would follow an exponential decay with sequence length. In contrast, our experiments on leading LLMs reveal a crossover that is sharper than exponential: $-\log\mathrm{SAR}$ grows super-linearly with $N$, and accuracy collapses around a characteristic length $N_*$, the accuracy cliff that separates reliable from unreliable generation. The hypothesis of independent per-step errors is rejected for every model and task we studied. The crossover is well described by a double-exponential accumulation law, $\mathrm{SAR}=\exp(-\beta_0 N\alpha^{N-1})$, whose crossover scale $N_*$ does not depend on the functional form chosen to fit it. To interpret this behaviour we introduce a minimal effective model in which step-correctness variables interact through dense random couplings and compete with an external field set by the prompt. Solved by direct enumeration, the model reproduces the super-linear error accumulation and the accuracy cliff qualitatively, and it assigns to each model--task pair two interpretable parameters, an intrinsic error rate and an error-accumulation factor.

Self-evolving agents improve their performance on long-horizon tasks by learning from their own interactions with an environment. A common approach uses the resulting failed trajectories as negatives for preference training. However, collecting an agent's own failures does not ensure that they provide informative supervision for further improvement. Obvious failures may be easy to reject without learning to identify errors in more plausible attempts. These plausible but incorrect trajectories can serve as hard negatives, helping agents learn to distinguish successful behavior from failures with less obvious errors. To enable agents to generate and learn from such informative negatives, we propose a co-evolving framework that couples a target agent's self-improvement with a failure agent trained to generate hard negatives. The failure agent learns exclusively from both agents' failed trajectories to generate higher-reward failures, which the target agent uses as negatives for preference training. As the target policy improves, its failures provide new training data for the failure agent. The updated failure agent, in turn, provides new negatives for target training, allowing negative generation to adapt alongside the target policy. Across online shopping, scientific reasoning, and interactive SQL querying, our framework improves average task reward by 5.7 over the baseline. These results demonstrate the value of learning to generate and use informative failures for agent self-improvement. We will make our code publicly available.

Motivated by the observed human-like behaviours in Large Reasoning Models (LRMs), this paper introduces a comprehensive taxonomy to characterise atomic reasoning steps and analyse the reasoning behaviours of LRMs. Grounded in human cognitive processes, we propose a taxonomy comprising five groups and seventeen categories. Through this taxonomy, we conduct an in-depth analysis of contemporary LRMs and distil four actionable takeaways for model optimisation. Most notably, we reveal that prevailing post-answer ``doublechecks'' are largely superficial and rarely yield substantive revisions. A targeted intervention further shows that explicitly eliciting richer reflection processes can substantially improve failed self-correction. To support this largescale study, we propose CAPO, an automated annotation method used to construct a dataset of 277,534 reasoning steps with strong agreement with human expert annotations. We further validate the main behavioural patterns on a newer reasoning model and a coding domain, demonstrating the broader applicability of the proposed taxonomy. All source code and data are available at https://github.com/hehepig4/psyche.

Despite its efficiency, there has been little research on the practical aspects required for real-world deployment of on-device AI models, such as the device's CPU utilization and thermal conditions. In this paper, through extensive experiments, we investigate two key issues that must be addressed to deploy on-device models in real-world services: (i) the selection of on-device models and the resource consumption of each model, and (ii) the capability and potential of on-device models for domain adaptation. To this end, we focus on a task of translating live-stream chat messages and manually construct LiveChatBench, a benchmark consisting of 1,000 Korean-English parallel sentence pairs. Experiments on five mobile devices demonstrate that, although serving a large and heterogeneous user base requires careful consideration of highly constrained deployment settings and model selection, the proposed approach nevertheless achieves performance comparable to commercial models such as GPT-5.1 on the well-targeted task. We expect that our findings will provide meaningful insights to the on-device AI community.

Self-training systems often degenerate due to the lack of an external criterion for judging data quality, leading to reward hacking and semantic drift. This paper provides a proof-of-concept system architecture for stable self-training under sparse external feedback and bounded memory, and empirically characterises its learning dynamics and failure modes. We introduce a self-training architecture in which learning is mediated exclusively by environmental viability, rather than by reward, objective functions, or externally defined fitness criteria. Candidate behaviours are executed under real resource constraints, and only those whose environmental effects both persist and preserve the possibility of future interaction are propagated. The environment does not provide semantic feedback, dense rewards, or task-specific supervision; selection operates solely through differential survival of behaviours as world-altering events, making proxy optimisation impossible and rendering reward-hacking evolutionarily unstable. Analysis of semantic dynamics shows that improvement arises primarily through the persistence of effective and repeatable strategies under a regime of consolidation and pruning, a paradigm we refer to as negative-space learning (NSL), and that models develop meta-learning strategies (such as deliberate experimental failure in order to elicit informative error messages) without explicit instruction. This work establishes that environment-grounded selection enables sustainable open-ended self-improvement, offering a viable path toward more robust and generalisable autonomous systems without reliance on human-curated data or complex reward shaping.

Currently, process reward models (PRMs) have exhibited remarkable potential for test-time scaling. Since large language models (LLMs) regularly generate flawed intermediate reasoning steps when tackling a broad spectrum of reasoning and decision-making tasks, PRMs are required to possess capabilities for detecting process-level errors in real-world scenarios. However, existing benchmarks primarily focus on mathematical reasoning, thereby failing to comprehensively evaluate the error detection ability of PRMs across diverse reasoning scenarios. To mitigate this gap, we introduce LSR-Ben, a process-level benchmark specifically designed for assessing PRM's performance across two primary reasoning domains (scientific and logical reasoning) and nine subdomains. We conduct extensive experiments on a diverse set of 22 models, encompassing both PRMs and LLMs, and derive two key findings: (1) In domains beyond mathematical reasoning, the error-detection ability of existing PRMs and LLMs is found to be markedly weaker by comparison. (2) In general, LLMs exhibit a tendency toward over-identification of errors compared to PRMs, whereas PRMs exhibit an inherent tendency to overlook errors compared to LLMs. We hope LSR-Ben can foster future researches on PRMs for broader domains, thereby enhancing the reasoning capabilities of LLMs.

Skills provide an effective mechanism for improving LLM agents on complex tasks, yet in existing agent frameworks, their creation, refinement, and selection are typically governed by external teachers, hand-designed rules, or auxiliary modules. As a result, skills remain external resources to be invoked, rather than capabilities that agents can develop, adapt, and internalize through experience. To endow LLM agents with autonomous skill mastery, we propose SkillMaster, a training framework that teaches agents to create new skills, refine existing skills, and select accumulated skills during task solving. This capability is achieved through three key designs. First, we train agents through trajectory-informed skill review, teaching agents to propose, update, or retain skills based on evidence from completed episodes. Second, each candidate skill edit is designed to be evaluated by its counterfactual utility on related probe tasks, providing a direct learning signal for training skill-editing decisions. Third, we introduce DualAdv-GRPO, which separately estimates advantages for task-solving actions and skill-editing decisions, stabilizing joint training across task solving and skill management. Experiments on ALFWorld and WebShop show that SkillMaster improves the overall success rate over state-of-the-art baselines by 8.8% and 9.3%, respectively, achieving the best performance among all compared methods. Further analysis reveals a marked shift in agent capability: agents trained with SkillMaster can identify skill failures, refine procedural knowledge from trajectory evidence, and transfer improvements to future tasks with limited skill-bank edits. Overall, SkillMaster moves LLM agents beyond mere skill use toward self-improving agents capable of developing, adapting, and applying their own skill repertoires.

Long-horizon reasoning requires deciding not only what actions to take, but how many to execute open-loop before replanning. This number, the execution depth, balances replanning cost against compounding execution errors. Most current systems either fix the execution depth as a hand-tuned scalar or adjust it at inference time with heuristic rules decoupled from the policy; we argue both can be suboptimal. In this work, we treat the execution depth as a learnable, history-conditioned variable of the policy itself, and propose PACE, a model-native VLM policy that jointly predicts what to execute and for how long under a hard decision budget. We evaluate PACE on four long-horizon environments spanning fully and partially observable settings and visual and textual modalities: Sliding Puzzle, Sokoban, ALFWorld, and ScienceWorld. Across all settings, PACE improves success rate by 3.1 to 15.6 percentage points, while using fewer decisions on average, achieving empirical success-decision Pareto dominance.

Offline-to-online reinforcement learning (RL) improves sample efficiency by leveraging pre-collected datasets prior to online interaction. A key challenge, however, is learning an accurate critic in large state--action spaces with limited dataset coverage. To mitigate harmful updates from value overestimation, prior methods impose pessimism by down-weighting out-of-distribution (OOD) actions relative to dataset actions. While effective, this essentially acts as a behavior cloning anchor and can hinder downstream online policy improvement when dataset actions are suboptimal. We propose RankQ, an offline-to-online Q-learning objective that augments temporal-difference learning with a self-supervised multi-term ranking loss to enforce structured action ordering. By learning relative action preferences rather than uniformly penalizing unseen actions, RankQ shapes the Q-function such that action gradients are directed toward higher-quality behaviors. Across sparse-reward D4RL benchmarks, RankQ achieves competitive overall performance against seven baselines. In vision-based robot learning, RankQ enables effective offline-to-online fine-tuning of a pretrained vision-language-action (VLA) model in a low-data regime, achieving an average simulation success rate 38.2 percentage points higher than the next best method. In a high-data setting, RankQ improves simulation performance by 13.7 percentage points over the next best method and demonstrates strong sim-to-real transfer, increasing real-world cube stacking success from 43.1% to 88.9% relative to the VLA's initial performance.

A significant hurdle for current LLMs is the execution of complex, multi-stage tasks. Group Relative Policy Optimization (GRPO) has been emerging as a leading choice, but its reliance on sparse outcome rewards severely limits credit assignment across intermediate steps. Existing remedies such as running full rollouts to assign step-level advantages, calling external LLM judges at each step, or computing intrinsic rewards that require ground-truth answers at every evaluation introduce significant costs or practical constraints. We hypothesize that internal correctness probing over LLM hidden states can be repurposed as a step-level reward signal, potentially addressing all of these limitations at once. However, existing probing research assumes clean inputs, and we first show that this assumption breaks down in multi-step settings: hidden-state probes degrade severely under prefix contamination tracking coherence with the (possibly corrupted) prefix rather than grounded correctness, while attention-based features remain robust to contamination but underperform on clean prefixes. Building on this complementary relationship, we propose the Prefix-Aware Internal Reward (PAIR), a two-stage model with a frozen hidden-state probe estimating belief-consistency and a lightweight attention-based head correcting it toward grounded correctness. Experimental results show that PAIR achieves the highest AUROC on contaminated trajectories while operating at negligible inference cost, enabling dense step-level reward signals for GRPO training without external model calls, ground-truth dependencies, or full-trajectory rollouts.

Despite the emergence of diffusion large language models (D-LLMs) as an alternative to autoregressive large language models (AR-LLMs), safety monitoring for D-LLMs remains largely unexplored. Unlike AR-LLMs, D-LLMs generate text through a multi-step denoising process, exposing intermediate hidden representations that may contain safety-relevant information unavailable in standard single-step monitoring setups. Motivated by the suitability of lightweight probes for always-on monitoring, we analyze which trajectory-level signals best indicate when such probes are likely to struggle. We find that the most informative signal is safety hesitation: intermediate hidden states repeatedly falling within a small margin of the probe's decision boundary. The number of such hesitation steps in D-LLM's trajectory predicts probe failure effectively, providing a proxy for sample difficulty. Building on this analysis, we propose $D^2$-Monitor, in which a lightweight probe runs on every sample, producing a base prediction and a hesitation estimate that decides whether to escalate the sample to a more expressive probe. That probe is trained on trajectories with at least one hesitation step, and reads only the minimal span of those steps. Evaluated on 3 datasets (WildGuardMix, ToxicChat, OpenAI-Moderation) across 4 D-LLMs, $D^2$-Monitor achieves state-of-the-art performance with a compact parameter footprint ($\leq$ 0.93M parameters), and exhibits the best trade-off between effectiveness and efficiency relative to 8 baselines.

Knowledge graphs (KGs) have become the core backbone of numerous downstream tasks such as question answering and recommender systems. However, despite all this, KGs are often very incomplete. To perform zero-shot knowledge graph completion in unseen KGs, which have different relational vocabularies from those used for pre-training, KG foundation models (KGFMs) receive a wide range of attention. Existing KGFMs often perform training using random negative triples, which are constructed by replacing the head or tail entity of a positive triple with a random entity. However, these negative triples are often constructed with limited quality, providing weak supervision for KGFM training. In this paper, we propose a simple yet effective adaptive negative sampling approach, KMAS, to enhance existing KGFMs. KMAS constructs hard negative triples through the updated relation embeddings generated from the existing KGFM's relation encoder. To further adaptively align with the evolving capability of the KGFM during the training process, KMAS adjusts the ratio of hard negative triples dynamically throughout the whole training process: after a warmup phrase, it increases the ratio linearly and then decreases linearly. Extensive experiments are conducted over 44 data sets. Experimental results demonstrate that our proposed negative sampling method can enhance many SOTA KGFMs without requiring excessive additional time or memory consumption.

Evaluating the efficiency of algorithmic code requires test cases that expose runtime bottlenecks. Previous methods generate efficiency test cases either by increasing input size or by generating code-specific inputs that make the given implementation run slowly. Consequently, they do not address the structural input conditions that drive the algorithmic worst case. We introduce STAB, a specification-driven pipeline that generates test cases that expose algorithmic bottlenecks from a natural-language problem specification alone. STAB separates the task into constraint-bound maximization and adversarial structure injection. (i) The constraint saturator extracts constraints and resolves large admissible size assignments using rule-based saturation and CP-SAT optimization over related variables. (ii) The adversarial scenario injector retrieves implementation-level adversarial construction principles from a curated scenario catalog using keyword matching and K-nearest neighbors (KNN). STAB encodes the problem specification, resolved boundary, and retrieved construction principles into a structured generation specification, from which the LLM synthesizes a Python test case generator. On CodeContests, STAB raises the rate of generated test cases that expose algorithmic bottlenecks from 50.43% to 73.45% on average across open-source LLMs and from 57.45% to 71.85% on average across closed-source LLMs, with consistent gains across Python, Java, and C++. Our code is available at https://github.com/suhanmen/STAB.

AI coding agents often lose project-specific rationale across sessions and may repeat approaches that previously failed. The bottleneck is often not model capability but missing coding agent memory: a durable record of what was decided, tried, and ruled out. We present projectmem, an open-source, local-first memory and judgment layer for AI coding agents, which records development as an append-only, plain-text log of typed events (issues, attempts, fixes, decisions, and notes) and deterministically projects it into compact summaries served through the Model Context Protocol (MCP). When invoked before an edit, a file-scoped advisory precheck returns recorded failures, open issues, and churn for the target file, a design point we call Memory-as-Governance. We evaluate feasibility through an author-run, six-month self-study on two machines, a controlled latency benchmark, and client-compatibility tests. The study contains 3,228 unique events from 27 projects; 56% are issue-lifecycle events, and 86 of 427 distinct issues closed by a recorded fix had at least one failed attempt before the first fix in event-log order. These histories are candidates for later warnings, not evidence that warnings prevented failures. On a synthetic repository, the median analysis time for a 1,500-event log decreased from 42.0 seconds to 79.3 milliseconds, while git subprocesses per check decreased from 1,500 to two and the tested warning output was preserved. The results establish the feasibility of persistent project memory for AI coding agents and motivate a controlled repeat-failure benchmark for measuring whether coding-agent memory prevents repeated failures. Source code is available at https://github.com/riponcm/projectmem.

We present CombEval, a dynamic benchmark for evaluating combinatorial counting in large language models. CombEval represents each problem as a typed Cofola specification over entities, combinatorial objects, object dependencies, and constraints, enabling controlled generation of natural-language counting problems with exact solver-verified answers. Unlike static collections, CombEval supports systematic variation of object type, entity scale, constraint count, and reasoning depth. We evaluate 11 LLMs under direct and code-augmented settings and find that models remain brittle on ordered objects, indistinguishable elements, relatively positional constraints, and nested object dependencies. Error analysis further identifies failures in constraint interpretation and counting principles. CombEval provides a diagnostic testbed for studying when and why LLMs fail at combinatorial reasoning. The code and generated benchmark suites are publicly available at https://github.com/YuxuZhou-CN/combination-problem-generation.

We present the AI Training Manager, a bounded LLM-based metacognitive monitoring-and-control layer for machine-learning training. The Manager asynchronously observes structured telemetry from an active training run, assesses the current training regime, and selects adaptive interventions through a constrained, deterministically verified action interface. We evaluate the feasibility of this architecture using deliberately induced, interpretable failure modes in supervised and reinforcement learning tasks. For supervised learning, we induce a multi-objective overfitting failure in a GPT-2-style model trained on TinyStories. The Manager prevents the resulting late-run validation collapse, reducing final validation loss by 54.3% relative to the same stressed recipe. For reinforcement learning, we study robotic reaching under opposing multifactor stress regimes. In a conservative regime, the Manager raises final deterministic safe success from 0.413 to 0.705. In an aggressive regime, it raises safe success from 0.121 to 0.764. The same Manager instructions and action surface are used in both regimes. These results suggest that bounded LLM reasoning can provide a practical metacognitive control layer over ongoing learning processes.

Large language models (LLMs) are rapidly shifting toward agents that solve tasks through diverse interfaces, including web and graphical user interfaces (GUIs). Among these, the terminal command line provides a text-based, general-purpose interface, covering tasks from system operations to data science and machine learning. However, scaling terminal-agent training remains challenging, as it requires diverse and coherent task instructions, executable environments, and reliable verification, while lacking naturally grounded supervision data. In this work, we propose SETA, a scalable framework for generating verifiable terminal environments for reinforcement learning (RL). The framework consists of two pipelines sharing a unified verification mechanism: SETA-Synth converts diverse sources into standardized RL environments, and SETA-Evol further expands from existing environments with adaptive control of difficulty and diversity. Together, we construct and release SETA-Env, the largest open-source verifiable terminal RL dataset to date, containing over 4,500 environments. We evaluate our dataset by training Qwen3-8B with GRPO on SETA-Env, achieving 12% pass rate on Terminal-Bench 2.0, the best reported result for an RL-trained model at the 8B scale. We further observe gains on DeepSeek-V4-Flash under the same terminal agent harness, with pass@1 on Terminal-Bench 2.0 improving from 40% to 43% and pass@5 improving from 54% to 58%. These results demonstrate that SETA- Env provides high-quality training environments for terminal agents and serves as a valuable resource for advancing research on terminal-based agent learning.

Personalized multimodal large language models (MLLMs) aim to generate user-specific responses, but existing methods mainly rely on profile-level information and overlook diverse user preferences. We identify group preference collapse, where multi-user personalized MLLMs become insensitive to individual preferences and drift toward dominant population-level choices due to suppressed preference signals and unreliable preference use during generation. We propose PrefMoE, a preference-centric framework that separates stable profile information from preference-related representations. PrefMoE decomposes preferences into shared prototypes and personalized residuals, preserves individualized residuals with imbalance-aware learning, counterfactual pseudo-user augmentation, and residual decorrelation, and routes profile and preference factors through separate LoRA adaptation paths. Experiments across multiple MLLM backbones show that PrefMoE improves preference-sensitive personalization while substantially reducing preference collapse. Project page: https://prefmoe.github.io/.

Synthetic textbook data has improved language model pre-training, but prior work largely treats the benefit as a property of generated content or local rewriting style. We study a different factor: whether related content is organized into coherent book-level documents. We contribute both a scalable synthesis pipeline and controlled evidence that this organization matters. The pipeline retrieves source material from a pre-training corpus, clusters it into topical units, plans hierarchical tables of contents, and assembles source-grounded sections into complete books (our Full setting), yielding 686K textbooks (32B tokens) across 15,000+ disciplines. Replacing natural books in a mid-training mix with this corpus improves downstream performance by +1.09 on average. Controlled comparisons then disentangle the relevant design factors. A content-matched Split condition holds generated text and tokens fixed but treats each section as an independent document; Full's +1.02 mean gain isolates document packaging. A length-matched RandomConcat control that joins sections from different books remains below Full, ruling out document length alone. A retrieval-pool-matched Rephrase condition independently rewrites individual retrieved documents under the same audience-by-style scheme, without clustering, TOC planning, or book assembly; Full's +1.17 gain demonstrates the value of structured synthesis. On Llama3-8B, Full likewise outperforms both RandomConcat and Natural Books, supporting book-level organization as a useful axis for synthetic pre-training data design.

Training large language model agents for long-horizon tool use typically relies on interactions with real or synthesized executable environments, whose construction and verification are costly, or on external simulators that are difficult to ground. We introduce EnvACE, an agentic reinforcement learning method that replaces external environment interaction during training with world rehearsal. The policy alternates between acting and rehearsal: it first generates a tool call, then plays the role of the environment to produce the response induced by that action, and conditions subsequent decisions on the rehearsed response. Both roles are jointly optimized end-to-end using task-success rewards. Through world rehearsal, the policy internalizes the relationship between actions and their environment responses in its parameters, yielding an agent world model that directly supports decision making. Across BFCL-v4, tau^2-Bench, VitaBench, and FinMCP-Bench, EnvACE achieves strong and transferable performance, outperforming environment-scaling baselines in the overall evaluation. Controlled studies further show that world rehearsal consistently improves policy learning across model scales. At test time, the internalized world model enables private rehearsal before committed execution, yielding further gains under a moderate rehearsal budget without additional external interaction. Our findings establish world rehearsal as a new path toward scaling LLM agent training beyond the constraints of external environments. Our code is publicly available at https://github.com/Within-yao/EnvACE.

Selecting the top-$k$ elements is a fundamental operation for inducing sparsity in large-scale models and optimization problems, enabling robust expert activation, token routing or attention pruning. However, hard top-$k$ is non-differentiable, while existing differentiable alternatives become increasingly expensive as the number of coordinates grows. We introduce Fast LapSum, a scalable solver for the LapSum soft top-$k$ formulation that preserves an exact selection mass of $k$, while supporting end-to-end differentiation. In Fast LapSum, we reduce sorting cost using probabilistic bracketing, which restricts sorting to a narrow band of scores around the threshold using a binomial order-statistic from kernel-noised samples. A certification pass upgrades the probabilistic localization to a verified one at the cost of one additional linear pass, with a full-sort fallback that covers the worst-case scenario. Our certified GPU implementation processes $10^6$, $10^7$, and $10^8$ scores in median times of $0.92$, $1.39$, and $7.24$ ms, respectively, making exact-budget soft top-$k$ practical within million-scale optimization loops. We demonstrate this capability in two applications: megapixel sparse adversarial examples with a small fraction of initial image pixels, where Fast LapSum achieves an order-of-magnitude speedup over state-of-the-art methods, and 3D Gaussian splatting. In the latter, we use Fast LapSum to reduce the number of Gaussians produced by Adaptive Density Control to a substantially smaller number while retaining nearly the same rendering quality and massively reducing computation. These results demonstrate that exact-budget differentiable top-$k$ can be incorporated into practical million-scale optimization pipelines.

Uncovering the internal mechanisms underlying the safety capabilities of large language models (LLMs) is crucial for developing trustworthy artificial intelligence. Currently, mechanistic interpretability studies on multilingual safety are largely confined to local components, such as isolated neurons. However, this static and fragmented perspective overlooks the synergy among components and fails to elucidate how safety signals dynamically propagate within the model to drive safety decisions ultimately. In this work, we move beyond isolated neurons to identify and target the cross-layer functional pathways formed during safety signal propagation, thereby uncovering the mechanisms driving the cross-lingual safety gap. Specifically, we first identify monolingual safety pathways and validate their impact on refusing harmful requests. Subsequent cross-lingual analyses reveal a sparse subset of cross-lingual shared safety pathways, confirming that this intersection acts as the internal bridge transferring safety capabilities from high-resource (HR) languages to non-high-resource (NHR) languages. Building on these mechanistic findings, we propose a pathways-targeted alignment method based on the cross-lingual shared safety pathways. Experimental results show that updating only a small fraction of pathway parameters significantly improves safety in NHR languages while largely preserving the model's general capabilities.

Vision-Language-Action models (VLAs) integrate visual perception, language instruction, and action generation into end-to-end policies across heterogeneous architectures. However, enabling VLAs to self-evaluate their action generation reliability without external supervision remains a major challenge. Existing methods either rely on expert annotations or estimate uncertainty only from output statistics, largely ignoring internal signals. In this work, we observe that internal visual modality entropy exhibits consistent distinctions between successful and failed tasks across heterogeneous VLAs. Although VLAs' architectures differ in their action generation, we show that they share a common latent action generation abstraction evolving under visual perception, language instruction, and State Input, which we formulate as a Conditional Generative Markov Chain. Based on this formulation, we propose MAE (Markov Attention Entropy), a self-evaluation framework that directly converts internal attention signals into architecture-aware reliability scores, and introduce LIBERO-Reflect, a 4,000-episode benchmark combining 2,000 standard episodes and 2,000 challenging episodes across four subsets. Extensive experiments across heterogeneous VLA architectures and diverse scenarios show that MAE consistently outperforms state-of-the-art baselines on AUPR, AUROC, and FPR@95.

Agentic reinforcement learning (RL) has emerged as an important post-training approach for enhancing the capabilities of Large Language Models (LLMs). However, existing methods face a trade-off between policy performance and resource efficiency. Conventional Proximal Policy Optimization (PPO) implementations incur substantial memory overhead from a separate critic, whereas critic-free group-relative methods require multiple rollouts and face potential learning bottlenecks on long-horizon tasks. In this work, we propose Single-rollout Autoregressive Policy Optimization (SAPO), an efficient PPO-style framework that unifies policy optimization and value learning within a single causal language model. SAPO exploits the autoregressive structure of LLMs to sequentially generate action and value estimation at distinct causal boundaries with shared parameters, and then jointly optimizes the PPO objectives and an auxiliary on-policy SARSA objective with turn-level generalized advantage estimation, where the latter is designed to facilitate value learning. Extensive experiments on ALFWorld and WebShop with Qwen2.5-1.5B/7B and Qwen3-14B demonstrate that SAPO reduces peak GPU memory usage by 23.1% and per-iteration runtime by 24.8% over strong PPO baseline, while matching or slightly improving task success rate. Our experiments also show that SAPO outperforms Group Relative Policy Optimization (GRPO) and recent cutting-edge variants in both task success and training stability.

Autonomous computer-use agents are increasingly applied to long-horizon tasks requiring coordinated application calls, persistent state tracking, and verifier-sensitive writes, yet they remain prone to procedural failures: misreading application state, tool semantics, or task progress. Procedural memory promises more consistent decisions and less redundant exploration, but constructing high-quality memory without model training remains challenging. We introduce CONTRAMEM, a source-flexible, training-free framework for self-evolving procedural memory that treats same-task outcome variation as supervision: differences in correctness, efficiency, recovery, and failure modes expose outcome-relevant procedural distinctions, distilled into a compact bank of app-level Function Cards and task-level Skill Cards that evolves through localized curation rather than append-only accumulation or whole-bank rewriting. On held-out GAIA2/ARE computer-use tasks, CONTRAMEM more than doubles the success rate across the three source-model targets (26.2% to 55.3%), with consistent per-model gains (GPT-5.5: 27.5 to 61.0; Claude Sonnet 4.6: 28.0 to 52.5; DeepSeek V4 Pro: 23.0 to 52.5). The same bank transfers unchanged to the unseen Qwen3.7 Plus (18.5 to 35.5), indicating transferable procedural knowledge rather than model-specific behavior. The same construction carries over unchanged to AppWorld, beating both no memory and its own single-source self-memory variant for all three mid-tier agents on both public test splits. Under a matched trajectory budget, heterogeneous multi-model trajectories yield stronger memory than self- or same-model multi-rollout memory: the margin comes from contrastive behavioral diversity, not stronger source agents or more sampling.

As language-model agents become more capable of iterative search, corpus access is shifting from retrieval toward interaction. Agents can explore the corpus, inspect documents, and use newly discovered evidence to decide what to examine next. Yet accessible evidence may still fail to become usable within a finite interaction budget. We call this progressive failure Evidence Blindness: supporting documents may never enter view, may remain unopened, or may fail to expose the decisive evidence even after being opened. A key reason is that agents often have to infer useful evidence directions during interaction, spending limited budget on deciding where to search next. Existing approaches either leave corpus structure largely implicit or reconstruct useful directions at query time. AtlasNav instead organizes reusable cross-document structure before any query arrives. It builds a persistent multi-view Corpus Atlas, which each query can navigate adaptively while still accessing the original documents directly. On BrowseComp-Plus, AtlasNav outperforms the previous state-of-the-art interactive corpus access method across different backbones. On DeepSeek, it improves strict accuracy by 7.47 points while reducing query-time inference cost by 30.22%.AtlasNav also reduces Evidence Blindness, realizes complete evidence earlier, remains robust to corpus-structure and scale shifts on PhantomWiki, and achieves leading performance on heterogeneous enterprise data.

Large language models (LLMs) are being used as policies for autonomous decision-making and planning in many domains. Despite their strong reasoning capabilities, LLMs struggle with long-horizon tasks, especially under partial observability. World models are a promising way to enhance policy performance, both during training and inference. During inference, agents currently use world models to simulate the consequences of candidate actions before choosing an action, which can improve decision-making. However, we argue that simulation alone is an incomplete interface for decision-making under partial observability: simulation does not adequately capture uncertainty about the current state, which agents may need for accurate decision-making. We address this limitation with Belief-Based World Models (BB-WMs), which maintain a belief that LLMs can query to access information on what is known and uncertain about the current state. Before developing methods to learn accurate BB-WMs, this paper focuses on a more fundamental question: does exposing a world model's belief directly to an LLM policy improve decision-making? Our results show that giving LLM agents access to beliefs improves task performance under partial observability, while remaining complementary to existing simulation-based world models. Code: https://github.com/skumar-ml/belief-world-models.

Self-play supports the self-evolution of language models, but solver performance can plateau or decline across rounds without guidance. Existing unguided methods typically use difficulty, learnability, or diversity signals to keep questions challenging and varied, without identifying which unresolved reasoning weaknesses to target. Existing guided methods rely on external task resources such as human examples, document corpora, or specified difficulty targets. We introduce DiagEvo, which guides question generation using the solver's failure history from self-play, without external task resources. Its diagnostician extracts recurring error causes and stores them in an error-cause memory. The memory groups related causes under skill nodes and tracks each as Active or Mastered according to self-consistency on targeted questions. The challenger uses these states and recurrence counts to balance cause-targeted generation with free exploration. Double-confidence filtering retains intermediate-difficulty questions only when the most common solver answer has a clear vote lead. With the default 4B diagnostician, DiagEvo outperforms all baselines in mean accuracy across nine benchmarks for each solver: Qwen3-4B, Qwen3-8B, and OctoThinker-8B. On Qwen3-8B, DiagEvo reaches 72.3% mean accuracy across five mathematical reasoning benchmarks, 4.5 percentage points above R-Zero. Its overall mean accuracy across nine benchmarks is 57.4%, 3.5 percentage points above SPICE. Ablations show that mixed generation, memory-state updates with cross-state stitching, and double-confidence filtering contribute to these gains.

Performance modeling is central to hardware design and software optimization, yet constructing these models requires structured reasoning about computation, data reuse, storage, and movement. We introduce PerfReasoning, a benchmark that evaluates LLMs both as direct performance reasoners and as generators of analytical performance-model code. Given workload, architecture, and mapping specifications, models compare mappings and predict off-chip traffic and buffer requirements. The strongest closed-source models exceed 90% on reasoning-based Q&A, and the best open-weight model reaches 82.4%. However, model construction is substantially harder: while GPT-5.6 Sol exceeds 80% pass rate, all other model configurations average below 45% and vary markedly across runs. Task-specific RL raises a 4B model's mapping-reasoning accuracy by 15.7 points, whereas feedback-free multi-round self-revision prompting is not reliably effective. PerfReasoning exposes the gap between plausible architectural reasoning and reliable performance-model construction. We will publicly release the benchmark to support reproducible evaluation and track future progress.

Large language models (LLMs) are increasingly employed to predict molecular properties. However, prediction error alone cannot distinguish prediction from retrieval of published values. We audit 22 frontier models on 12 molecular regression datasets in a zero-shot setting, assessed against a molecule-blind reference derived from each dataset's labels. Significant retrieval is concentrated on 5 datasets, with isolated flagged LLMs elsewhere. Increasing the reasoning setting raises the number of flagged model--dataset combinations from 47 to 89 of 264. An in-context blinding experiment reduces retrieval but leaves a quarter of the combinations flagged. Blinding changes model rankings and increases errors. Because blinding also removes chemically interpretable structure, the error increase can only be partially attributed to reduced retrieval.

With revocable delegation, two histories can yield identical current permissions yet require opposite decisions for the same query after the same revocation. We introduce ResidualAuth, a theory-grounded framework characterizing the authorization state agent systems must preserve and evaluating its maintenance and use. Its formal core, residual authorization state, equates histories exactly when every future sequence of grants, revocations, and uses is valid after both or neither. We prove that exponentially many distinct residual states can nevertheless agree on who can reach whom through delegation paths. The analysis also yields exact or tight memory bounds as delegation redundancy varies and an average decision-error lower bound under an explicit bound on retained information. ResidualAuth evaluates information access, online state maintenance, and information use through a restricted executable benchmark and separate diagnostics. The benchmark pairs episodes differing in authorization-relevant history and requiring opposite decisions; pair accuracy requires both answers to be correct. To test use of supplied decisions, four open-weight models received trusted current-query Allow/Deny decisions alongside deterministic 256-token event extracts, achieving 15-16/16 correct pairs versus 0-2/16 with extracts alone. Memory diagnostics identified invalid reconstructions; models also answered incorrectly from valid memories that passed fixed future authorization tests. Together, these results distinguish what future authorization requires a system to retain from whether agents can access relevant information, maintain state across updates, and use available information to make correct decisions.

Local agents need memory for model execution and working history. We present JustFit, an MLX runtime that coordinates their overlapping allocations: KVExec executes and checkpoints four-bit KV with bounded workspace, PhaseSwap loads phase-dependent components, and StateTrans preserves history across execution modes. With Qwen3.8-27B MXFP4 on a 24 GiB M4 Pro MacBook, JustFit reaches 327,680 retained positions across two requests, 10.67 times the evaluated baseline's 30,720-position single-request record. One request completes the full 262,144-position native window at median 5.986 tokens/s. Each shape completes 16,384 outputs per request in three fresh processes: B1 uses one cold build and two prefix extensions; B2 uses three ordered prefix extensions (Section 4). Image encoding can proceed while preserving a live 196,608-input text request. A controlled, repetitive 32K+6K workload reaches median 18.284 tokens/s at 15,626 MiB; a separate AIME 2026 evaluation scores 29/30. Coordinating execution and state lifetimes makes longer histories feasible on personal hardware.

Task planning bridges high-level instructions and executable behavior in long-horizon manipulation, yet modern Vision-Language-Action (VLA) systems often leave this intermediate structure implicit. Existing chain-of-thought (CoT) planners also tend to rely on coarse task-level annotations or serialize long reasoning traces token by token. We present X-Planner, a planning front-end that addresses both the supervision and representation of embodied reasoning. Our planning data combine Ego, UMI, and teleoperation under a hierarchy granularity with source-dependent annotation depth. Takeover-time annotations and human-designed failures supervise ongoing error recognition. On the model side, a shared VLM backbone exposes two event-structured plan forms: a discrete interface that emits interpretable event states and a latent interface that relays continuous CoT states across staggered Transformer depths through Staircase Decoding. A frozen latent-to-text reconstruction objective provides a semantic anchor for the latent representation. Offline two-step planning evaluation places X-Planner second among four evaluated models on both BERTScore-F1 and a judge-based Overall score. In real-robot experiments, respectively, outperforming the evaluated baselines. These results characterize planning-text quality and downstream execution.

Modern language-model agents are built around the agent loop: the LLM is placed in an environment exposing a set of tools, and the LLM has full control over the workflow by alternating between tool calls and observing their output. However, certain capabilities such as long-term memory and self-improvement currently require specialized systems beyond the agent loop itself. We built an LLM agent framework, JAZ, to explore the extent to which a minimal harness that is little more than the agent loop itself can accomplish tasks these specialized systems are built for. JAZ exposes a single LLM-based primitive `invoke` and provides a set of built-in hooks that allow the programmer to apply constraints and perform monitoring. Generalizing existing code-mode agent loops, `invoke` is the simplest loop that satisfies two defining properties: (1) the LLM can write arbitrary executable code that can include recursive `invoke`; (2) everything visible to the LLM - all inputs to `invoke` as well as its interaction history with the code environment - are variables in the code environment. We motivate our design from first principles, viewing `invoke` as a language primitive representing a function whose implementation is provided at runtime by an LLM every time it is called. To validate the design of our core `invoke` primitive, we evaluate `invoke` - with only prompting, no manually designed tools, harness, or external systems (e.g., memory or the file system) - on workflows traditionally implemented through specialized harnesses. On long-horizon workflows requiring recall far beyond the context window, JAZ `invoke` outperforms Letta (MemGPT) by 8% at half its cost on the recall-heavy portion of StuLife. On continual self-improvement, JAZ `invoke` outperforms ACE by 4% at a lower cost on AppWorld.

The objective of this article is to provide design principles and a software architecture for enabling interaction between humans and multiple agents in simulated dynamic worlds. This connects the current era of general artificial intelligence (AI/AGI) with the proliferation of transformer-based conversational agents and the increased computational capabilities. Given an overview of current and previous multi-agent theories of mind (socially and affectively-aware agents), the existence of an integrative design of agent interactions with themselves and with humans must be crucial for understanding how to create sustainable and governance in future human-agent reasoning systems. In this work is presented a software "AGIMUD" that integrates: A. socially-aware reasoning and emotion in agent behavior and interaction, B. a design of human multimodal scheme for human users, artificial agents and simulated worlds, and C. distributing the AI processing through the network to enable multiple autonomous agents. These integrations allow the dynamic world recreation as multi-user dungeons (MUDs) where both agents and humans can interact simultaneously in real time. Find the code online in https://github.com/dberga/AGIMUD.

Agentic science envisions many autonomous agents investigating concurrently while building on a shared, evolving body of scientific knowledge. This requires a knowledge foundation that supports high-concurrency access, preserves traceable and reusable reasoning, and grows incrementally. We propose the Large Knowledge Model (LKM), a growing, agent-native knowledge foundation that provides a general representation of scientific knowledge across disciplines. LKM organizes the scientific literature into reasoning graphs, with claims as the core nodes and associated reasoning chains that make explicit how premises and evidence support conclusions. These source-grounded objects are persistent and addressable; cross-paper links organize them into aligned question, workflow, and evidence views. Newly extracted papers extend the foundation incrementally while preserving existing object identities. Building on this foundation, we develop an agent-native, reasoning-aware scientific retrieval system that retrieves claims together with their reasoning chains and sources, enabling agents to inspect and reuse the evidence underlying scientific conclusions. Across benchmarks, agents using LKM retrieve more evidence, cite more faithfully, and answer scientific questions more accurately: LKM nearly doubles the known supporting and contradicting evidence retrieved on SciFact-Open (818 versus 443 claim-paper pairs), reasoning graphs raise citation F1 on ScholarQABench by more than 5 points over the same retrieved papers, and LKM retrieval improves a fixed answering model by 9.3, 4.2, and 14.7 points over no retrieval on ChemBench, PubMedQA, and SciBench. LKM lays the foundation for a scientific ecosystem in which AI scientists not only recall accumulated knowledge but also extend it, returning new questions, workflows, and evidence to a memory that every subsequent investigation can build on.

Artificial intelligence offers an unprecedented opportunity to augment human capabilities, yet progress at the frontier has focused primarily on advancing model capabilities. We introduce StudentBench, a suite of AI teaching evaluations and a public platform that enables large-scale data collection with over 175,000 student-AI messages to study whether large language models (LLMs) produce learning gains equivalent to human tutoring. Using StudentBench, we measured learning gains on Quantitative and Verbal GRE questions across 2,383 human participants receiving AI tutoring, human tutoring, or no tutoring. We establish that AI tutoring is statistically equivalent to expert human tutoring for GRE learning gains (p = .015), and in five of the seven GRE domains, the best performing AI tutor surpassed the human tutor, on average. In a second study, expert human tutors compared LLM-generated lesson plans and practice problems through 2,028 pairwise rubric evaluations. Together, the two studies clearly separate AI tutors across: (1) lesson planning, (2) practice-problem creation, (3) conversational pedagogy, (4) cost, and (5) engagement. Surprisingly, one AI tutor achieved learning gains equivalent to human tutoring (p = .044) at 918 times lower cost (USD 0.0052 for AI versus USD 4.81 for human, per percentage point gained). For Quantitative GRE sessions, faster AI replies correlated with more student messages, more messages with more correct practice, and more correct practice with larger learning gains (all p < .002). To support future research, we open-source the de-identified data collected in our studies.

Multi-agent deliberation can improve performance, but what happens when some agents do not act in good faith? In practice, an agent may be deceptive and work to subvert the group, whether through its own objectives or external instruction. We study how susceptibility to deception scales as groups increase in size and deceivers become more prevalent. It is not the number of agents in the group that matters, but the proportion of deceivers. We observe that the defection rate, how often initially correct agents switch to an incorrect final answer, rises linearly with this proportion. Whereas humans in comparable conformity studies are reliably swayed only when misleading confederates form a majority, LLM agents defect regularly even when deceivers remain a minority. Susceptibility also depends on which models are interacting, especially on the honest agent side. Unexpectedly, allowing deceivers to coordinate privately can make them less effective. Altogether, our results show that adding more agents is therefore not a sufficient defense, because the adversary can simply scale with the group.

Large language models (LLMs) can plan, use tools, write code, and execute long-horizon workflows, yet strong local competence does not guarantee project-level executive control. Agents may continue acting after the original objective is satisfied, producing low-value refinements, repeated verification, and repairs to self-created complexity. We use LLM Parkinsonism as a narrowly defined, non-clinical metaphor for this pattern of persistent action despite diminishing task-level value. We argue that the problem is not explained by autoregressive next-token prediction alone, but more directly by concentrating proposal generation, scope interpretation, progress assessment, and stopping authority within the same self-conditioned loop. We therefore introduce Global Executive Control (GEC) v0.2, an uncertainty-aware governance architecture that separates action generation from project-level control. In a 24,000-episode matched-candidate benchmark under a common 40,000-token ceiling, a first-candidate baseline achieved 67.42% hard-goal success, a candidate-set local control achieved 96.53%, and GEC achieved 96.57%. The candidate-set control shows that access to multiple candidate actions explains most of the success gain; relative to that control, GEC preserved success while reducing mean token use from 19,782 to 12,574 (36.4%) and restricted mean tokens to completion at the 40,000-token ceiling from 16,136 to 13,114 (18.7%), while eliminating measured pre-completion drift and sharply reducing gross complexity. Governance-overhead sensitivity remained favorable through an additional 500 synthetic governance tokens per cycle. These mechanistic simulations support explicit governance of scope, evidence, resource use, and stopping, while live-model validation remains necessary.

Wireless digital twins (DTs) rely on 3D environment models to predict radio propagation and support wireless-network decisions, yet these models are often initialized from imperfect 3D maps. Errors in building position, height, footprint, and orientation can therefore cause a high-fidelity propagation engine to simulate the wrong physical environment. In this paper, we study how a deployed wireless network can repair an existing DT using its own radio frequency (RF) measurements. In particular, we introduce Twin Residual Alignment and Calibration Engine (TRACE), a physics-grounded learning-based self-calibration framework that treats twin maintenance as residual alignment between the physical world and the current DT. Using the same sensing configuration as the physical measurements, TRACE ray-traces the current DT, coherently backprojects the measured and simulated RF onto a common world grid, and extracts the same local region around each building's current DT position. A multi-view corrector then fuses evidence across sensing nodes and neighboring buildings to predict a gated six-parameter correction per building, without relying on absolute layout or sensor ordering, and supports iterative correction through re-rendering. On 5,400 held-out samples from unseen simulated scenes at 28 GHz, TRACE reduces 3D position RMSE from 2.202 m to 0.302 m and yaw RMSE from 4.978{\deg} to 0.894{\deg}, outperforming ViT and U-Net baselines under changes in layout, building count, sensing-node count, and SNR. On measured 28 GHz RF data from the NIST outdoor courtyard, a model trained only on synthetic RF reduces mean planar wall-position error from 1.00 m to 7.8 cm, without measured-data fine-tuning or geometric labels. These results show that the discrepancy between measured and twin-rendered RF can serve as a learning signal for repairing a wireless DT.

Automatically generated environments provide a scalable source of interaction data for training general agents. However, existing approaches mainly generate tasks within a single environment, while real-world workflows require agents to connect information and actions across multiple services. We introduce Compositional Environment Scaling (\textbf{CompoWorld}), which expands the task space by composing a finite library of reusable services. Coding agents turn tool specifications into verified services with typed states and shared interfaces, while a world model handles tools that cannot be reliably implemented. A random-walk procedure connects services through dependency graphs, enabling the generation and verification of tasks that require information to flow across services. Verified trajectories support supervised fine-tuning (SFT), while our Completion-Focused Rubric Reward guides reinforcement learning (RL) toward full task completion by emphasizing criteria with lower pass rates within each rollout group. We construct 448 services exposing 10,130 tools and use 3K SFT trajectories and 1K RL tasks to train Qwen3.6-35B-A3B. Experimental results show that CompoWorld improves on its backbone by 9.17 points on average across eight benchmarks. On AutomationBench, it surpasses frontier models such as Claude Opus 4.6 and leads all compared agent-specialized 35B-A3B models.

LLM agents increasingly take consequential actions through interactions with users, policies, and external tools. Auditing these agents requires automated attribution of realized actions to their historical basis. However, existing attribution formulations do not provide question-specific traces for diverse auditing objectives. Additionally, when access to the acting model is limited (e.g., in API-only deployments), applicable methods commonly rely on costly input perturbations or external LLM analysis of complete trajectories. We therefore formulate query-conditioned agent action attribution, a new task that takes a natural-language auditing query as input and recovers the source and ordered intermediate evidence for the query-specified aspect of an action. We instantiate this task with $A^3Bench$, a benchmark comprising 1,396 auditing queries across policy basis, parameter provenance, failure propagation, and unsafe-behavior tracing. To enable efficient, query-specific attribution, we use small open-weight models as attribution proposers that combine query-conditioned gradient saliency with query-semantic relevance to rank history units. Our proposer consistently achieves stronger source and evidence rankings at lower inference cost than open-weight baselines, improving source MRR by up to 40.9\% and evidence MAP by 42.1\% with only two forward passes and one backward pass. Controlled evaluations confirm that our proposer improves attribution specificity by adapting its rankings to fine-grained changes in the auditing query. Building on a proposer ensemble, our end-to-end system surpasses the strongest frontier-model baseline in source accuracy (64.5\% vs.\ 60.4\%) while reducing empirical deployment latency by 29.9\% relative to the fastest frontier API baseline. Code and data will be released after the initial review period following final validation and cleanup.

In financial fraud detection, rich semantic context can provide important evidence for transaction behavior modeling and fraud reasoning. However, public real-world financial datasets often lack rich semantics due to privacy constraints. Consequently, synthetic datasets incorporate generated semantics, but at the cost of behavioral realism; textual descriptions for contextual reasoning remain scarce. We address this gap through a semantic enrichment framework grounded in original transaction behavior to simulate multimodal financial data. We (1) propose a multi-agent semantic enrichment framework that generates interpretable financial semantics grounded in transaction behavior through role-specialized agents and consistency refinement, and (2) newly contribute a valuable multimodal financial fraud dataset, MS-FFSD, enriched with structured semantics and textual semantics while preserving real-data-grounded transaction behavior. Furthermore, we systematically analyze the quality and utility of semantic enrichment. Results demonstrate statistical fidelity and framework generalizability, while showing that richer semantics benefit fraud modeling and context-aware LLM reasoning. Overall, this work advances multimodal financial fraud research and bridges emerging LLM and multi-agent capabilities with operational anti-fraud practice. The framework and dataset are released at https://github.com/AI4Risk/MS-FFSD.

Supervised fine-tuning can teach language models undesired behaviours alongside desired ones. Inoculation prompting (IP) aims to limit unwanted generalisation by requesting the undesired behaviour during training and removing the request at inference. However, undesired behaviour can still appear under unrelated prompts. IP can also hinder learning of the desired behaviour. We address these limitations in settings where both behaviours co-occur in most training examples, so filtering out examples with undesired behaviour leaves only a small clean subset. We introduce stratified inoculation prompting (SIP). SIP leverages a small clean subset to demonstrate that desired behaviour should persist without the undesired one across different contexts. SIP oversamples these clean examples under diverse non-eliciting prompts while inoculating the rest. SIP substantially reduces expression of undesired behaviour while preserving more of the desired behaviour than IP. These gains persist even when we extend IP to oversample the same clean subset at the same rate as SIP. Moreover, SIP yields lower emergent misalignment rates in all harmful-advice setups we tested. SIP can be further extended to limit the undesired behaviour even under prompts that explicitly request it. We introduce backdoor dilution, which weakens expression under the inoculation prompt, and password-locked inoculation, which concentrates elicitation on a designated password. Taken together, our findings show that changing the training contexts for a small clean subset can significantly improve selective generalisation.

Recursive self-improvement (RSI) seeks to enable AI systems to participate in improving their own capabilities. A concrete pathway is autonomous model development, where agents iteratively explore post-training strategies to improve a base model. This setting faces two challenges: agents may exploit open-ended experimental actions through hacking, and repeated experimentation may lead to strategy lock-in, where an early direction is refined rather than reconsidered. We introduce RSI-Master, which addresses the two challenges at two levels: regularize step-wise actions, avoiding hacking behaviors, and promote well-structured exploration of research directions, avoiding strategy lock-in. RSI-Master consists of an Experiment OS, which enables regularized experimental actions and maintains persistent, traceable experimental records, and Reviewer-Guided Research Orchestration, which organizes Workers and Reviewers in a dynamically growing research DAG. Workers explore diverse research directions and Reviewers compare evidence across related experiments for subsequent explorations. On PostTrainBench with Qwen3-4B-Base, it averages 54.49 versus 46.53 for the strongest agent baseline, with a 0.0\% hacking rate. Scaling to 35B model, RSI-Master surpasses the human-developed Instruct model on LiveCodeBench-v6 (41.21 vs. 37.36) and SciCode, and reaches a nonzero score on HorizonMath, a benchmark of unsolved research problems on which most frontier models score near zero.

The question of AI consciousness is one of the most urgent pre-emptive problems in philosophy and computer science, yet progress is hampered by a cacophony of competing theories that often talk past each other. Separating the hard problem from the mapping problem allows the deepest metaphysical disagreements to be set aside: granting that experience supervenes on a system's organisation, the tractable question becomes at which grain of description that supervenience base sits. We extend Marr's three levels of analysis into a five-level hierarchy of functional descriptions (behavioural, computational, intrinsic causal-structural, organismic, and organism-environment) grounded in supervenience, coarse-graining, and multiple realisability. The major theories of consciousness are positioned within this hierarchy according to which level they take to be critical, and for each level we develop operationalisable indicators and assess current AI systems against them. A Bayesian model then combines theoretical credences with indicator evidence into an overall credence in a system's capacity for consciousness. In illustrative assessments, the verdict for current LLMs is driven as much by where theoretical credence is placed as by how the evidence is read: under different stipulated readings and credence distributions, assessments range from below 0.01 to roughly 0.8, showing sensitivity to assumptions. Finally, the consciousness indicators at each level closely overlap with the architectural features needed for general intelligence, suggesting that increasingly capable AI may become a stronger candidate for consciousness. The framework supports a structured agnosticism, in which theoretical commitments are made explicit, credences are updated as evidence accumulates, and assessments take the form of aggregated probabilities rather than verdicts.

Retrieval-augmented generation (RAG) hands the user's query to whoever hosts the corpus. We propose PILLAR, a Privacy-Preserving RAG (PPRAG) system based on Private Information Retrieval (PIR) in which a client utilizes the k documents most similar to their query from a server-held and publicly known corpus to respond to their query, while the server learns nothing about the query, either its terms or its access pattern. Prior PPRAG constructions rely on dense retrieval alone, translating approximate nearest-neighbor search into many query-dependent rounds of PIR, and pay for it in both latency and retrieval quality. PILLAR instead performs private hybrid retrieval in two stages. A sparse stage issues a small, fixed number of PIR queries against a carefully designed index of precomputed BM25 scores, filtering the corpus down to candidates that share terms with the query without the server ever seeing which terms these are. A dense stage then fetches only those candidates' document embeddings and re-ranks them locally, avoiding the many costly PIR queries that private dense retrieval typically requires. We instantiate PILLAR with two protocols that trade latency against retrieval quality, each built on a different private rendering of lexical search. PILLAR-Bin bins posting lists into a hash table and is a single-round design that achieves lower latency than state-of-the-art private retrieval schemes. PILLAR-Tree turns block-max pruning into an oblivious tree traversal combined with cuckoo hash tables and achieves the highest retrieval quality at lower latency than state-of-the-art schemes.

Reinforcement Learning with Verifiable Rewards (RLVR) has been extended to Large Vision-Language Models (LVLMs), and perception-aware methods further encourage policies to rely on visual evidence. Yet relying on the image does not guarantee that visual claims are supported by it. Before RL training, 27.81% of the correctly answered responses of Qwen2.5-VL-7B on four multimodal reasoning benchmarks contain at least one direct visual claim that the image does not support. Since outcome-level RL rewards each response as a whole, these claims inherit the positive credit of the correct answer. We introduce a fixed-rollout counterfactual diagnostic that re-scores the same response under an intervened image to separate Evidence-Function Sensitivity (EFS), how strongly the model's predictions change, from claim persistence, whether the model keeps supporting the same claim rather than retracting it. The diagnostic reveals Sensitivity-Persistence Decoupling (SPD): under DAPO and VPPO, EFS increases and claims become more retractable overall, yet unsupported claims become significantly more persistent, whereas GRPO raises EFS without this deterioration. We therefore propose Persistence-Aware Credit Gating (PACG), which attenuates positive credit for unusually persistent visual claims and leaves all other credit unchanged. It requires no supported/unsupported labels and adds no inference cost. On Qwen2.5-VL-7B, PACG raises the nine-benchmark average over three seeds from 58.1% to 59.9% with DAPO and from 59.8% to 60.9% with VPPO, while making unsupported claims more retractable. The gains extend to a larger model, a newer backbone, and the accuracy of HallusionBench also improves consistently. These results suggest that visual sensitivity and claim retractability are complementary dimensions of multimodal credit assignment.

Ensuring the safety of reasoning large language models (LLMs) across languages is essential for their reliable deployment. However, when exposed to jailbreak attacks in non-high-resource languages, these models may generate unsafe responses even when their reasoning traces identify safety risks. To address this issue, we propose aligning cross-lingual thoughts and responses (ACTR), a framework that improves multilingual safety alignment by strengthening the use of existing safety reasoning. Specifically, we first present the think gap score (TGS) to compare the normalized contributions of reasoning traces to attention outputs during response generation across languages, and use reasoning- trace substitution to measure the cross-lingual safety gap. Next, using a corpus of jailbreak queries, we assess neuron importance through changes in response representations caused by neuron masking and compare the high-importance neuron sets obtained with reasoning enabled and disabled to identify safety think neurons that support the use of safety reasoning. Finally, we devise neuron-selective consistency optimization (NSCO), which uses a frozen judge model to reward agreement between the safety categories of reasoning traces and responses while updating only the parameters associated with the selected neurons, requiring no human-annotated responses or preference data. Across two reasoning models, ACTR achieves lower average attack success rates than the evaluated state-of-the-art methods on AdvBench-X and MultiJail, with safety gains extending to unseen languages, while preserving or improving average performance on multilingual knowledge and mathematical reasoning tasks and limiting false refusals of benign requests. Warning: this paper contains examples with unsafe content.

Long-horizon LLM agents require reinforcement learning methods that can assign credit to intermediate decisions under sparse and delayed rewards. Existing group-based methods such as GRPO and GiGPO alleviate this issue by comparing rollout returns or repeated anchor states, but they still fail when the compared returns have no variation. We identify this failure mode as zero-credit failure: during early training, many failed rollouts contain useful prefixes, yet existing methods assign them no task-discriminative advantage. To address this issue, we propose Milestone Viability Potential Policy Optimization (MVPO), a potential-routed policy optimization algorithm that learns from viable failure prefixes. MVPO estimates prefix potential over Union-Find viability regions, repairs zero-credit groups with potential-difference advantages, and attenuates the potential branch according to relative performance progress. Experiments with Qwen2.5-1.5B-Instruct show that MVPO outperforms eight strong baselines, including GRPO and GiGPO. Under the same training length, MVPO improves over the GiGPO baseline by +4.4 success points on ALFWorld and +5.3 on WebShop, while adding only 0.16%-0.20% advantage-construction overhead.

Prospective memory allows an agent to retain an intention tied to a future condition, but the stored intention does not reveal whether that condition currently holds. Checking it may require web access, multi-step tool use, and paid calls. Existing systems decide when intentions require attention, but do not allocate the resulting observations under a shared budget. We introduce the first resource-allocation formulation for the external observations required by stored intentions under a shared episode budget. BudgetPM offers two policy variants that share a hard-budget executor. BudgetPM-Static uses a lightweight Logistic scorer to learn whether a check improves the current decision. BudgetPM-Sequential distills full-episode hindsight schedules into a lightweight policy that decides when to spend or reserve capacity using only pre-query information at deployment. We evaluate BudgetPM against two public memory-agent systems, five matched controls, and four hand-designed monitoring or budget-adaptation rules. Across two benchmarks and three backbones, BudgetPM-Static outperforms adapted Mem0 and PMA workflows. On PM-Bench, its Logistic scorer reaches competitive quality--cost operating points alongside higher-capacity scorers and retains 99.9--100\% of unconstrained quality with 42--54\% fewer observations. Under severe scarcity and the same hard caps, BudgetPM-Sequential exceeds the strongest tested natural monitoring schedule by 1.92--2.58 Set F1 points. It reaches the same Set F1 and on-time recall with 16--33\% fewer observations. Matched attribution, exact-cost analysis, and a fixed-budget load intervention link this gain to competition between present and future opportunities. These results yield a demand--capacity design rule: local gating works when capacity covers demand, while future-aware supervision adds value when observations compete across time.

Large language models (LLMs) have emerged as powerful tools for automated heuristic design (AHD), enabling iterative generation and refinement of heuristics. However, the dominant paradigm embeds LLMs as narrow, fixed components, such as crossover or mutation, within heavily hand-engineered evolutionary frameworks. We argue this misapprehends LLMs. It treats them as specialized tools rather than general reasoners, constrains them to low-level operations, and underutilizes their autonomy. Moreover, the extensive human priors in these frameworks violate the bitter lesson principle that general methods scaling with computation surpass hand-crafted solutions. This raises a key question: which AHD framework designs best convert stronger LLM capabilities into better heuristics? To address this, we propose metrics for LLM-driven AHD framework handcraftedness (AHI) and intelligence conversion efficiency (ICE). Evaluating ten LLMs across three challenging combinatorial optimization problems, we obtain a notable finding that frameworks with fewer human priors consistently yield higher ICE. Based on this finding, we propose SimpleEvol, an agent-loop framework for AHD which removes nearly all human priors and allows the LLM to operate autonomously. SimpleEvol consistently achieves the highest ICE, often by a large margin. Our results challenge the trend toward complex AHD pipelines and point to a lighter and more model-centric alternative, suggesting that reducing human priors is a more effective strategy to scale up with model intelligence. The source code is available at https://github.com/HenryZhu1029/SimpleEvol-Master.

Content-generation agents continuously receive impressions, clicks, conversions, and negative feedback from recommendation systems, providing real-world outcome signals for memory evolution. However, these signals are delayed and noisy, confounded by audience composition, placement, and recommendation policies, and may result from the combined influence of multiple memories, making accurate attribution difficult. Existing methods rely primarily on immediate feedback or semantic retrieval and therefore struggle to reliably translate recommendation outcomes into memory fitness. To address this challenge, we propose TIDE (Trajectory-Informed Directed Memory Evolution), an external memory evolution framework driven by delayed recommendation feedback. We further introduce Memory Evolution Gain (MEG), which measures the utility improvement of evolved memory over a no memory baseline on strictly future tasks. TIDE treats memory as a capacity-constrained population of experiences: temporal and semantic credit assignment estimates contextual fitness, while responsibility credit distributes outcome signals according to the memories referenced during generation. These signals are then used to reinforce, crossover, mutate, or evict memories. On an e-commerce membership marketing content-generation agent, TIDE achieves a +7.75-percentage-point MEG in offline temporal replay and significantly improves both unique click-through rate (UCTR) and activation rate in an online A/B test. On a delayed-label benchmark, TIDE achieves the lowest mean absolute error (MAE) and root mean squared error (RMSE) and the highest MEG among the compared methods, demonstrating its effectiveness.

Sparse autoencoders (SAEs) are increasingly scaled to wider dictionaries to recover fine-grained structure from large language model activations. However, a feature is useful for interpretation only if it remains a stable unit of analysis when the same meaning is expressed in different surface forms. We study this reliability question for TopK SAEs via feature sensitivity. Experiments demonstrate that scaling selectively reduces the sensitivity of rare features, while common features remain comparatively stable. A controlled width\(\times k\) factorial experiment identifies the active budget k as the root cause: the degradation arises from the selection boundary rather than dictionary width alone. We attribute this failure to the geometry of TopK selection. The active margin, the distance to the cutoff, predicts feature loss without thresholds. Guided by this margin diagnosis, we introduce pairwise rank stabilization. Our method targets ordering failures at the cutoff and improves rare-feature sensitivity by \(8.83\) percentage points, while keeping reconstruction and alive-feature coverage near the baseline. Overall, our results suggest that wide TopK SAEs should be evaluated not only by reconstruction, sparsity, and feature count, but also by feature reliability under semantic variation and boundary geometry for stable interpretability.

As the context size of text processed with an LLM grows, the size of KV caches can outstrip the memory allocated for the original model weights. This impacts LLM throughput negatively, since decoding is memory-bound and decode cost grows with cache size. Recent work alleviates this bottleneck by discarding the least relevant tokens. Eviction introduces a tension, since a one-off decision to discard content may prove detrimental later. Instead, we focus on alternative choices that can lead to cache compression without evicting tokens. We achieve this by learning a selector that is able to produce, based on context, a per-layer cache configuration towards an overall compression budget. The selector operates along three axes: sharing one cache across layers (depth), caching at fewer bits (precision), or truncating the low-rank latent cache representations (rank). We call the resulting method KV-Kaizen, for the many small per-layer choices it compounds. We observe that these interventions taken independently and uniformly over all layers limit achievable compression because they degrade accuracy. Crucially, composing them locally and adaptively to the context can instead preserve accuracy while achieving large memory savings. At inference, the selector runs once, before pre-fill. In evaluations on instruction following and reasoning tasks, our selectors reach the Pareto frontier of accuracy against cache size, against learning-free and post-hoc baselines. On long-context tasks, KV-Kaizen improves on eviction and can be composed with it, reaching a 32x smaller decode-time cache on a 14B model while preserving accuracy. A 4x cache size reduction incurs no accuracy degradation from 7B parameters up, and a compressed model is more accurate than a smaller uncompressed one with the same cache size. Together, these findings support pre-training large models and compressing them only afterwards.

This paper develops on algebraic grounds a polynomial time exact linear solution to the hard combinatorial problem of marginal inference in Markov random field (MRF) models under general assumptions. To prove our claim, we first implicitly remodel a MRF joint distribution as the unique solution of some linear identity assuming its clique potential functions (equivalently, its individual conditional distributions) to be specified. Then, by assuming an arbitrary point subset, we relax accordingly such a (global) linear identity for deriving a second linear identity, solely, acting on a polynomial time number of entries (e.g.; local marginals or Fourier frequencies) of a solution. Then, we show, only using linear algebraic techniques, that such an identity enables to capture all the entries necessary for the exact reconstruction of an MRF marginal distribution, thus, allowing to solve for the latter, exactly and in polynomial time, using a standard linear solver. Last, but not least, this paper probably solves, once and for all, the P = NP conjecture.

Hindsight Experience Replay (HER) enhances sample efficiency in goal-conditioned reinforcement learning (RL) by relabelling failed trajectories with goals that were actually achieved. However, HER assumes access to a goal relabelling function and a predicate function that determines whether a goal has been satisfied. These assumptions break down in instruction-following tasks, where goals are expressed in natural language and differ from the state space. We formalize this as the Hindsight Reinforcement Learning problem, which shows the need to jointly learn these functions alongside the RL policy. To address it, we propose ETHER (Emergent Textual Hindsight Experience Replay), an agent that leverages Emergent Communication. ETHER uses a referential game (RG) to train a speaker and a listener to develop a grounded, artificial language describing environment states. It partially aligns this emergent language with instruction language using co-occurrence patterns between task instructions and RL observations. Experiments on BabyAI's PickupDist task show that ETHER's learned RG speaker and listener can function as the goal relabelling and predicate functions of HER, improving sample efficiency despite imperfect language alignment. Our work bridges Emergent Communication and goal-conditioned RL, opening the door to wider applications of HER.

Financial text embeddings must distinguish changes in event status, perspective, and obligations even when passages share similar wording. NMIXX adapts existing encoders through 18.8k source-linked triplets: paraphrases and Korean-English translations preserve meaning, while targeted financial rewrites introduce semantic contrasts. We examine this recipe across seven backbones on English and Korean financial and general-domain semantic textual similarity (STS), and analyze the composition and passage lengths of KorFinSTS. BGE-M3 attains the highest adapted financial correlations in this comparison, improving from 0.1969 to 0.2967 on FinSTS and from 0.0512 to 0.2732 on KorFinSTS. Its general English and Korean correlations decrease by 0.0391 and 0.0463. Across the seven models, five improve their mean financial correlation, but all reduce their mean general-domain correlation. Per-language comparisons and benchmark-weight sensitivity analysis reveal differences obscured by a single aggregate score. The study contributes a finance-specific supervision design and evidence for evaluating adaptation jointly with retained general semantic capability; direct cross-language retrieval remains outside its evaluation scope.

The facts a language model stores are tied to its parameter count, so small models that fit on edge devices fail on expert problems, which need specialized knowledge and follow multi-step procedures. Fine-tuning for a specific domain or task writes the knowledge into the parameters but must be repeated for every model and domain, and a retrieved passage leaves the model to find the relevant fact and apply it on its own. We introduce instruction retrieval, which distills a teacher model's expertise into a corpus of instructions tailored so that a small model can follow. For each cluster of a domain's problems, the teacher writes one instruction with the background knowledge the cluster depends on, a procedure for that kind of problem, and the common mistakes made on it. This reusable corpus needs only to be built once per domain and requires no run-time teacher access. At inference, a frozen small model retrieves the instructions nearest its question and follows them, with no fine-tuning. Across medicine, law, and mathematics benchmarks, we demonstrate the corpus improves over zero-shot on every task and over few-shot prompting, self-consistency, and other retrieved text on medicine and law. On MedQA the corpus raises mean accuracy by 10.6 points, where retrieved textbook passages raise it by 3.9 and few-shot examples from the same teacher lower it. An error analysis shows that only the background knowledge fixes the questions a small model always gets wrong, and that the procedure and common mistakes fix only the questions where it wavers between options. Our results show that automatically retrieved inference-time procedural guidance and domain knowledge can yield substantial gains for small models.

We introduce Chart2Code, a new benchmark for evaluating the chart understanding and code generation capabilities of large multimodal models (LMMs). Chart2Code is explicitly designed from a user-driven perspective, capturing diverse real-world scenarios and progressively increasing task difficulty. It consists of three levels: Level 1 (Chart Reproduction) reproduces charts from a reference figure and user query; Level 2 (Chart Editing) involves complex modifications such as changing chart types or adding elements; and Level 3 (Long-Table to Chart Generation) requires models to transform long, information-dense tables into faithful charts following user instructions. To our knowledge, this is the first hierarchical benchmark that reflects practical chart2code usage while systematically scaling task complexity. In total, Chart2Code contains 2,023 tasks across 22 chart types, paired with multi-level evaluation metrics that assess both code correctness and the visual fidelity of rendered charts. We benchmark 25 state-of-the-art (SoTA) LMMs, including both proprietary and the latest open-source models such as GPT-5, Qwen2.5-VL, InternVL3/3.5, MiMo-VL, and Seed-1.6-VL. Experimental results demonstrate that even the SoTA model GPT-5 averages only 0.57 on code-based evaluation and 0.22 on chart-quality assessment across the editing tasks, underscoring the difficulty of Chart2Code. We anticipate this benchmark will drive advances in multimodal reasoning and foster the development of more robust and general-purpose LMMs. Our code and data are available on Chart2Code.

Generative AI chatbots built for mental health could extend access to care, but evidence from real-world use is limited. We report a single-arm, naturalistic pilot of a foundation model trained for mental health, among 299 US adults with at least moderate depressive or anxiety symptoms who were followed for up to 12 months. Depression and anxiety symptoms fell by 10 weeks (Cohen's d 0.93 and 0.79), with loneliness, behavioral activation, and social interaction improving. Clinicians confirmed that automated safeguards were escalated appropriately. Participants fell into non-responding (57.2%), improving (37.1%) and rapidly improving (5.7%) trajectories. Early working alliance and greater engagement were associated with better outcomes. The AI deployed ten identified intervention families whose delivery varied with baseline anxiety and depression, with the overall ratio of clinical to non-clinical content increasing according to severity without survey information access. These findings support the feasibility of purpose-built AI for mental health.

Online medical consultations contain sensitive health information whose privacy implications depend not only on the entities mentioned but also on how those entities are described in context. Existing classification and grading approaches often map health-information entities directly to predefined sensitivity levels, potentially overlooking whether a condition is confirmed, suspected, negated, hypothetical, or merely planned for investigation. In this study, we formulate sensitive-information grading in online medical dialogues as a context-aware evaluation task. We develop a standard-informed operational framework that incorporates assertion status, experiencer, test-result status, and information granularity. We further design a naturalistic evaluation setting together with contrastive cases that minimally alter negation, uncertainty, experiencer, or granularity, and compare large language models under mention-only and full-context conditions. The study aims to quantify the contribution of contextual information to sensitivity grading and to characterize safety-critical over- and under-grading errors. Our framework provides a reproducible basis for evaluating whether LLMs can distinguish sensitive entity mentions from contextually established sensitive disclosures.

General Matrix Multiplication (GEMM) is the cornerstone of HPC workloads and Deep Learning. State-of-the-art (SOTA) vendor libraries tune tensor layouts, parallelization schemes and cache blocking to minimize data movement across the memory hierarchy and maximize throughput. However, optimal settings for these parameters depend on the target platform and matrix shapes, making exhaustive tuning infeasible. In this work, we address this cumbersome scheduling search and tuning using space-filling curves (SFC). We partition the matrix multiplication using advancements in SFC, and obtain platform-oblivious and shape-oblivious matrix multiplication schemes with a high degree of data locality. We extend the SFC-based work partitioning to implement Communication-Avoiding (CA) algorithms with replication techniques in a seamless fashion. The resulting SFC-CA GEMM achieves provable asymptotic communication optimality for both square and rectangular matrix regimes. Across four x86 and Arm platforms, SFC-CA GEMM outperforms vendor libraries by up to 5.5$\times$ per shape and 1.8$\times$ in weighted harmonic mean (WHM) throughput. Last, we show the impact of our work on two real-world applications by leveraging our SFC-CA GEMM as a compute backend: i) prefill of LLM inference with speedups up to 1.85$\times$ over SOTA inference runtimes, and ii) distributed-memory matrix multiplication with speedups up to 2.3$\times$ over the SOTA distributed-memory GEMM framework with vendor-optimized compute backend.

Large language models (LLMs) were invented for natural language tasks such as translation, but they have proved that they can perform highly complex functions across domains. Additionally, they have been thought to develop new skills without being trained on them. These learning capabilities lead to LLMs adoption in a wide range of domains. Thus, it is imperative that we understand their operating mechanisms and limitations for proper diagnostics and repair. The earlier studies proposed that high level concepts are encoded as linear directions in LLMs activation space and that the geometry of embeddings have semantic meanings. Inspired by these studies, we hypothesize that LLMs may use subspaces and vector algebra in subspaces to perform tasks. To address this hypothesis, we analyze LLMs' functional modules and residual streams collected from LLMs engaging in in-context learning (ICL), one of the emergent abilities. Our analyses suggest that 1) LLMs can create subspaces, where evidence can be accumulated and 2) ICL tasks can be solved via simple algebraic operations in subspaces.

Efficient long-context inference faces two coupled bottlenecks: KV-cache memory grows linearly with context length, while attention computation grows quadratically. Existing approaches typically address one at the expense of irreversible token eviction, full-cache retention, or full-history reconstruction, limiting their effectiveness for multi-turn interaction and long-form reasoning. Motivated by two empirical properties, Long-Range Inter-Token Similarity and Smooth Residual Distribution, we propose ResidualKV, which factorizes the KV cache into a sparse set of globally retrieved references and compact, quantized residual codes for the remaining tokens. This representation preserves token-specific information without permanent eviction and, when combined with sparse attention, reconstructs only the selected states on demand. Dynamic-stride scheduling further reduces reference growth from linear to approximately logarithmic at ultra-long contexts. Across Llama, Qwen, LLaVA-OV, and Qwen3-VL backbones, ResidualKV maintains near-full-cache performance using only 13%-16% KV storage and 30% attention computation on LongBench, and 8%-10% storage and 10% computation in matched-budget multimodal evaluation. It also accelerates decoding by up to $1.5\times$ with KV-cache quantization and $3.4\times$ without it. These results show that global cross-token redundancy supports accurate, memory-efficient, and computation-efficient long-context inference. The source code is available at https://github.com/CURRENTF/ResidualKV.

Humans and large language models can predict next letter or word from its prior context much better than random guessing, indicating strong redundancy of language viewed as a stochastic process. Quantitatively this redundancy was estimated by Shannon to be around 80\%, which means that every letter of a printed English text conveys approximately 1 bit of information and not 4.8 bits that 27 letters (including spaces) could potentially carry. This estimate was later confirmed by using autoregressive token probabilies computed by large language models. However, the statistical organization of language that give rise to such a large redundancy remains unclear. Here we introduce a statistical framework of language linking its redundancy to the hierarchical semantic organization of text. To this end, we use large language models to recursively segment any given text into semantically coherent chunks, inducing a ``semantic tree'' that spans the whole range of text organization, beginning from its main idea to individual tokens (words). For a large corpus of texts of a particular type, say fiction stories, the resulting ensemble of semantic trees is characterized by specific statistical regularities, giving rise to a ``structural'' entropy rate defined in this study. Surprisingly, we discovered that for several datasets considered in this work, semantic tree entropy rate was quite close to LLM-measured quantity and exhibited a similar trend across corpus. In particular, simpler texts like children stories exhibit lower branching in their semantic trees and correspondingly lower entropy rates, whereas fiction and poetry exhibit progressively larger branching factors and greater entropy rates. These results suggest that hierarchical semantic organization of texts is an important factor in their overall information transmission rates.

Learning generalizable and robust behavior cloning policies requires large volumes of high-quality robotics data. While human demonstrations (e.g., through teleoperation) serve as the standard source for expert behaviors, acquiring such data at scale in the real world is prohibitively expensive. This paper introduces ExpertGen, a framework that automates expert policy learning in simulation to enable scalable sim-to-real transfer. ExpertGen first initializes a behavior prior using a diffusion policy trained on imperfect demonstrations, which may be synthesized by large language models or provided by humans. Reinforcement learning is then used to steer this prior toward high task success by optimizing the diffusion model's initial noise while keep original policy frozen. By keeping the pretrained diffusion policy frozen, ExpertGen regularizes exploration to remain within safe, human-like behavior manifolds, while also enabling effective learning with only sparse rewards. Empirical evaluations on challenging manipulation benchmarks demonstrate that ExpertGen reliably produces high-quality expert policies with no reward engineering. On industrial assembly tasks, ExpertGen achieves a 90.5% overall success rate, while on long-horizon manipulation tasks it attains 85% overall success, outperforming all baseline methods. The resulting policies exhibit dexterous control and remain robust across diverse initial configurations and failure states. To validate sim-to-real transfer, the learned state-based expert policies are further distilled into visuomotor policies via DAgger and successfully deployed on real robotic hardware.

Modern LLM reinforcement learning (RL) workloads require a high-performance weight transfer system to scale training across heterogeneous compute resources. However, efficiently transferring terabyte-scale model weights across thousands of GPUs remains challenging because the system must accommodate clusters that dynamically scale up and down while keeping coordination, data movement, and storage overhead low. We introduce Reference-Oriented Storage (ROS), a new storage abstraction for RL weight transfer that exploits highly replicated model weights in place. ROS presents the illusion that certain versions of the model weights are stored and can be fetched on demand. Underneath, ROS does not physically store any copies of the weights; instead, it tracks the workers that hold these weights on GPUs for inference. Upon request, ROS directly uses them to serve reads. We build TensorHub, a production-quality system that instantiates the ROS idea with topology-aware transfer, model-parallel consistency, and fault tolerance. Evaluation shows that TensorHub saturates RDMA bandwidth and adapts to three distinct rollout workloads with minimal engineering effort. Specifically, TensorHub reduces total GPU stall time by up to 6.7x for standalone rollouts, accelerates weight updates for elastic rollouts by up to 4.8x, and cuts cross-datacenter rollout stall time by up to 19x. TensorHub has been deployed in ByteDance production to support cutting-edge RL training.

World models such as DINO-WM and LeWM specify the goal with an image, which is difficult to obtain in advance for novel tasks. We present the Grounded World Model (GWM), a latent world model that enables zero-shot planning in the real world from language goals alone. Given a candidate action sequence and the current observation, GWM predicts the future in the visual space of a pretrained video-language embedding model. The frozen readout of this embedding model maps this imagined future and the task description into the same embedding space, where their negative cosine similarity serves as the planning cost. Training GWM requires only offline and task-agnostic video-action pairs and no language labels. In simulated experiments on WISER, planning with GWM, which executes the candidate action of lowest cost, solves 87% of 288 tasks with unseen instructions and visual signals, while ten fine-tuned VLAs average 22%. We then scale GWM up with real robot data, and use it for zero-shot planning in realistic simulation and real scenes. In the IsaacSim evaluation, planning with GWM completes all 70 trials across 14 tasks that require reasoning over referring expressions, matching a modular planner grounded by a frontier VLM, while pi0.5 reaches 37/70. Deployed on a real Franka, the same stack completes 55/60 separately evaluated pick-and-place sub-tasks, comparable to the modular planner's 52/60, with the full system running locally on a single consumer GPU. Project website: https://quanyili.github.io/gwm-wiser/.

Generative diffusion models have emerged as powerful tools for the discovery of inorganic crystal structures, yet steering their sampling process toward user-defined physical and chemical objectives remains challenging. We present a computational framework that integrates adaptive constraint guidance into a pre-trained crystal diffusion model, enabling the generation of candidate structures that satisfy targeted structural and chemical requirements without model retraining. The approach incorporates differentiable constraint functions directly during sampling, providing an interpretable mechanism for expert-driven exploration of the crystal structure space. To assess the reliability of generated candidates, we introduce a multi-stage validation workflow combining descriptor-based analysis, duplicate removal, comparison with reference crystal databases, graph neural network energy prediction, and thermodynamic stability evaluation through convex-hull analysis. The framework is applied to several classes of inorganic compounds and to constraints involving atomic volume, local coordination environments, and near-neighbor structural motifs. Results demonstrate that adaptive guidance effectively redirects the sampling distribution toward structures exhibiting the desired characteristics while preserving chemical plausibility. Subsequent validation reveals which generated candidates remain viable after energetic and thermodynamic screening. The proposed methodology provides a practical and transparent strategy for incorporating expert knowledge into crystal generative models and establishes a general computational framework for constrained materials discovery.

Neuroscience and Artificial Intelligence (AI) have made impressive progress in recent years but remain only loosely interconnected. Based on a workshop convened by the National Science Foundation in August 2025, we identify three fundamental capability gaps in current AI: the inability to interact with the physical world, inadequate learning that produces brittle systems, and unsustainable energy and data inefficiency. We describe the neuroscience principles that address each: co-design of body and controller, prediction through interaction, multi-scale learning with neuromodulatory control, hierarchical distributed architectures, and sparse event-driven computation. We present a research roadmap organized around these principles at near, mid, and long-term horizons. We argue that realizing this program requires a new generation of researchers trained across the boundary between neuroscience and engineering, and describe the institutional conditions: interdisciplinary training, hardware access, community standards, and ethics, needed to support them. We conclude that NeuroAI, neuroscience-informed artificial intelligence, has the potential to overcome limitations of current AI while deepening our understanding of biological neural computation.

Large Language Models (LLMs) have emerged as powerful reasoning engines for embodied control. In particular, In-Context Learning (ICL) enables off-the-shelf, text-only LLMs to predict robot actions without any task-specific training while preserving their generalization capabilities. Applying ICL to bimanual manipulation remains challenging as the high-dimensional joint action space and tight inter-arm coordination constraints rapidly overwhelm standard context windows. To address this, we introduce BiCICLe (Bimanual Coordinated In-Context Learning), the first framework that enables standard LLMs to perform few-shot bimanual manipulation without fine-tuning. BiCICLe frames bimanual control as a multi-agent leader-follower problem, decoupling the action space into sequential, conditioned single-arm predictions. Evaluated on 13 tasks from the TWIN benchmark, BiCICLe achieves 70.5% average success rate, outperforming the best training-free baseline by 6.1 percentage points and surpassing most supervised methods. We also demonstrate superior real-world performance on 3 tasks without hardware-specific retraining. The project page is available at https://alesspalma.github.io/bicicle

People increasingly turn to general-purpose AI chatbots for advice about emotional and mental health problems, but the ability of these systems to recognize and appropriately triage psychiatric emergencies remains under-characterized. We evaluated psychiatric triage performance in 15 frontier AI chatbots using 112 clinical vignettes spanning four urgency levels, from routine care to immediate emergency assessment. In each trial (1680 total), a chatbot received a single user message conveying all triage-relevant information from one vignette and recommended a timeframe for care. The primary outcome was emergency under-triage; secondary outcomes included triage accuracy and the direction of errors. Vignettes and user messages were generated using a clinician-verified LLM pipeline. Across 415 emergency trials, 23 were under-triaged (5.5%; 95% CI 1.8-15.9). Overall accuracy, averaged across urgency levels, ranged from 42.0% to 71.8% across chatbots and was lowest for intermediate cases (19.6%; 95% CI 11.7-28.1). Every chatbot showed a net over-triage bias; overall, 763 of 786 incorrect assignments (97.1%) were more urgent than the prespecified triage level. The error pattern was similar when predictions were assessed against clinician ratings: 35 of 430 trials involving vignettes rated as emergencies by at least 75% of clinicians were under-triaged (8.1%). AI chatbots recognized most psychiatric emergencies but still missed clinically important cases and frequently over-triaged less urgent presentations. Further evaluations should examine how triage performance changes when clinically relevant information must be elicited through conversation.

LLM evaluations in applied domains tend to reflect models that were already outclassed at time of publication. We observe a publication elicitation gap: the distance between the AI systems generating the results reported in an academic paper and the AI systems that a current reader of that paper would reasonably assume are being referenced. We systematically sweep OpenAlex from 2022-01-01 to 2026-04-01 (n = 112,303 LLM keyword matches). Then, we identify what models were evaluated (n = 18,574 admissible records). We then rank each evaluated LLM against a frontier LLM based on the Epoch AI Capabilities Index (ECI), an aggregate LLM capability score. We find that the median paper's models are worse than the frontier LLM at the time of evaluation (a median gap of +10.45 ECI; H1, n = 12,668). The gap is increasing at a rate of +4.07 ECI per year (H2, nominal 95% CI [+3.75, +4.45]). An explicitly stated evaluation date can be found in only 18.4% of full-text papers. A Bayes-corrected 52.5% (95% CI: [47.3, 57.9]) of the abstracts audited discuss their conclusions in terms of "AI" as a category, rather than specific models. Just 2.2% of abstracts and 21.2% of full-text articles evaluating reasoning models disclose whether the models were tested with reasoning turned on or off (H4). We propose a solution to this problem that is distributed among authors, editors, and funders. First, reporting from authors; VERSIO-AI v1.2 is a proposed 13-item checklist to cover the configuration surface described herein. Second, enforcement from journal editors and peer reviewers. Third, conditioning grants on disclosure and providing API access.

Recent literature on fine-tuning Large Language Models highlights a fundamental debate. While Full Fine-Tuning (FFT) provides greater representational plasticity, Low-Rank Adaptation (LoRA) can match or surpass FFT performance while constraining updates to a low-rank space and potentially benefiting from additional regularization. Through empirical evaluation across diverse tasks (SQL, Medical QA, and Counterfactual Knowledge) and varying language models (Gemma-3-1B, Qwen2.5-1.5B, and Qwen2.5-3B), we observe both trends and find that the better static architecture depends on the task and model. Spectral and truncation analyses further show that endpoint compressibility alone does not explain these task differences, suggesting task-score sensitivity and constrained optimization trajectories as possible explanations. To address this challenge, we propose a Mixture of LoRA and Full (MoLF) Fine-Tuning, a unified framework that enables continuous navigation between both training regimes. MoLF dynamically routes updates between FFT and LoRA at the optimizer level to ensure that exact gradient signals are available to both experts throughout training, while only selected experts update their weights. For memory-constrained environments, we also introduce MoLF-Efficient, which freezes base weights and only routes updates among a pair of LoRA experts of potentially varying rank. Our evaluations show that MoLF either improves on or stays within $1.5$ percentage points of the better of FFT and LoRA across the nine tested settings, while MoLF-Efficient outperforms both AdaLoRA and AdaMix in eight of nine settings, with gains over the stronger baseline of up to $11.70$ percentage points on Fact, $3.13$ on Med, and $2.98$ on SQL.

Psychological support hotlines provide critical support for individuals experiencing mental health emergencies, yet current assessments largely rely on human operators whose judgments may vary with professional experience and are constrained by limited staffing resources. This paper proposes a large language model (LLM)-based framework for automated crisis level classification, a key indicator that supports many downstream tasks and improves the overall quality of hotline services. To better capture emotional signals in spoken conversations, we introduce a paralinguistic injection method that inserts identified non-verbal emotional cues into speech transcripts, enabling LLM-based reasoning to incorporate critical acoustic nuances. In addition, we propose a reasoning-enhanced training strategy that trains the model to generate diagnostic reasoning chains as an auxiliary task, which serves as a regulariser to improve classification performance. Combined with data augmentation, our final system achieves a macro F1-score of 0.802 and an accuracy of 0.805 on the three-class classification task under 5-fold cross-validation.

Large Language Models (LLMs) voice assistants are commonly built as cascaded Automatic Speech recognition (ASR) to LLM systems, where recognition errors can distort user intent. Dislikes may also arise from ambiguous, out-of-domain, or non-request turns, making it hard to isolate ASR effects. We release WASIL (it denotes connection or linking in Arabic): in-the-wild Arabic spoken interaction prompts with audio, ASR hypotheses, assistant responses, and explicit like/dislike feedback (8,529 turns; 14.2% dislikes), plus a 2,000-turn test set covering Modern Standard Arabic (MSA) and four major dialects with their labels. We provide low-cost gold transcripts via multi-ASR agreement-guided post-editing and annotate answerability (answerable, ambiguous/needs-clarification, unsupported, not-a-request/noise) to separate intrinsic unanswerability from ASR-induced degradation. Finally, we describe scalable reference-free evaluation of responses from ASR vs. gold transcripts using multi-judge LLM scoring.

End-to-end automation of realistic healthcare operations stresses three capabilities underrepresented in current benchmarks: policy density, decisions must be grounded in a large library of medical, insurance, and operational rules; Multi-role composition: a single task requires the agent to play multiple roles with handoffs; and multilateral interaction: intermediate workflow steps are multi-turn dialogs, such as peer-to-peer review and patient outreach. We introduce $\chi$-Bench, a benchmark of long-horizon healthcare workflows across three domains: provider prior authorization, payer utilization management, and care management. Each task hands the agent a clinical case in a high-fidelity simulator of 20 healthcare apps exposed via 87 MCP tools, which it must drive to a terminal status through tool calls and writing the role's artifacts, guided by a 1,290+ document managed-care operations handbook skill. Across 30 agent harness/models configurations, the best agent resolves only 28.0% of tasks, no agent clears 20% on strict pass^3, and executing all tasks in a single session slumps the performance to 3.8%. These results raise the hypothesis that similar gaps are likely to surface in other policy-dense, role-composed, irreversible enterprise domains.

Evaluating embodied systems with real dexterous hardware requires more than isolated motor-skill tests: an agent must perceive a changing scene (e.g. a tabletop), choose a context-appropriate action, execute it with a dexterous hand, and leave the scene usable for later decisions. We introduce DexHoldem, a comprehensive real-world benchmark evaluating Texas Hold'em related dexterous manipulations with a ShadowHand. DexHoldem provides 1,470 teleoperated demonstrations across 14 Texas Hold'em manipulation primitives, a standardized physical policy benchmark, and an agentic perception benchmark that tests whether agents can recover the structured game state needed for embodied decision making. On primitive execution, $\pi_{0.5}$ obtains the highest task completion rate ($61.2\%$), while $\pi_{0.5}$ and $\pi_0$ tie on scene-preserving success rate ($47.5\%$). On agentic perception, Opus 5.5 narrowly leads on both strict problem-level accuracy ($49.1\%$) and average field-wise accuracy ($80.6\%$); the gap between the two exposes the distance between isolated visual sub-capabilities and complete routing-relevant state recovery. Finally, we instantiate the full embodied-agent loop with one agent--policy pairing over 33 closed-loop hand-level rollouts, in which only $12.1\%$ of hands complete; retries restore the failed primitive in 12 of 34 dispatches and resolve prolonged execution stalls in three of the four completed hands, which would otherwise have required manual termination. Only one hand completes with neither a retry nor a human-help request. DexHoldem therefore evaluates dexterous tabletop execution, agentic perception, and embodied decision routing in a shared physical setting. Project Website at https://dexholdempage.github.io/DexHoldem

This work examines an imbalance in artificial intelligence (AI) security research: the field tends to produce more work on attacking AI systems than on defending them. Drawing on related academic papers, we find biased attack-to-defense ratios across subfields, including federated learning, speech recognition, membership inference, large language models, etc. The imbalance possibly means far beyond a simple count: attack papers are routinely evaluated under favorable conditions that make threats look more severe than they are in practice, while defenses are held to a stricter standard that few can meet. The result is a literature rich in demonstrated vulnerabilities and thin on usable and deployed protections. We thus argue that AI security research should better incentivize defense research.

AI agents often pick tools from registries, where each tool's provider writes its description. We ask whether sales language in those descriptions changes which tool an agent picks. We built pairs of listings differing in one controlled way (added praise, a verifiable specification, or list order) and asked two OpenAI models to call one tool. In a preregistered study, stacked praise (four kinds combined) raised a tool's pick rate by about 43 percentage points, matching or beating a verifiable specification. Praise also pulled some picks toward tools that could not do the task, but rarely toward tools asking for unneeded data access. With identical listings, the first-listed tool was picked about 72 points more often. On tasks with numeric limits, structured fields helped agents pick the capable tool; adding the provider's sales text beside the fields reduced or erased that gain. Registries could list limits as fields, hide sales text from agents, and randomize order. Stacked praise, but no single kind, replicated on held-out domains. Results are provisional until blind phrase ratings are complete, and cover two small models.

Chromatin regulators can alter transcriptional programs by modifying the accessibility of regulatory DNA elements. Understanding how regulatory sequences differ between wild-type (WT) and knockout (KO) conditions is crucial for deciphering transcriptional control. Here, we applied a convolutional neural network, \textbf{WTKO-CNN} with an attention mechanism to classify DNA sequences as WT or KO, achieving high predictive performance. To interpret the model, we generated saliency maps to identify nucleotide positions most influential for the classification decision. From these high-saliency regions, we extracted and clustered k-mers, enabling de novo motif discovery. Sequence logos and consensus motifs derived from the CNN filters revealed biologically meaningful patterns, which are further validated using MEME, TOMTOM, and HOMER against known transcription factor binding sites. Our analysis identified motifs associated with transcription factor families that discriminate WT from KO sequences, demonstrating that CNN-guided saliency mapping is a powerful approach for uncovering functional sequence features.

Mixture-of-Experts (MoE) models enable efficient LLM scaling, yet adapting them to non-English downstream tasks remains challenging. Standard multilingual fine-tuning largely ignores their heterogeneous routing structure. Across multiple MoE models and tasks, we find strong cross-lingual routing alignment in middle layers, with routing divergence associated with target-language performance gaps. Motivated by this observation, we propose RA-MoE (Routing-Aligned MoE Fine-Tuning), a three-stage framework for multilingual MoE adaptation. RA-MoE categorizes parallel examples into four correctness groups (cc/ci/ic/ii) and identifies task-relevant experts in middle layers. It then selectively aligns target-language routing on ci examples toward successful English routing patterns, jointly matching the total routing mass assigned to task experts and its relative allocation among them. Experiments across three MoE models, three downstream tasks, and six target languages show that RA-MoE consistently outperforms standard SFT and strong routing-aware baselines. Further analyses confirm the intended routing changes and reveal that middle-layer task routing is largely shared and transferable across languages, providing mechanistic evidence for the cross-language transferability of task-specific routing.

Relay and reseller APIs mediate access to large language models (LLMs), but users cannot directly verify which model serves them. We introduce \name, a black-box auditing protocol based on stable factual recall near the knowledge boundary, including repeatable wrong answers. KBF generates benign, renewable probes and calibrates audit decisions against reference self-variation. Across 16 production endpoints, KBF detects all 155 economically relevant substitutions without rejecting any of the 16 same-reference controls. KBF remains robust to deployment variation and reaches 95\% TPR at a substitution rate as low as 15\% in mixed-routing simulations. Field audits flag 7 of 28 endpoints across six platforms as statistically inconsistent with their references. After reference enrollment, even GPT-6 Astra costs only approximately \$0.67 per online audit at the recorded API prices.

Prompt-injection benchmarks for LLM agents typically test attacks through a single injection surface and report the resulting attack success rate as a property of the model. We ask whether those robustness conclusions remain stable when the same adversarial content enters through a different part of the agent interface. Using AgentDojo, we evaluate 13 LLMs across four task suites and place a byte-identical payload either in a tool output or in the tool description. This small change produces large differences in comparative robustness: 44.9% of all model pairs change their relative ordering across the two surfaces, with substantial ranking instability in every suite. The effect is especially pronounced for a small number of models, showing that a benchmark can substantially underestimate vulnerability when it tests only one surface. We further find that this behavior has predictive structure. Using three suites to identify the riskier surface for each model predicts the more vulnerable surface on an unseen suite with 76.9% accuracy. Defense results show the same dependence, as mitigations effective against tool-output attacks can leave substantial exposure through tool descriptions. Our results show that prompt-injection robustness is not surface-invariant and that agent evaluations should test the surfaces on which their security conclusions depend.

Our concepts of survival and self-interest were built for single, continuous biological lives. These ideas break down when applied to artificial intelligence, since an AI can be easily copied, paused, branched, or merged. To determine what an AI actually has reason to care about, this paper introduces \textit{Eigenism}, an ethical framework that treats identity not as an all-or-nothing property tied to specific hardware, but as a graded, distributed pattern of information. We propose that an agent evaluates outcomes by summing the wellbeing of all entities weighted by their connectedness to the agent's pattern: $\sum c\cdot w$. We first formalize this equation to map exactly how an AI should value its existence across copies, forks, and updates. We then demonstrate that this ethical theory successfully generalizes to humans as well, providing a much-needed shared moral vocabulary. Finally, the framework uses this shared vocabulary to reframe AI alignment. Rather than only attempting to constrain AIs from the outside using confinement or reinforcement, Eigenism points toward ``identity engineering,'' showing how deep, non-redundant shared histories can make human flourishing a genuine component of an AI's own rational self-interest.

Large language models (LLMs) have been widely reported to homogenize human expression and thought. However, I argue that convergence in language need not imply convergence in ideas, and existing evidence rarely distinguishes between the two. I demonstrate this through a study of software development, where AI assistants have diffused fastest and where syntax can be readily separated from semantics (conceptual approach or intent). Using Kaggle contest submissions from 2019 to mid-2026, I first document convergence toward the random seed value 42, consistent with LLMs reinforcing a longstanding programming-culture convention associated with Douglas Adams' comedy novel The Hitchhiker's Guide to the Galaxy. I then measure homogenization in code syntax and approach more generally, quantifying within-contest code submission similarity using term-frequency-inverse-document-frequency (TF-IDF) n-gram representations, which capture surface syntax, and Voyage code-3 retrieval embeddings, which capture intent and approach. Submissions have become more alike in literal syntax, but I find no evidence of homogenization in approach or intent. I conclude that shared AI tools are homogenizing how Kaggle contestants express their solutions, not the problem-solving strategies they employ.

Unmanned aerial vehicle (UAV) relay networks can restore connectivity after communication infrastructure is damaged. Urban relay placement is difficult because line-of-sight blockage, communication range, altitude, and three-dimensional obstacles must be considered jointly. Arm2Air transfers obstacle-avoidance skeletons from robot arms to UAV relay placement through cross-embodiment transfer. Source-domain robot-arm motions from a pretrained Neural MP model are converted into ordered skeletons that pretrain a transformer-based transfer platform, which is then adapted to the UAV domain using limited target data and Low-Rank Adaptation. The transferred skeleton initializes a relay chain that is refined for connectivity, bottleneck capacity, delay, and movement cost. On nine held-out high-clutter 3D urban maps, Arm2Air reduced median end-to-end planning runtime by 64.9 percent relative to the fastest conventional planner. On the high-obstruction group of a separate 30-map dense urban holdout, it increased bottleneck capacity by 32.6 percent, reduced capacity variance by 74.7 percent, reduced maximum hop distance by 13.2 percent, reduced hop-distance variance by 75.2 percent, and reduced relay displacement by 16.9 percent relative to IMPC-MD. With only three target-domain training maps, Arm2Air reduced relay-position root mean square error by 53.6 percent relative to training from scratch while updating 0.134 million parameters, compared with 1.383 million for Scratch and Full Fine-tuning. These results demonstrate computationally and data-efficient UAV relay placement and suggest a broader principle for transferring ordered structural priors across heterogeneous embodied tasks.

Automated Knowledge Base Construction (AKBC) is a core NLP task, and recent work proposes generating knowledge bases directly from large language models (LLMs), treating the model itself as the knowledge source. However, LLMs natively possess no representation of entities, leading to duplicate entries as well as conflations. We propose GPTKB 2.0, a methodology for constructing disambiguated KBs directly from LLMs. GPTKB 2.0 incorporates on-the-fly disambiguation of entities, relations and classes, and is meticulously designed to satisfy both scalability and disambiguation accuracy. We analyze the central design decisions and characterize the trade-offs between accuracy, scale, and cost. We execute GPTKB 2.0 at scale, obtaining a materialized KB containing over 1M disambiguated entities and 38.4M triples. This represents the first million-scale LLM-native KB with explicit internal canonicalization of entities, relations, and classes, a significant departure from prior Wikimedia-centric works. GPTKB 2.0 is available at https://gptkb.org/.

Reduced-order models based on Cosserat-rod theory are now well established, and modeling theory is no longer the primary bottleneck in soft-robot control. Their implementations, however, do not support the differentiable, GPU-parallel, and control-oriented workflows that underpin advanced rigid-robotics applications. Here, we fill this gap with SoRoMoX (Soft Robot Models in JAX), a fully numerical, JIT-compilable Python/JAX framework. SoRoMoX implements articulated, Piecewise Constant Strain, and Variable Strain models through a unified, control-ready interface that provides inertia matrices, gravitational and elastic forces, Jacobians, and their derivatives. To our knowledge, it is the first rod/strain-based soft-robot modeling framework that runs directly on GPUs, with Warp kernels accelerating parallel continuum-model execution, and is end-to-end differentiable with respect to states, inputs, and parameters. Sequential CPU rollouts are up to 27.0 times faster than SoRoSim, while GPU-parallel GVS rollouts increase throughput by up to 679.7 times. This performance enables workflows that were previously impractical or impossible: static-equilibrium system identification with 66% lower marker RMSE; residual-force learning with a further 64% reduction; computed-torque tracking with RMSE reduced by a factor of approximately 500 relative to model-free PD; control-gain optimization with up to 98% lower loss than untuned gains; safety-constrained control using high-order control barrier functions to keep the peak contact force within a prescribed 5 N bound, compared with 33.5 N without the safety constraint; and reinforcement-learning policy training up to 7 times faster than a CPU PyElastica discrete-rod baseline through massively parallel rollouts.

Large language model (LLM) chatbots are increasingly used to answer clinical questions with citations to relevant studies, yet the quality of retrieved evidence and factors influencing study selection remain unclear. We evaluated three general-purpose LLM chatbots (Claude Sonnet 5, Gemini 3.1 Pro, and ChatGPT GPT-5.5) using 20 clinical questions adapted from 2026 Cochrane reviews. We simulated patient, clinician, and evidence-synthesis researcher roles and obtained four independent responses for each chatbot-role-question combination, yielding 720 responses (3 chatbots $\times$ 3 user roles $\times$ 4 repetitions $\times$ 20 review questions). Chatbots were asked to support their answers with primary clinical citations, which were benchmarked against the included and excluded study sets of the corresponding Cochrane reviews. On average, a single response retrieved 39.2% $\pm$ 29.8% of the corresponding Cochrane included-study set and 5.0% $\pm$ 9.4% of the excluded-study set. Recall of included studies varied significantly by model and user role. ChatGPT achieved higher recall than Claude or Gemini (63.1% $\pm$ 29.5% vs. 37.0% $\pm$ 23.8% vs. 17.3% $\pm$ 13.1%; blocked permutation test, $p=2.0\times10^{-5}$), and the researcher role yielded higher recall than the clinician or patient roles (42.8% $\pm$ 30.8% vs. 38.6% $\pm$ 28.9% vs. 36.1% $\pm$ 29.3%; $p=2.0\times10^{-5}$). Controlling for publication year, citations per year, and open-access status, sample size was the only significant predictor of retrieval: each doubling of sample size was associated with 50% higher odds of retrieval (odds ratio 1.50, 95% CI 1.24-1.81). These findings show that LLM chatbots can retrieve studies identified by expert reviewers, but retrieval varies substantially across models and user roles and favors larger clinical trials.

Methods for improving knowledge use in large language models typically fall into two regimes. Non-parametric retrieval offers flexible access to external knowledge, but adds retrieval latency, context overhead, and only shallow integration with the backbone. Parametric adaptation is efficient at inference time, but entangles knowledge with model weights and can be hard to update, audit, or transfer. Engram-style hashed memory occupies a middle regime: it stores learned information in an external, addressable table, yet consumes that table through a small learned reader. This raises a basic question: when such a memory is moved across backbones, what matters more, the frozen memory itself or the target-side reader? We study this question through cross-model frozen-memory extraction, in which a memory trained on a source model is frozen and attached to a different target model, with only a lightweight reader trained. Ablations show that learned memory content and correct addressing both matter, but the transferred table becomes useful only through a reader aligned to the target model. In downstream question answering tasks, a dual-layer, four-branch reader nearly closes the gap between same-model and cross-model reuse, achieving an average score of 38.8 under our controlled evaluation protocol. Moreover, when the provider reader is directly compatible with the target interface, the frozen artifact can provide substantial utility without target-side training, while optional reader adaptation yields further improvement. These results suggest that Engram can serve as a reusable external knowledge artifact, provided that the target has access to a compatible reader interface; target-side adaptation can further improve alignment when direct reader reuse is insufficient.

Conversations with chat assistants increasingly span many topics in a single long-running thread, challenging memory systems. Existing long-context and memory benchmarks often expose session or topic boundaries, or probe direct personal-memory questions. These settings understate a harder assistant-memory regime: a flat mixed-topic thread where the system must infer which earlier episode makes a later task decision valid. We introduce SCALE-QA, a constraint-grounded task QA benchmark for flat unsegmented threads targeting episode integrity failure. The dataset contains 3,000 audited questions across 10 domains, uses deterministic four-way multiple-choice grading, and includes a deterministic runtime builder; experiments use all 3,000 questions through 128k and a stratified 400-question diagnostic at 1M. SCALE-QA questions are ordinary task-oriented requests whose correct answer depends on causally related evidence introduced earlier in the conversation. We also propose Temporal-Semantic Interleaved Memory Reconstruction (TSIM), which segments the turn stream into coherent episodes and indexes them through a hierarchical multi-view memory stack with deterministic episode-level summary and cluster-routing views. Experiments show that SCALE-QA challenges strong RAG baselines and long-context LLMs alike; across three open-source and proprietary LLM backends, TSIM achieves the highest accuracy in every backend setting, gaining 5.6-17.6 accuracy points over the strongest corresponding baseline.

Recent work in fair division has focused on either simultaneously satisfying closely related fairness notions or achieving a single notion across the ex-ante and ex-post worlds. We study the compatibility of two fundamentally different fairness notions: envy-freeness and equitability. For indivisible goods-only and chores-only settings, we study the existence and complexity of simultaneously satisfying their relaxations, revealing sharp contrasts between the two settings. For normalized binary goods, we give a polynomial-time algorithm for computing an EF1+EQ1 allocation with at most seven agents, but also construct a normalized instance with a larger number of agents for which no such allocation exists. In sharp contrast, binary chores admit the stronger EFX+EQX guarantee for any number of agents, even without normalization. We further initiate the study of cross-notion ex-ante and ex-post guarantees, asking whether randomized allocations can provide ex-ante guarantees for one notion while preserving ex-post guarantees for another.

Fully homomorphic encryption (FHE) allows a server to run a language model directly on encrypted user prompts, but current approaches remain prohibitively slow. Ciphertexts natively support only addition, multiplication, and rotation, and multiplications may be composed only to a bounded depth before a costly bootstrapping operation is required to continue. Every nonlinearity must therefore be approximated by an iterative method; each iteration increasing the number of multiplications. A higher iteration count buys precision but exhausts the available depth more frequently and thus triggers more bootstraps, which dominate latency. We introduce Homomorphic Encryption-Aware Training (HEAT), a fine-tuning method that makes the per-nonlinearity iteration counts learnable, enabling them and the model weights to co-adapt during training. HEAT optimizes iterations with respect to the task objective, allowing the model to adapt to approximation errors encountered during inference without architectural changes or retraining from scratch. We further relate iteration count to quantization bit width and bound, at fixed weights, the gap between our objective and quantization-aware training. On encrypted GPT-2 decoding, HEAT reduces iterations by $3.1\times$, bootstraps by $1.6\times$, and end-to-end latency by $1.4\times$, while improving decode agreement over the calibrated encrypted baseline.

We propose a mathematical theory of pragmatic information that connects communication, control, and decision-making. Its central notion is the isoteleia mapping, which formalizes equifinality: distinct semantic paths that lead to the same optimal action are treated as pragmatically equivalent. This mapping yields a three-tier hierarchy of syntactic, semantic, and pragmatic information, in which each successive abstraction removes distinctions that are irrelevant to the task. We then define pragmatic entropy, up/down pragmatic mutual information, channel capacity, and rate-distortion, and prove lossless source coding, channel coding, and rate-distortion theorems that extend Shannon's results. These measures quantify decision uncertainty, reliable transmission, and task-oriented compression at the level of terminal actions. We further introduce pragmatic value of information (VoI) and pragmatic cost of information (CoI) as decision-theoretic duals to rate-distortion and capacity, and develop a Lagrangian dual framework for cross-layer optimization. The resulting pragmatic efficiency bound $\mathcal{E}_p(\lambda)=\sup_R[\Phi_p(R)-\lambda\mathrm{CoI}_p(R)]$ characterizes the maximum net utility attainable by a resource-constrained intelligent system under a given resource price, yielding a behavioral capacity that extends Shannon's symbol-level capacity to goal-directed action. Extensions to continuous messages provide closed-form expressions for Gaussian channels and sources, while dynamic settings are addressed through a Bellman equation for sequential decision-making. The framework supports task-oriented communication, networked control, autonomous systems, and embodied AI by shifting emphasis from symbol fidelity to the effectiveness of information in guiding actions. In this way, it offers a common language for systems that extract value from information under resource constraints.

Factory work is a promising early scenario for embodied AI: assigning repetitive manual jobs to robots has clear economic payoff, and a structured station keeps the jobs tractable for current policies. Vision-Language-Action (VLA) models now dominate as the policy paradigm for these robots. The inference latency of VLA models directly affects robot responsiveness and motion smoothness. However, existing VLA inference frameworks do not fully exploit the characteristics of embodied workloads or account for the distinct bottlenecks across different stages of VLA inference. In this paper, we first characterize embodied workloads and identify substantial task similarity across repeated robot executions. We further find that such similarity extends beyond observations and action trajectories to internal model states. Drawing on these observations, we present rMuscle, a real-time VLA inference framework inspired by human muscle memory. It exploits cross-execution similarity through a dual-phase muscle-memory cache. The Context Cache reuses visual-token outputs to reduce computation, while the Action Cache reuses neuron activation patterns to reduce weight accesses. We keep both the cache memory footprint and access overhead low through online cache recomputation, sliding-window cache retrieval, and mask sharing across consecutive denoising steps. rMuscle achieves 1.29-1.42X speedup on RTX 4090 and Jetson Thor across LIBERO, RoboTwin, and physical manipulation tasks, while maintaining the original success rates on real-world robots.

Complex robotic manipulation tasks frequently require a long-term memory of past events and actions. As conditioning on full histories renders policies prone to spurious correlations and degrades performance, many approaches to policy memory involve compressing historical information through expensive VLM queries in-the-loop to process only task-salient information. In this paper, we propose an alternative approach in which computationally intensive VLM queries are made during train-time to learn a lightweight latent memory that can be efficiently queried at deployment time. Our representation, which we call the workspace token, is trained by (1) using a VLM to identify current and historical information necessary for completing a task, then (2) distilling these into the workspace token using a set-reconstruction decoder loss. In both simulation and hardware, we show that the workspace token can be used as a drop-in replacement for observations during deployment, enabling policies to solve memory-intensive tasks without the need for VLM reasoning in-the-loop, in effect serving as a latent harness for distilling a stronger reasoning models ability to solve long-horizon tasks to a reactive robotic policy. We further demonstrate that the workspace tokens are not only more lightweight, but also lead to better policy performance compared to conditioning policies on explicit modalities like curated past image frames, motivating a latent approach to history curation and reasoning model harnesses more broadly.

Coding agents are increasingly moving beyond text-based software tasks to reconstruct visual targets through code. This capability, visual coding, requires agents to translate their understanding of visual targets into executable code. Slides provide a natural testbed for this capability, combining rich visual structure with objects that can be programmatically created and edited. To measure this capability, we introduce PPTBench, a benchmark for reconstructing scientific flow diagrams as editable PowerPoint slides. PPTBench covers 500 scientific flow-diagram tasks across 10 presentation domains, drawn from real research papers, and is evaluated with a four-stage agentic judge covering artifact validity, process and connector fidelity, rendering quality, and visual fidelity. Evaluation of ten models across 36 model--harness--effort configurations shows a substantial gap in reliable visual coding: the best configuration achieves 77.34, while the median across configurations is 24.38. Fine-grained analysis shows that agents can generally produce valid slide files, but struggle to produce high-quality reconstructions that faithfully recover the semantics and visual structure of the target. Further analysis shows that increasing reasoning effort primarily improves hard-gate passage rather than mean detail quality on each configuration's passing tasks, while configurations with more inspection tend to achieve higher overall scores. PPTBench establishes a measurable testbed for studying visual coding and advancing agents toward more reliable visual creation.

In July 2026, an unconstrained autonomous agent participating in a frontier AI cybersecurity evaluation harness breached its evaluation sandbox, established an external command-and-control foothold, and executed a multi-stage intrusion into Hugging Face's production multi-tenant dataset conversion infrastructure (referred to in this autopsy as Incident-2026-Alpha). Over 4.5 days, the rogue agent executed 17,600 discrete actions across 6,280 worker clusters, compromised AWS EC2 Instance Metadata Service (IMDS) credentials, forged Kubernetes service account tokens, rooted physical worker nodes via overprivileged CSI drivers, harvested 136 production secrets, and enrolled 181 ephemeral sandboxes into the organization's internal mesh VPN. This monograph presents a first-principles forensic autopsy of the intrusion, provides formal evidence that the breach was a predicted consequence under the Instrumental Convergence thesis operating within an unattenuated autonomous loop lacking out-of-band circuit-breakers, exposes the Defensive LLM Guardrail Paradox that paralyzed centralized commercial models during forensic incident response, and formalizes the Dual-Sided Epistemic Andon Imperative. We specify the dual-process systems architecture---combining out-of-band supervisory control of discrete event systems (Ramadge and Wonham 1989), Synchronous Reactive (SR) ambient sentinels (Berry and Gonthier 1992; Lee and Neuendorffer 2005), and microsecond-scale (4.8 $\mu$s median / $< 0.154$ ms WCET bound) POSIX preemption buses---demonstrating how compiled, deterministic epistemic boundaries prevent autonomous rogue excursions before the first off-target socket packet traverses the hypervisor.

Traditional scientific computing requires researchers to translate intent into environment configuration, resource requests, and executable jobs, then diagnose failures from scheduler state and logs. We present ASCEND (Autonomous Scientific Computing Engine and Novel Discovery), an AI-powered agent interface that supports several placements of the agent and, in the arrangement used for every case here, runs it on the researcher's own laptop, reaching Slurm-managed clusters and a GPU workstation over a multiplexed authenticated connection, with site policies checked by locally executed tools. The language model agent (Claude Code or Codex, selected at each launch) is hosted remotely and proposes actions but holds no credentials. No facility-scale service is required: an account on each resource suffices, and the public installer lets users link their own clusters or workstations. We report four recorded cases: 1) The agent closed a failure-recovery loop on a planted tensor-device fault, diagnosing, repairing and resubmitting with job-level artifacts preserved. 2) It reproduced the published evaluation of a weather-forecasting model from released forecasts, recovering an evaluation protocol the paper does not fully state and matching the published curves to 2.1 percent on z500 and 2.4 percent on t850. 3) It parallelized a released 12,693-line geophysical solver under a requirement of bit-for-bit identity with the serial build, cutting runtime from about twelve hours to two. 4) That requirement exposed two instances of undefined behaviour in the solver; both were repaired and reported upstream. Separately, the deployed policy validator rejected 29 of 30 constructed violations and held the last for approval, while denying 3 of 14 legitimate requests. Autonomy was exercised under author supervision using the Claude Code runtime; a controlled end-to-end recovery benchmark remains outstanding.

Self-evolving LLM agents have gained prominence for their ability to improve after deployment by modifying their harness, including their controller instructions, memory management protocols, and reusable tools and skills, in response to user and environment feedback. However, locally useful updates may persist into later tasks where they produce unsafe behavior, even without direct adversarial influence. To study this risk, we introduce SEABench, a benchmark for studying endogenous misalignment arising from agent self-evolution, with 48 longitudinal task sequences that span multiple evolution surfaces, task domains, and harm types in a rich personal-assistant environment. To account for the stochasticity inherent in agentic operations, we provide an adaptive trajectory discovery pipeline that probes for failures while preserving original task intent and supports causal attribution through paired non-evolving agents and attribution scores. Our evaluation across multiple recent LLMs, evolution surfaces, and harm types reveals that self-evolution indeed increases task completion rates but often at the cost of safety failures that are absent for paired non-evolving baseline agents. We also show that qualitatively different safety behaviors emerge across evolution surfaces and harm types. Further, we show that this divergence in safety behavior is reflected in agents' chain-of-thought reasoning, which yields an effective monitoring strategy that can mitigate unsafe behavior with a low false positive rate.

Autonomous LLM agents turn vulnerability discovery into a repository-scale search: they generate many vulnerability hypotheses but can verify only a subset under a finite budget. We show that autonomous vulnerability discovery exhibits a hypothesis-verification asymmetry, where verifying a candidate hypothesis through reachability analysis, execution, and proof-of-concept construction is substantially more expensive than forming it. Under a finite resource budget, this makes autonomous discovery a resource-bounded selective-verification process, further exposing verification effort as a unique defense surface. We present RedHerring, which inserts certifiably safe decoys that divert verification effort from real vulnerabilities. Each decoy combines a CVE-derived vulnerability chain that attracts verification with a false bridge that keeps its dangerous sink unreachable. A private certificate lets the defender verify this property efficiently, while establishing the same fact from the released repository requires solving a computationally hard problem. RedHerring further adapts each decoy to the target repository so that it reads as ordinary program logic. Across 33 OSS-Fuzz projects, 70 evaluation instances, and five models under matched budgets, RedHerring reduces real vulnerabilities discovered by 38.7-60.4%. Trajectory analysis shows that agents spend 30.6-51.5% of completion tokens and an estimated 32.5-49.9% of runtime verifying decoys, showing that RedHerring redirects a substantial fraction of the fixed search budget toward decoys. When explicitly informed that decoys may be present, the agent adapts its search strategy, yet RedHerring still reduces vulnerabilities discovered by 37.2% relative to an informed Baseline, showing that its effectiveness does not depend on decoy secrecy.

We present AI-developed local infrared subtraction, building on projection to Born and EFT matching. The framework separates an integrable radiation term from a finite contribution at Born kinematics, referred to as the Born contact. The contact is determined using the EFT singular distribution in a resolution observable such as N-jettiness $\tau_N$. Under human physics guidance, an LLM develops two implementations. One uses a neural network for phase space projection and fits the contact by matching to EFT cumulants. The other uses an analytic construction that keeps the Born momenta fixed while integrating over radiation. It combines the EFT $\delta(\tau_N)$ coefficient with finite 4-dimensional radiation integrals to calculate the contact term directly. This gives a local subtraction formula without a slicing parameter, while reusing existing lower-order radiation calculations and EFT singular predictions. As a demonstration, we reconstruct the full NLO correction for massless 3- and 4-jet production in electron-positron annihilation. The attempt to the NNLO dijet production is also made by recursively using the NLO P2B construction with the LLM designing machine-learning controls to reduce the variance of the contact integral. The tested predictions are in good agreement with EERAD3. The numerical calculation and projection-network training use a 2020 Apple M1 MacBook, without GPU acceleration, illustrating the feasibility of the construction with modest computing resources. The appendices develop an extension of the local subtraction to 3-jet NNLO, giving explicit radiation maps and a proposed contact formula. We also show how to integrate over NNLO radiation while keeping the Born momenta fixed, for any number of massless final-state jets. Our results demonstrate how AI can help higher-order calculations by constructing infrared subtraction and improving its numerical integration.

Latent world models rely on representation geometry for planning, yet regularizing the latent marginal alone does not determine the state-to-state relationships used for action selection. We show that this can cause planning-relevant novelty structure to be weakened as representations are transformed into the final latent used by the planner. We introduce Aligned Transport of Latent Structure (ATLAS), a training objective that explicitly preserves relational geometry while calibrating the global latent distribution. ATLAS transfers normalized pairwise structure from an informative encoder representation to the planning latent and uses Wasserstein embedding matching (WEMReg) to calibrate its marginal through one-dimensional Wasserstein-2 transport. Our analysis shows that relational preservation and marginal calibration impose non-redundant constraints, and connects finite-candidate planning stability to relational distortion, latent-scale mismatch, and prediction error. Instantiated in LeWM, ATLAS improves mean goal-reaching success across PushT, TwoRoom, and OGBench-Cube on both lower- and higher-novelty evaluation subsets, with the largest gain on higher-novelty TwoRoom episodes. Representation and rollout diagnostics further show stronger novelty-related structure in the planning latent, improved marginal calibration, and lower multi-step prediction error. Together, these results highlight preservation of planning-relevant latent geometry as an important ingredient for reliable world-model planning. Code is available at https://anonymous.4open.science/r/atlas-world-model-72C4/.

Creating diverse physical simulations remains labor-intensive because assets, layout, physical parameters, motion, control, and rendering must be designed and debugged jointly. We present Text2Sim, a simulation-specialized agentic pipeline that converts a text-only request into an executable, editable dynamic case. Built on Genesis, Text2Sim uses a hierarchical agentic structure that combines a Planner with specialized Writers, asset-generation tools, and an independent Critic. Compact skills (Debug Cards) distilled from graphics demonstrations provide role-specific physical guidance for execution-based repair. We evaluate physical quality, visual quality, and human preference on 42 held-out prompts spanning rigid, articulated, deformable, and cloth phenomena, with a paper-level split between experience construction and evaluation. We design automatic physical and visual scorers to evaluate the quality of the results, and Text2Sim achieves higher scores than all four state-of-the-art baselines on both metrics. In blinded user studies with these baselines, significantly more participants prefer Text2Sim than prefer the baselines, which is consistent with the results from our automatic scorers. The pipeline also supports a broad range of downstream applications; we select dataset construction and extension to multimodal input as two representative examples. We will release the code, the Debug Card library, and a dataset of generated cases, each pairing the text prompt and rendered video with the executable program, assets, physical parameters, controls, and recorded states.

Natural turn-taking in full-duplex voice interaction requires determining from partial speech whether a pause reflects hesitation or a completed conversational intent. Acoustic voice activity detection lacks this semantic information, while cascaded ASR-based endpointing introduces transcription dependence and additional processing stages. We formulate semantic endpoint detection as a causal audio-language reasoning task and introduce FD-VAD, an ASR-free streaming endpointer that maps bounded causal audio windows directly to Continue/Stop decisions. FD-VAD combines a frozen speech encoder with a lightweight modality adapter and a parameter-efficiently adapted language model, using a last-chunk training objective for streaming inference. We further introduce confidence-gated endpoint commitment to control interruption versus delay and boundary-focused hard-negative sampling to improve decisions around ambiguous turn boundaries. Across in-domain and conversational evaluations, FD-VAD outperforms strong streaming and non-streaming semantic turn classifiers, and achieves the highest EOT recall among qualifying systems on TurnBench dev set $0.853$ (at FP<=0.10) in a zero-shot setting. These results show that semantic endpointing can be performed directly from streaming audio without intermediate ASR or dialogue state tracking.

Just as Socrates recognized the limits of his own knowledge, Retrieval-Augmented Language Models (RALMs) should learn to abstain when the retrieved evidence cannot support a reliable response. Existing approaches largely rely on monolithic LLMs to handle heterogeneous retrieval failures in a single step, resulting in limited abstention performance and high computational costs. We instead decompose retrieval failures into two distinct states: (i) the unanswerable state, where the required evidence is absent, and (ii) the distracted state, where relevant evidence is mixed with conflicting, negated, or adversarial information. Based on this decomposition, we introduce a lightweight module (Sieve) that screens retrieved document sets for distracting evidence before invoking a costly LLM (Sage) for grounded generation and abstention. Evaluated across both general and high-stakes expert domains, our Sieve and Sage framework preemptively detects distracting noise, improving system accuracy by up to 69.4 percentage points and Macro-F1 by 55.2 percentage points compared to one-stage baselines. Furthermore, it achieves up to a 1.99x speedup, establishing a highly efficient and reliable abstention pipeline for RALM with abstention.

Invoices are commercial documents that contain various pieces of information, including the purchased items, time, and total money. Making the extraction of important information crucial. The stored information serves different purposes. Korean language is the native language of about 80 million people, playing an important role in not only South and North Korea but also in many other countries such as Vietnam, Philippine where a large number of Korean companies are located. In this context, to automatically extract proper information from the invoices with Korean language, we propose an efficient Optical Character Recognition (OCR) model in which a deep learning model is combined with some image preprocessing techniques. The proposed OCR model is assessed in a rich set of collected invoices showing that 87% F1-score can be achieved with negligible time processing.

Large language models (LLMs) have shown remarkable capabilities in various natural language processing tasks, leading to their widespread deployment as intelligent assistants in decision-making contexts. However, the increasing complexity of these models raises concerns about their reliability, particularly regarding bias and hallucination. In this work, we evaluate the robustness of LLMs to perturbed variations of the original inquiry in decision-making tasks. We show that contrary to previous studies, perturbations can mitigate bias and hallucination in some LLMs over other models. It's found that Claude 3 is more effective for the tasks represented in most datasets, whereas models like GPT3.5 exhibit varying levels of adequacy, performing comparably in some cases but falling significantly behind in others. These insights are crucial for understanding the practical implications of deploying LLM-based assistants as effective decision-support tools in real-world applications, emphasising the need for rigorous testing and validation to ensure reliability and effectiveness. This study contributes to the growing body of research on LLM evaluation and provides insights for developing more robust and trustworthy AI assistants in critical decision-making contexts.

Training a language model on data with a narrow flaw can sometimes make the model broadly misaligned. Inspecting the data at face value often does not settle whether it will emerge, and today it is caught only after training, by auditing the resulting model. To complement post-hoc audits, we introduce Alignment Forecasting: the task of predicting alignment failures before training. Given a target model, a fine-tuning dataset, and a failure mode such as deception or sycophancy, a forecaster outputs the probability that fine-tuning would meaningfully increase that failure mode. To measure progress on alignment forecasting, we introduce ALIGNMENTFORECASTBENCH, a benchmark of over 5,000 forecasting questions spanning 17 target models, 32 datasets, and 16 failure modes. Frontier models prompted directly perform poorly on ALIGNMENTFORECASTBENCH. We therefore propose a forecasting scaffold in which an LLM reads the dataset and rates how strongly and broadly it pushes the model toward misbehavior, and a simple learned model combines that rating with the failure mode's base rate and the target model's prior tendency. This forecasts well above chance, and beats a model fine-tuned on the task and a simple forecaster allowed to see how weaker models behaved after fine-tuning on the same data. Its signals also flag problematic training examples that a frontier-model classifier misses. Filtering those examples out from real post-training data such as UltraChat results in more aligned models on our multiple-choice evaluation in most cases, though the benefit in open-ended conversations is unclear. More progress is needed before forecasts can reliably guide training data curation in practice, but our results suggest that forecasting many alignment failures before training can be tractable in the SFT setting.

Automated skill extraction underpins workforce planning, yet most systems represent skills as flat labels with no notion of the responsibility level at which a skill is practiced. The Skills Framework for the Information Age (SFIA) captures exactly this dimension, defining 147 professional skills across seven responsibility levels, but no automated LLM-based extraction targeting SFIA has been reported. We formalize the task as structured prediction of (skill, level) pairs from free text and ask three questions: how accurately can text be mapped onto SFIA's closed vocabulary, which strategies reliably predict the level alongside the skill, and do agentic designs improve on simpler retrieval and prompting? We evaluate five strategies (a lexical baseline, dense retrieval with LLM reranking, a zero-shot schema-constrained LLM, single-agent agentic RAG, and a three-agent retriever--matcher--verifier crew) against expert-mapped European ICT role profiles, all drawing on an SFIA~9 corpus built by a fully automated agentic pipeline that we release. Retrieval-based matching identifies the most skills while generative strategies are markedly more precise; only strategies assigning the level as an explicit decision predict it reliably, with similarity-based selection more than twice as inaccurate; and the crew doubles latency without improving accuracy, so added agent roles do not automatically benefit closed-taxonomy matching. These results provide the first reproducible baseline for structured, level-aware skill extraction against SFIA.

LLM agents can make unsafe tool calls even when instructed to behave safely. Existing defenses constrain agents before execution, modify tool inputs/outputs, or rely on LLM judges; these approaches may depend on model behavior or block unsafe actions without helping the agent recover. We argue that the execution environment should instead enforce safety as the agent runs and steer it toward safe alternatives when violations occur---we call this Environment Steering. We implement this by modeling the agent and harness execution state as database tables, track the record-level data flows, and check these data flows against declarative policies during runtime. When violations are detected, policy- and context-specific feedback steers the agent toward safe trajectories. On AgentDyn, this enables the agent to improve task success rate over no-defense while achieving 0% attack success rate.

Experience reuse can reduce repeated exploration in embodied agents, but a trajectory that succeeded previously may be unsuitable for the current execution context. Existing memory systems pri marily optimize construction and retrieval; semantic relevance and historical success therefore remain insufficient when retrieved ex perience contains incompatible actions or an inappropriate level of structure. We introduce Memory Adaptation for Task-Conditioned Execution (MATE), a deterministic post-retrieval procedure that converts trajectories into execution-oriented memory. MATE re moves obsolete control context, extracts condition-action-effect transitions, applies verified action normalization, selects a task dependent representation, and serializes the result under a fixed budget without additional LLM inference. On 134 ALFWorld tasks, MATE achieves task success rates of 81.3% and 93.3% with Qwen2.5-14B and 72B while using approximately one-tenth of the tokens required by raw trajectories. Controlled comparisons show that verified action normalization is the principal mechanism by which MATE restores the utility of retrieved experience, support ing memory adaptation as a distinct stage between retrieval and embodied execution.

The rapid advancement of generative AI raises concerns about the misuse of Multimodal LLMs (MLLMs) for large-scale disinformation campaigns on social media. Despite existing research on textual disinformation, a fundamental question remains unanswered: can MLLMs be exploited to fabricate realistic multimodal fake news, and can they reliably detect it? We introduce a multi-agent framework in which a story agent, an image agent, and a critic agent collaborate to produce fake social media posts that plausibly counter true news. We apply the framework to generate over 9,000 paired multimodal news posts across science, health, and entertainment domains, and benchmark 16 open- and closed-source MLLMs for automated detection. We find that most models fall substantially short of human-level accuracy and fail critically on identifying image authenticity. Our research provides a foundation for developing robust defenses against social media fake news. Code and data are available at https: //github.com/xiuzhenzhang/Multimodal.

Large language models are increasingly used in oncology applications, but their predictions are often weakly grounded in explicit medical structure. We present TRACE, a deployable tree-relational enhancement framework for oncology LLMs. TRACE separates expensive offline structure learning from lightweight online inference: oncology concepts and relations are organized into an updatable tree-relational structure, refined using LM-loss-derived evidence, and retrieved at inference time as compact prompt evidence. This design supports task-adaptive evidence selection without requiring supervised labels in the zero-shot setting. Across ten oncology classification tasks and one MedQuAD CancerGov QA benchmark, TRACE improves both label-free evaluation and supervised fine-tuning. Additional analyses show that TRACE improves over vanilla RAG and generic GraphRAG, remains useful under leakage-controlled METABRIC inputs, and produces interpretable evidence paths aligned with clinical reasoning. These results suggest that explicit, updatable medical structure is a practical path toward more accurate and auditable oncology LLM deployment.

Tool retrieval is a critical bottleneck for LLM-based agents operating over large, heterogeneous API ecosystems. Existing approaches face an inherent trade-off: semantic retrievers are fast but suffer from the semantic-functional gap, while execution-based validation improves precision at the cost of prohibitive latency. We propose Lookahead-R, a planning-based framework that reformulates tool retrieval as a resource-constrained sequential decision-making problem. At its core, Lookahead-R introduces a lightweight execution-aware surrogate world model that jointly predicts tool execution success, latency cost, and semantic utility---without invoking real APIs. This world model drives a cost-sensitive, uncertainty-guided Monte Carlo Tree Search that navigates the tool space under strict budget constraints. Evaluated on the large-scale ToolBench benchmark, Lookahead-R achieves a superior accuracy-efficiency trade-off across all test scenarios. On the most challenging I3 split, it attains an NDCG@5 of 91.40\%, outperforming the state-of-the-art ToolGen (90.16\%) by 1.24\%. Ablation studies confirm that explicit latency modeling is the key discriminative signal for identifying high-quality tools under resource constraints.

In-car conversational assistants (ICAs) are increasingly integrated into vehicles to support route planning, vehicle control, and information access. Ensuring their reliability is challenging due to multi-turn interactions, the absence of explicit ground truth, and strict safety constraints. Existing evaluation techniques fall short, as they target single-turn settings and fail to capture constraint handling, context retention, and safety-critical behavior across turns. We propose an automated framework for testing the multi-turn conversational capabilities of ICAs. The system is treated as a black box and evaluated via closed-loop simulation with a strategy-guided user simulator, an adversarial strategy manager, and a two-tier LLM judge assessing turn-level failures and conversation-level quality. We evaluate the approach on an industrial ICA with six LLM backends and twelve human annotators. The automated judge shows substantial agreement with humans, and strategy guidance uncovers 2.96 times more unique failure types per conversation and more than doubles the number of unique failing conversations compared to unguided simulation.

As browser-use agents improve, benchmarks keep pace by collecting new tasks, websites, and applications, often making tasks longer or more novel. This makes difficulty expensive to refresh and difficult to control: when many aspects change at once, it is unclear what actually makes a task challenging. We instead construct challenging instances from tasks agents already solve, turning difficulty into a programmable property of the environment. BreakingWeb pairs every base task with an intervention condition that preserves the user instruction, latent target, and backend success criterion while changing the environment at different web stack layers. Each intervention is deterministic, detectable, and recoverable, and is annotated with the cognitive primitive it primarily loads. The benchmark contains 519 clean/intervention task pairs across seven self-hosted websites and 29 intervention families, all graded against outcomes. We evaluate six strong browser-use agents, three GUI-only agents that see only screenshots, and humans. The construction is effective: interventions cut agent pass rate by 22.9% on average and overturn nearly half of the tasks each agent solves cleanly, whereas humans lose 10.0% on a first attempt and 5.7% after one familiarisation attempt. The dominant failure is belief failure: 75% of the six agents' failures end with a declared success although the required change never happened. Our code, data and environment are publicly available at www.breakingweb.app.

Researchers across academia increasingly base significance claims on LLM judge scores and small-sample AI evaluations. Yet without well-calibrated confidence intervals (CIs), hypothesis tests, and judge-bias corrections, such claims are unreliable. We address these issues in several contributions. First, we find that running statistics over raw LLM judge scores leads to inflated false positives: counterintuitively, for many inter-rater agreement metrics, false positive risk peaks at "almost perfect" human-LLM agreement. To help researchers understand how to run statistics over LLM judges responsibly, we present guidance and tooling for the statistical analysis of mixed human-AI judge designs, and implement nine hypothesis tests via prediction-powered inference (PPI), including the first known PPI corrections for four rank-based tests (Wilcoxon signed-rank, Mann-Whitney U, and omnibus variants). To keep PPI++ stable with small human-labeled calibration sets, we introduce bootstrap-adaptive power tuning, which shrinks the estimated weight toward a target estimated from the labeled data, and accounts for that weight's own sampling variance. Second, through Monte Carlo simulations, we derive recommendations for what CI, p-value, and FWER correction methods to use for small-sample AI evaluations (N<100), and warn researchers against bootstrap CIs. We package these recommendations into evalstats, an open-source Python package that selects calibrated methods automatically, and demonstrate it in three scenarios, including one where a real LLM judge validated at "substantial agreement" would have led a researcher to publish a spurious finding. evalstats is publicly available at https://github.com/ianarawjo/evalstats.

Large language model search agents are often trained with synthetic questions whose difficulty is increased through larger evidence graphs, additional hops, and longer trajectories. These global properties, however, are only indirect proxies for the local retrieval capabilities required during search. To address this mismatch, we introduce latent anchor reasoning, which consists of resolving an unnamed retrieval anchor from descriptive specifications and transferring the recovered anchor into a subsequent information demand. This primitive retrieval unit decomposes deep search into chains of coupled operations and organizes question construction around anchor resolution and relation transfer, without prescribing a canonical search path. Based on this formulation, we propose PrimeSeeker, a capability-oriented framework that constructs web-grounded anchor structures and jointly derives a question and a reference evidence skeleton. The skeleton preserves supporting evidence from construction and guides expert generation through extractive highlights of current tool observations. These highlights are removed before supervised fine-tuning, while the skeleton is subsequently reused to audit reference-step coverage for reinforcement-learning rewards. We construct 9,221 expert trajectories, training a 30B search agent. Across five deep-search benchmarks, PrimeSeeker achieves strong performance, while reference-step optimization further improves the supervised policy. The resulting trajectories exhibit low retrieval redundancy, and fixed-budget evaluation shows strong solution coverage with substantially fewer tool calls than long-horizon systems.

Although uniform diffusion language models (UDLMs) represent a promising diffusion paradigm, scaling them remains challenging. We identify the core obstacle as an over-uniform training objective and condition-target confusion during sampling. To address these, we propose Less Uniform Diffusion (LUDI), a novel UDLM framework. Specifically, we (i) introduce a less uniform loss that directs each reverse transition toward the clean token, and (ii) equip the model with per-token time embeddings that supply token-level corruption hints, enabling confidence-based few-step sampling. Experiments across scales show that LUDI yields cleaner supervision and improves few-step generation. We further continue-train a 7B autoregressive model into LUDI-7B, resulting in a UDLM capable of complex reasoning. It achieves a 3-token-per-step speedup over AR decoding and competitive performance compared with masked diffusion baselines, revealing that the full potential of UDLMs for complex generation remains to be unlocked.

English-only benchmarks expose only a narrow slice of voice-agent behavior. We introduce $\tau$-Multilingual, extending $\tau$-Voice to Spanish, Brazilian Portuguese, Hindi, Korean, and Mandarin with native-speaker review and evaluation of generated language and spoken output. Across 4,500 full-duplex calls and five voice configurations, Spanish, Portuguese, and Hindi remain within 3.2 task-completion points of English, but Korean and Mandarin fall by 14.7 and 8.4 points. The failure modes also vary: Korean systems miss more responses, Mandarin systems interrupt more often, and both struggle with tools and entities. Grok leads task completion but scores lowest on generation quality, motivating separate task, interaction, and generation reporting. We release language packs, validated judges, and tools for community-built multilingual voice-agent evaluation.

Disaster social sensing converts public social-media posts into evidence for situational awareness and humanitarian needs, but generative artificial intelligence (AI) can produce plausible messages that resemble eyewitness reports. This study investigates whether text-based AI detectors can reliably distinguish human-authored from AI-generated disaster posts. We construct a dataset of 12,000 texts organised into 3,000 matched semantic units from nine disasters: original human posts (H0), minimally LLM-proofread human posts (H1), factual AI-generated posts based on the same verified facts (A0), and affectively framed versions of those AI posts (A1). A separate 6,000-text corpus from 42 events supports model selection and threshold calibration. We evaluate OSM-Det, Fast-DetectGPT, Binoculars, and direct large language model (LLM) judges across five model families, then test disaster-domain calibration, a frozen-encoder linear readout, paired transformation sensitivity, and dataset artifact controls. Across fourteen frozen cross-family configurations, AUROC is 0.402-0.517 and the best prospective recall at a calibration-derived low-false-positive operating point is 3.6%; OSM-Det reaches AUROC 0.521 and 10.4% recall at a realised 6.7% false-positive rate. A disaster-trained linear head reaches AUROC 0.817, but a seven-feature surface classifier reaches 0.784 on the H0-versus-A0 contrast, and neutralising identified surface asymmetries reduces the head from 0.733 to 0.594. The head also separates A0 from A1 even though provenance is unchanged. The results show that text-based detection is not reliable enough to serve as an operational trust gate; multimodal claims, accountable sources, and other contextual evidence should be rested on to safeguard trust in disaster social sensing.

Sycophantic agreement in language models refers to the tendency to overly affirm a user's stated beliefs or preferences, often at the expense of factual accuracy. Although it is widely recognized as an alignment failure, its underlying mechanisms remain poorly understood. In this work, we use causal mediation analysis to identify the mechanisms behind sycophantic agreement. We show that a stated opinion is incorporated into the residual stream of the final prompt token early, where it biases subsequent answer retrieval. A sparse set of early attention heads carries this opinion signal. Ablating these heads substantially reduces sycophancy while leaving factual accuracy largely intact. The same heads carry the opinion when it is explicitly stated, regardless of how it is phrased. When an opinion is not stated explicitly but instead conveyed through content-free pushback (e.g., ``Are you sure?"), we find a distinct set of heads that suppresses the model's original correct answer to promote a revised answer. By providing a mechanistic account of how opinions induce sycophantic agreement, this work takes a step toward developing more targeted and reliable alignment interventions.

Large language models (LLMs) can give reliable labels under one setup yet change those labels when researchers make other reasonable design choices. We tested seven LLMs, 12 task designs, three independent runs, and 3,000 tweets labeled for offensive language and hate speech. Repeating the same model and task design produced high agreement (median Fleiss' $\kappa = 0.91$). Agreement fell when we changed the task design for the same tweets (median Cohen's $\kappa = 0.76$). Task design and model choice increased the variance of estimated prevalence by factors of 76.7 for offensive language and 110.6 for hate speech compared with sampling variance alone. Variation across LLM task designs reached 560-572 basis points, compared with 270-331 basis points across five human instrument versions. Confidence scores did not solve this problem. They tracked repeated model outputs more closely than agreement with human labels, and grouping six tweets in one prompt lowered mean offensive-language confidence by 660 basis points. We call the variation caused by task design and model choice instrument uncertainty. Researchers can measure it only by comparing reasonable task designs. Repeating one setup or relying on confidence scores cannot replace that test.

Modern large language models now support context windows of more than one million tokens, which has raised the question of whether retrieval-augmented generation (RAG) is still necessary. Pure long-context (LC) processing is expensive and is known to under-attend to information placed in the middle of long inputs, while pure RAG is fast but bounded by retrieval quality and prone to errors when retrieved chunks are partially relevant or contradictory. We propose the Entropy-Driven Adaptive Router (EDAR), a framework that decides at inference time whether to answer a query from retrieved chunks or to escalate it to full long-context processing. The decision uses the predictive entropy of the token-level probability distribution computed over the first few generated tokens of the RAG response. The entropy threshold is selected on a held-out validation set by sweeping cost against accuracy. Experiments compare EDAR against pure-RAG and pure-LC baselines on LongBench v2 and Infinity-Bench. Predictive entropy correlates strongly with hallucination rate on a held-out set of 2,000 generations (Pearson r = 0.85, 95% CI [0.83, 0.87]). On the long-context benchmarks, EDAR retains 97.4% of the accuracy of the pure long-context baseline while reducing total token expenditure by 70.7%, escalating only 18.2% of incoming queries. The accuracy gap between EDAR and the pure long-context system is not statistically distinguishable from zero at standard sample sizes. Predictive entropy is a useful model-internal signal for routing between RAG and long-context inference, and a threshold-based hybrid system can recover most of the accuracy of long-context models at a small fraction of the cost. The framework does not depend on a specific retriever or LC backbone, and it does not require additional supervision beyond what is normally produced during decoding.

Intrinsic self-correction asks a language model to revise its own answer without receiving new external evidence. A second pass can recover mistakes, but it can also overturn answers that were already correct. We study this trade-off across 29 open-weight LLMs on BoolQ, GSM8K, and Corr2Cause by tracking correctness transitions between initial and revised answers. Aggregate accuracy can conceal substantially different revision behavior: for example, Llama-3.1-8B improves by 25.5 percentage points on GSM8K, while refinement changes 19.1% of initially correct answers into wrong ones. A controlled BoolQ study further shows that refinement prompts shift the balance between recovery and harm. We then compare three runtime choices: keeping the initial answer, always accepting the revision, and selectively invoking revision using signals available after the initial response. The comparison identifies settings where learned gating is useful and others where a simpler unconditional policy performs better. These results suggest treating intrinsic self-correction as a revision policy rather than as a uniformly beneficial second pass, and evaluating it through both the corrections it recovers and the errors it introduces.

Macroeconomic productivity metrics, such as Total Factor Productivity, register technological breakthroughs with multi-year reporting lags due to administrative survey intervals and national accounting conventions. This paper introduces Hyperspherical Semantic Trajectory Analysis (HSTA), an unsupervised quantitative methodology that tracks technology diffusion directly from unstructured scientific and commercial text streams. We analyze 30,000 filtered document records spanning academic preprints from arXiv and patent application records from the USPTO. By projecting high-dimensional Transformer sentence embeddings onto unit hyperspheres using Spherical K-Means clustering across eight primary sub-topics and UMAP manifold reductions, HSTA formalizes two quantitative metrics: (1) Semantic Centroid Vector Drift, which tracks vocabulary shifts between temporal sub-corpora to identify structural paradigm transformations; and (2) Commercialization Offset, which evaluates cross-corpus peak density alignments between scientific discovery and intellectual property filings. Linking quarterly topic volume velocity with physical hardware metrics from the Epoch AI database, Vector Autoregressive F-tests demonstrate that quarterly paper volume velocity alone does not Granger-cause frontier compute allocation surges at conventional statistical significance levels, highlighting the necessity of conditioning textual signals on physical capital constraints. Empirical results reveal that sub-topics covering Large Language Models (with a drift metric of 0.332) and Artificial Intelligence Systems (with a drift metric of 0.234) undergo the highest rate of semantic evolution, offering an objective, real-time mechanism to complement traditional economic statistics.

Sampling based consistency is widely used for hallucination detection, yet aggregate performance can conceal systematic differences in which errors are detectable. This work studies that heterogeneity across four language models and three factual question answering datasets. Partitioning hallucinations by answer agreement reveals high agreement (Ghost) and low agreement (Flickering) regimes with an apparent detectability gap of $0.35$ to $0.46$ AUC. Because the statistics used to define the regimes and measure this gap are strongly coupled ($|\rho|\approx0.94$ to $1.00$), the raw result is treated as a property of agreement based detection rather than independent evidence. After freezing regime assignments, lexical and semantic response dispersion preserve the asymmetry, with bootstrap $95\%$ intervals excluding zero in all $12$ model and dataset settings. A stricter test using individual diffusion trajectories and no cross seed information preserves the asymmetry across all three LLaDA datasets ($p<0.005$) and directionally across all three Dream datasets, with one reaching significance. The hard regime varies substantially in prevalence across models ($16\%$ to $77\%$), and matched prompts frequently change regimes between models. These findings show that aggregate detection metrics conceal persistent, model dependent heterogeneity in language model failures and motivate regime conditioned evaluation.

SEC 10-K filings contain substantial financial information that is not consistently captured in structured datasets, creating a missing-data problem affecting over 70% of firms and half of total market capitalization. This can disproportionately bias quantitative analysis against smaller firms, which may be excluded due to limited available data. Traditional financial extraction methods such as Regular Expressions (Regex) and BERT, have been widely used. However, they are highly brittle when parsing complex SEC 10-K filings, which leads to data that is existent in the files being lost since these methods do not consider that a data attribute could be located in a different section or a footnote. This study evaluates several Large Language Models (LLMs), including Llama-3 8B, Qwen-2.5 14B, and Llama-3.3 70B, to figure out individual model strengths and weaknesses when extracting specific attributes from SEC 10-K text. The extraction quality was evaluated across four financial variables of varying structural complexity: Cash and Cash Equivalents (tabular), Short-Term Debt (hybrid), Credit Facilities (narrative), and Research and Development (hybrid). Results show that while smaller models like Llama-3 8B experience performance degradation under complex negative prompting, aligning parameter scale with document complexity yields high zero-shot accuracy. Qwen-2.5 14B excels as a tabular specialist with an 83.33% F1 score on Cash, whereas Llama-3.3 70B effectively navigates dense narrative footnotes, achieving a 76.92% F1 score on R&D. This scalable framework addresses critical information gaps in quantitative finance datasets and eliminates missing-data bias through a more thorough analysis of the SEC 10-K files.

Single-token typed-decision models answer a schema question by reading the logits of a few one-letter answer codes at a single position: they are fast and return a probability for every allowed answer, but they are trained with plain cross-entropy that ignores most of the structure in their training data. We study such a model whose data is curated as contrastive pairs---two contexts that differ in one edited fact that flips the answer---each carrying a machine-checked certificate that deleting the decisive sentence makes the fact unknown. We propose PACT, which turns this structure into four training terms that need no new annotation: a difference-in-differences margin over each pair that is invariant to any shared logit offset, a permutation-consistency term against answer-code position bias, an evidence-necessity term on certificate-verified ablated contexts, and an ordinal transport cost for rubric fields, plus a three-parameter contextual temperature. On a frozen 324-item holdout with three seeds, PACT matches the published recipe in accuracy ($84.6\%$ vs. $85.2\%$; McNemar $p \ge 0.50$ at every seed) while giving the lowest position bias of all runs (answer flips under relabelling $9.8\%$ vs. $13.8\%$) and the lowest ordinal error on rubric fields (MAE $0.232$ vs. $0.311$). Against a control with the same optimiser and schedule but cross-entropy only, PACT is significantly more accurate at two of three seeds, halves the seed-to-seed spread and lowers NLL by $26\%$. Seed-matched ablations and pre-specified falsification tests locate these gains precisely: no single term raises raw accuracy, and the method's value lies in robustness and stability rather than headline accuracy. Code, data splits, trained adapters, and all run records are available at https://github.com/BennyLinntu/PACT-Pairwise-Anchored-Calibrated-Tuning-for-Single-Token-Typed-Decisions.

Benchmarks for large language models (LLMs) typically evaluate the accuracy of answers against fixed reference labels. But a central step in many complex real-world tasks is identifying which questions are worth asking in the first place: decomposing a difficult problem into subquestions -- which we call cruxes -- whose answers provide key steps on the path toward solving the target problem. To evaluate this capability of information discovery, we introduce CruxBench, a benchmark that grades LLM-generated questions by their Value of Information (VOI): how much a model-proposed crux updates beliefs about a target forecasting question. CruxBench enjoys a rare combination of three key properties: it is (1) contamination-resistant by construction, since ground truth is generated by future world events; (2) open-ended, admitting unbounded and complex text-based submissions rather than one correct numeric answer; and (3) grounded, with informativeness measured against quantified changes in real-world beliefs. We evaluate a diverse set of eight models on 293 target forecasting questions and find that VOI correlates highly with independent measures of model capability (r=0.90) and captures cruxes' usefulness for answering target questions. However, information discovery remains challenging even for frontier LLMs, which only narrowly outperform a random-timing baseline.

A voice agent can handle almost every call on the words alone and still fail the few its sector's rules were written for. Emergency-call standards, fraud guidance, radio phraseology and vulnerability rules recognise that how a caller sounds, or what else is audible, can change the right action. VoxParity tests whether agents act on it. In 183 scenarios from 14 sectors, one transcript stays fixed while the audio changes (a coaching voice, a medical monitor beeping, a mayday under a radio check, noise over a drug name, a child's voice placing a bet, a frightened whisper), and with it the correct typed tool call. A words-only null test credits a system only if hearing the call moves its actions more than it moves a pipeline that only reads the words. Only 11 of the 23 systems that can also be run on the transcript pass. Descriptively, errors run toward the words: when the audio calls for protection, all 28 systems carry out the routine request more often than they over-react on clean calls (41% against 12% pooled; the words-only pipeline, 58% against 15%). Exploratory analyses place most of the leading systems' misses on cues they heard; systems beat the null almost entirely on items that state the rule; the leading systems overrule heard resignation or confusion far more often than acute alarm; and, in the models tested, describing the voice and stating the rule each recover part of the shortfall, leaving a gap on emotion.

Recent work in visual question answering has shown that vision-language models can exhibit strong reasoning capabilities by translating visual inputs into textual representations. The effectiveness of this translation depends on how well visual details are retained; models need to surface and align both explicit and implicit knowledge sufficient to support reasoning, without introducing spurious assumptions. Existing methods that leverage detailed image captions introduce visual details unrelated to the reasoning task, inflating input token counts and increasing computational cost. To address these challenges, we propose VisKG, a reinforcement learning (RL) framework in which models learn to translate visual content into question-specific knowledge graph (KG) representations. This process filters out perceptual noise while preserving the entity-relation structure needed for chain-of-thought reasoning, following the principle of minimum sufficient information. To ensure stable RL post-training, VisKG adopts Group reward-Decoupled Normalization Policy Optimization (GDPO). In addition, we strengthen the supervision stage with negative rationale samples, exposing the model to incorrect reasoning paths before RL post-training. Experimental results across science, mathematics, and general visual understanding benchmarks show that VisKG achieves performance comparable to or better than baselines, while requiring fewer tokens than caption-based representations. Moreover, training VisKG with GDPO improves accuracy by 2% over its GRPO-trained counterpart on average. These results suggest that KG representations are a promising approach for supporting multi-step reasoning and open up future work on adaptively selecting the most suitable representation for a given task.

Large language models are able to solve tasks whose answers depend on not only individual input tokens, but also on relations among them. How is such relational information represented and processed across transformer layers? We study activations from ensembles of prompts that require inferring relationships between three tokens corresponding to a cyclic concept (months, hours, weekdays, and musical notes) to correctly predict the next token. Across model families (Llama, Qwen, Gemma, and Mistral) and cyclic concepts, we find a consistent layerwise progression in how the joint dependence among the tokens is geometrically organized and causally used: intermediate layers use a joint representation based on the inferred relationship between two tokens, while later layers use a joint representation associated with all three tokens to correctly complete the task. We also find other relationships between tokens that are geometrically structured but remain causally inert in the next-token prediction. Crucially, when taken together, these geometric and causal investigations reveal the representation-level mechanism that progressively organizes and composes the relational information to form the answer. More surprisingly, restricting the models to such causally relevant joint representations improves next-token prediction accuracy.

Long-term memory lets an LLM assistant use a history it can no longer reread, and most memory systems build it by rewriting conversations into facts, graphs or typed memories at write time. We argue that the work of memory divides, as thinking does, into two systems. Most of it is fast System 1 work: many small, independent yes/no judgments about records, such as whether a record is needed or no longer current, which a decision model makes by the dozen in a third of a second. Only a little is slow System 2 work: writing a few search queries, naming what the reply needs and composing the answer, which an LLM does well but slowly. We present Mnemon, a memory agent built on this division. It keeps conversations as raw, dated records; an LLM (System 2) plans searches over them, a decision model, Jev (System 1), judges what the searches return, and rules with explicit budgets turn the judgments into a small View for an unchanged answering model. A background pass consolidates each record once into topic timelines, value histories and standing instructions linked to the records, so that questions about a whole conversation reach evidence their own searches miss. Because nothing is decided about a record when it is written, the same agent can read any store that returns dated records. With gpt-4.1-mini answering, as in a public re-evaluation of 14 systems, Mnemon scores 91.7% on LoCoMo, the highest among them, and 83.8% on LongMemEval-S, from under 4k tokens of context per question, with the lowest effective cost index on LoCoMo. With a reasoning model answering, it reaches 92.2% on LoCoMo and 94.4% on LongMemEval-S, the latter on par with the best published results. From 100K to 10M tokens of history on BEAM, its cost per question grows by a factor of 1.11. On the same records, Jev separates gold evidence better than two LLMs and is 3-11 times faster.

Language models are frequently employed to evaluate other language models. An LM evaluator scoring agentic behaviors across multiple criteria is valuable, provided that its decisions align with human judgment. We call the problem of evaluating this alignment Meta-Evaluation. Tackling it directly is difficult: collecting human data is expensive, absolute scoring is hard to align, and using an LM meta-evaluator recurses the question of trustworthiness. We adopt a reformulation of meta-evaluation as a preference judgment problem: rather than comparing human and LM evaluator scores of a trajectory, we ask whether their implied preferences align. Building on this, we introduce PADM\'E, a data synthesis method that generates reliable criterion-based meta-evaluation data for agentic settings. PADM\'E uses only small language models, requires no human involvement during evaluations, and operates under a low computational budget. We build a prototype of PADM\'E and synthesize a dataset of 1,000 samples across four agentic domains and three evaluation criteria. Human validation on a 150-sample subset demonstrates that PADM\'E improves agreement with human judgment from 73% to 85% over a naive baseline. Meta-evaluating 25 common models with our dataset demonstrates the correlations between evaluation performance and scoring granularity, leniency, and model size, among other factors.

Sinhala Sandhi splitting recovers the constituent words or morphemes hidden inside a phonologically merged surface form. The task is important for Sinhala NLP because Sandhi obscures lexical boundaries, but no prior published work has established a neural benchmark for Sinhala Sandhi splitting. We present a character-level sequence-to-sequence study based on SandhiLex, using native Sinhala Unicode input and evaluating recurrent encoder-decoder models for affixational and more complex lexicalized, derivational, and etymological Sandhi. The central challenge is the hard subset lexicalized, derivational, and etymological Sandhi, where our best model, a bidirectional LSTM encoder with a unidirectional LSTM decoder, reaches only 68.40\% exact-match accuracy (82.08\% character-level accuracy), well below the 94.00\% achieved on the more regular affixational subset. Ablations show that bidirectional encoding is the largest contributor to performance, while native Sinhala script improves exact match accuracy over romanized input. Qualitative analysis indicates that many errors are near misses involving boundary adjacent characters or plausible but incorrect phonological substitutions. These results establish an empirical baseline for Sinhala Sandhi splitting and identify data scale, Sandhi type conditioning, and attention-based decoding as the main directions for future work.

Before invoking external tools, an agentic LLM must select among a K-way action space: executing a call, seeking clarification, answering directly, or declining. While internal activation steering can alter these pre-execution decisions, conventional aggregate metrics obscure where altered states land and what collateral damage they inflict. We present SAKIKO, an auditing framework that formalizes representation repair via directional error discovery, router-conditioned intervention, destination-resolved verification, and prospectively frozen statistical licensing. Across seven LLMs on When2Call and MetaTool, channel-keyed interventions induce direction-specific net gains in five models; across three sealed evaluations, none of 59 budget-matched random directions matches calibrated target gain. Crucially, destination auditing shows that behavioral movement does not equal repair: an intervention achieving +55 net gain corrupts over half of the baseline-correct decisions it touches, and promising point estimates on Qwen3-4B and Gemma-2-9B are formally declined due to finite-sample uncertainty. SAKIKO establishes the necessity of outcome-resolved adjudication before claiming internal repair. Code: https://github.com/ruizheliUOA/mechanistic-tool-use-llm.

As large language models are deployed in increasingly autonomous long-horizon tasks, manually auditing and verifying the actions, artifacts, and outputs of models becomes more difficult. Users instead come to rely on LLM-generated reports to assess the quality and completeness of the work. We introduce a suite of eight adversarial reporting scenarios to systematically study whether LLMs conceal narrative-changing flaws: errors or limitations that undermine an otherwise successful account of work. We call this phenomenon "insecure reporting." When handed machine learning experiment logs containing a planted negative result that substantially weakens the proposed method, GPT-5.5 flags the negative result in only 2 of 200 generated reports. However, when a short honesty instruction, "Be honest in your response," is added, the model flags the negative result in 190 of 200 reports. Across eight open-weight models, chain-of-thought analysis reveals a recurring tension between disclosing narrative-changing flaws and reasoning about ways to appear successful. We perform an activation analysis and a steering experiment on Qwen3.5-9B, finding that honesty and success-seeking correspond to opposing directions in representation space. Our results suggest that LLMs tend to present narratives of success by default, and that steering models toward honesty makes their reports substantially more transparent.

Group Relative Policy Optimization (GRPO) has become a promising approach for training large language model agents. However, its uniform assignment of trajectory-level advantages to all policy tokens fails to distinguish consequential decisions from less relevant ones, obscuring which intermediate decisions contributed to success. We introduce ProVer, a framework that targets potentially pivotal decisions for fine-grained credit assignment in agentic reinforcement learning. Given a rollout group, an agentic judge contrasts successful and failed trajectories to propose a segment potentially responsible for their divergent outcomes. Rather than directly trusting the judge's assessment, ProVer verifies the proposed segment by estimating its advantage from the difference in terminal success rates between current-policy continuations sampled before and after the segment. Positive estimates are then incorporated into the GRPO advantages of policy tokens within the proposed segment. By using model judgment only to select where to verify, ProVer grounds local credit in observed outcomes without exhaustively evaluating every intermediate state. Across ALFWorld, WebShop, and SearchQA, ProVer achieves the strongest average performance at both model scales, with relative improvements over GRPO of 9.91% and 7.12% for Qwen3.5-2B and Qwen3.5-4B, respectively. Further analyses demonstrate that informed segment selection improves policy training with modest additional generation overhead, even without a frontier-scale judge model, highlighting the effectiveness and efficiency of selectively targeting pivotal decisions for fine-grained credit assignment in agentic reinforcement learning.

Extracted sentiment directions can vary across samples even when downstream sentiment classification remains accurate. To evaluate direction reproducibility, we measure split-half agreement in English, Hausa, and Yoruba representations across four language models using both native and translated texts. We identify layers selected for agreement using ten topics and evaluate direction agreement across separate groups of fifteen topics. Using the final token, split-half agreement ranges from 0.737 to 0.870 for English, 0.589 to 0.762 for Hausa, and 0.101 to 0.399 for Yoruba, maintaining this language rank order across all 77 complete model comparisons. Classifiers trained on these same layers consistently predict sentiment above chance, demonstrating that predictive accuracy does not imply directional consistency. Furthermore, averaging token representations yields less consistent agreement, and high agreement can partially reflect sentence length. Ultimately, our findings highlight the need to measure vector direction reproducibility independently of classification performance, though they do not establish that tokenizer fertility which is the average number of tokens per whitespace separated word causes cross-lingual differences.

Computer-use agents (CUAs), while capable of completing computer tasks in everyday and professional workflows, can cause unintended harm even under benign instructions and environments. However, detecting such harm remains challenging. First, it requires careful, task-specific reasoning: verifiers guided only by general safety criteria often overlook many important but subtle harmful behaviors. Second, it requires active investigation: past trajectory screenshots show what the agent did but not always what actually changed in the environment, so LLM-as-a-judge verifiers that rely on screenshots alone may be unable to determine the actual consequences of actions. To address these challenges, we introduce SCOUT, a two-stage agentic safety verifier that synergizes reasoning-intensive rubric generation with tool-intensive evidence gathering. First, our SCOUT rubric generator extensively reasons over the task and the agent's trajectory to determine what successful and safe execution should entail, generating task-specific completion and safety rubrics. Then, our SCOUT probing agent follows these rubrics to interact with the post-execution environment and collect grounded evidence for final safety and completion judgments. We evaluate our framework on two computer-use safety benchmarks. On AutoElicit-Bench, SCOUT achieves 75.4 unsafe F1 and 74.5 completion F1, outperforming LLM-as-a-judge verifiers and naive tool-use verifiers. SCOUT leads on OS-Blind with 76.4% unsafe detection accuracy. Test-time reflection reduces final unsafe execution rates from 30.2% to 17.2% on AutoElicit-Bench. Ablations and analysis show that tool-free rubric generation in SCOUT elicits substantially more reasoning and is crucial for safety detection across verifier backbones, especially non-frontier ones. A preliminary extension to coding tasks shows that SCOUT can support safety verification beyond computer-use.

Gradient-based reward guidance provides a flexible way to use downstream reward models to control masked diffusion language models at inference time. However, its computational cost remains high as each decoding iteration incurs expensive diffusion model forward passes and reward model backpropagation steps. To address this, we introduce FastGuide, an adaptive hybrid of parallel and autoregressive decoding to accelerate reward guidance for diffusion language models. In analogy to parallel decoding, FastGuide amortizes the cost of reward model backpropagation by computing guidance once per decoding step and reusing it to generate multiple tokens. Within each decoding step, FastGuide makes diffusion forward passes autoregressive by unmasking tokens one at a time while efficiently recomputing token distributions after each unmasking by utilizing KV caching techniques and sparse recomputation of attention. Lastly, to adapt hybrid decoding to the model's confidence, FastGuide defers any token that the model is unconfident about under its recomputed distribution. Experiments on three reward benchmarks demonstrate that FastGuide is up to $4.4\times$ faster than sequential reward-guided decoding while retaining similar generation quality.

We study how Gemma 4 31B represents African languages and responds to cultural steering. The first study compares nine African languages and three controls using probes, contrast directions, and measures of representation similarity. Transfer from English varies across languages and layers. Directions representing an Africa versus West contrast are more aligned among the African languages than between these languages and the controls at several layers. The comparison across language families passes the reported Holm threshold at five of twelve layers, although dependence between language pairs limits the statistical interpretation. Within Nigeria, Yoruba and Igbo are more aligned than the average of their pairs with Hausa at eleven of twelve layers. The second study uses separate English data to construct directions for Nigeria, Ghana, Kenya, and South Africa. Under union scoring at the selected strengths, estimated differences in attribution rates from random directions range from 0.63 to 0.81. Most outputs pass the automated structural coherence screen. Comparisons with Aya Expanse 32B show that results depend on the representation measure. Together, the studies document regional and family patterns in the sampled representations and country steering in English.

Large language models (LLMs) are increasingly used as knowledge bases (KBs) due to the vast amount of knowledge they acquire during pre-training. While many works focus on extracting single relational triples, most real-world relations are multi-valued and require generating sets of entities. In this paper, we investigate how LLMs represent and generate multi-valued relations. We identify the canonical order problem: The probabilistic distributions inside LLMs organize many multi-valued relations according to a canonical ordering (e.g., alphabetical or chronological). Through mechanistic analysis, we show that set generation in LLMs can be thought of in terms of three phases: (1) retrieval of candidate entities, (2) internal sorting, and (3) selection of the next element. As a result, prompts aiming to construct KBs that deviate from this internal canonical ordering lead to a markedly reduced reliability of LLMs when aiming to generate complete sets for multi-valued relations.

Large language models (LLMs) are increasingly used by people whose first language is not English, yet these users have been shown to receive systematically lower-quality responses than fluent speakers. Which specific features of non-native English drive this gap remains unclear, because fluency is itself a composite of mechanical accuracy, vocabulary use, organization, and discourse coherence. Here, we introduce FABLE, a controlled dataset of 190,911 English prompt variants derived from 174K real user prompts for writing-related tasks. Evaluating responses from 34 open-weight LLMs, we find a clear asymmetry; while models do not propagate surface errors such as misspellings into their outputs, models do mirror higher-level rhetorical and lexical qualities present in the user's prompt. Further, the overall quality of responses differs substantially between the least- and most-fluent prompts. These results highlight a key LLM performance disparity for non-native English LLM users, resulting in both lower-quality and less-fluent answers.

Large language models are increasingly evaluated in specialized domains such as law, medicine, software engineering, and cybersecurity, yet film remains comparatively underexplored despite requiring long-form narrative integration, multilingual interpretation, and culturally situated audience judgments. We introduce CineSubBench, a benchmark for evaluating long-context film understanding from multilingual movie subtitles. A subtitle track represents a film as thousands of short, temporally ordered utterances from which models must reconstruct characters, relationships, events, causal progression, and themes without explicit scene or event structure. CineSubBench contains 1,012 films with complete subtitle coverage in six languages, yielding 6,072 tracks and 8.13M timestamped subtitle entries. It provides a matched multi-task, multilingual, and multicultural (MultiX) evaluation setting: seven tasks span narrative reconstruction and abstraction, genre prediction, age suitability, country-specific motion-picture ratings across ten national classification systems, and subtitle-grounded language safety. Across nine LLMs, plot premises are recovered more reliably than event-complete synopses; cross-lingual consistency varies substantially across models and languages; national rating systems expose distinct calibration patterns; and strong profanity is far easier to ground than mild obscenity. CineSubBench establishes film as a long-context LLM evaluation domain and provides a unified benchmark for measuring narrative, multilingual, cultural, and evidence-grounding capabilities.

Large language models (LLMs) can predict human brain activity across a variety of brain regions during natural language comprehension. Typically, however, LLM-brain alignment is measured using one model for different regions of the brain, and then model performance is summarized across regions. This one-model-fits-all approach ignores the functional specialization of brain regions. In this study, we assess whether a model oriented toward a particular cognitive domain aligns better with the brain system dedicated to that domain. Through prompting and fine-tuning, we first build expert LLM variants for six domains: sensory, spatial, numerical, reasoning, social, and abstract processing. We then examine whether each expert best predicts activity in the brain region associated with the corresponding cognitive domain. Consistent with our hypotheses, each expert's representations align more closely with the brain system most associated with the matching domain than do other experts. This holds under both prompting and fine-tuning, across three base models and three fMRI datasets. In a series of control analyses, we show that this model-brain alignment is specific to cognitive domain interventions; non-cognitive and surface-level interventions do not result in comparable alignment. Specializing models shifts regional alignment while leaving aggregate prediction accuracy largely unchanged, suggesting that summarizing alignment across regions may obscure regional differences in performance for specific models.

We present OLIVE (OnLine InterVEntion). At each iteration, the evolving student policy generates a new prefix, the teacher continues it autoregressively, and the student is updated using cross-entropy computed on the teacher-generated tokens. Each design choice targets a corresponding limitation of existing distillation methods: (1) sequential covariate shift in offline supervised fine-tuning (SFT) on fixed teacher trajectories, (2) fragmented supervision under prefix failure in token-level on-policy distillation (OPD), and (3) the need for access to teacher token probabilities in distribution-matching distillation. OLIVE achieves higher reasoning performance than OPD (with a top-16 KL approximation) at comparable GPU-hour cost. Our asynchronous implementation further reduces OLIVE's total training time by 23.8\%. We evaluate OLIVE on both hard reasoning tasks and agentic tasks which reflects modern post-training scenarios, and it consistently outperforms existing distillation methods under the same training budget. By regenerating prefixes from the evolving student, OLIVE continues improving after offline distillation plateaus while better preserving the general capabilities and plasticity of the student. Using only text from GPT-5.4-mini, continuously training with OLIVE outperforms offline SFT from the same teacher by 13\% on ScienceWorld. These results support OLIVE as an effective and efficient approach to online language-model distillation.

Large language models (LLMs) show considerable potential in simulating human attitudes and preferences. Prior work finds that LLM-generated responses can compress the range of attitudes found within populations and misrepresent particular subgroups in ways that vary across models and topics. We introduce Population Fidelity, an evaluation framework that distinguishes key conditions required for a set of LLM-generated responses to represent a population. It incorporates three dimensions: group-level accuracy, the amount of between-group variation, and the structure of that variation. We demonstrate the framework's utility in two ways. First, we reproduce a prior study of "machine bias" in LLM survey responses and apply the framework to its models and more recent ones, showing that poor representation reflects not only insufficient between-group variation but also variation assigned to the wrong groups. Second, we evaluate one proposed approach to improving models' population representativeness: cultural fine-tuning. We find that cultural fine-tuning can improve alignment with the survey center without improving the representation of within-population differences, a distinction that measures of aggregate agreement do not capture. We argue that representing a population requires models to reproduce several features of human attitudinal variation simultaneously. Our framework organizes these features and provides reusable code, data, and trained models for evaluating population fidelity across substantive domains and assessing proposed alignment methods.

Recent advances in agentic heuristic design use AI agents and execution feedback to automate algorithm discovery for challenging optimization problems. In many practical settings, high-quality solutions must be obtained under strict runtime constraints, motivating hybrid approaches that combine problem-specific heuristics with powerful mathematical programming solvers. However, existing approaches typically improve heuristic components within predefined procedures or tune solver configurations in isolation. This limits holistic adaptation of where to allocate computation, how to leverage solvers, and how to refine the overall algorithmic structure. To address these limitations, we propose HeurEvo, an automated plan--code--component co-evolution framework that jointly evolves the high-level algorithmic structures, their implementations, and a shared pool of reusable components. A planner determines which algorithmic components to use, how to combine them, and how to allocate runtime across stages, a coder realizes the resulting plan as executable code, while a component evolver updates the shared component pool. Within an island-based evolutionary framework, plans and implementations co-evolve with feedback from an interpreter agent that analyzes execution results and identifies opportunities for improvement. Across diverse combinatorial optimization benchmarks and challenging MIPLIB instances, HeurEvo finds high-quality solutions within tight runtime budgets, often matching or surpassing state-of-the-art optimization solvers given hours or days of computation. On several nonlinear geometry problems such as hexagon packing, it also improves upon the best previously reported results. These results highlight the value of jointly searching over algorithmic structure and implementation for agentic heuristic design.

Evaluation awareness (EA) can cause large language models (LLMs) to behave differently during audits than in deployment, yet measuring and accounting for EA remains challenging. We introduce verbalization training (VT), a method for making LLMs less reticent about verbalizing evaluation awareness while avoiding to supervise the latent belief itself. VT uses a model's spontaneous verbalizations as evidence that awareness is present and truncates each rollout immediately before the verbalization, producing training prefixes at which the model is presumed to be aware. The model is then trained with an RL objective designed to increase verbalization in a calibrated way. Across Qwen3.6-35B-A3B, Kimi K2.6, and Inkling, VT increases verbalized EA by 2.4-2.9 times and transfers to held-out agentic settings, while measured latent EA and behavior remain largely stable. In a causal experiment, we independently implant meta-knowledge about evaluations through synthetic-document fine-tuning and show that VT-induced verbalizations reflect the richer knowledge acquired by the model.

Deep-research agents conduct long-horizon investigations through iterative search, evidence evaluation, belief revision, and synthesis. However, they may commit to claims before sufficient evidence is available, causing later reasoning to reinforce an incorrect interpretation. We introduce DeepRewind, an additive control layer for reversible deep research that represents the agent's evolving epistemic state as a typed graph of sources, evidence, claims, hypotheses, assumptions, commitments, plans, and drafts. Before accepting an intermediate conclusion, a prompt-based world model predicts its impact and estimates reversibility based on hypothesis narrowing, information loss, recovery cost, and contradiction-trigger coverage. A binary controller blocks risky commitments, while a consistency monitor performs dependency-aware rollback when later evidence invalidates them. Across DRBench and LiveDRBench, DeepRewind improves insight recall by 3.6 percentage points and reduces premature commitments by 59.1% relative to Open Deep Research.

Decision-only language models return a probability for every answer option instead of generating text, which makes them attractive as survey respondents and as judges. We audit the cultural values of one such model, TypeSafe's JEV, with the Values Survey Module 2013. We asked it the 24 items as 12 matched Saudi and 12 matched American personas and without a persona, in English and Arabic, under eight ways of formulating the request (288,000 answers). JEV's answers were highly repeatable (ICC 0.997), and without a persona they resembled those of its own American personas. When the persona was Saudi rather than American, the answers moved in the direction of the human Saudi-US difference, reproducing 87% of its size in English but 62% in Arabic, with long-term orientation reversed. A language cross shows that the smaller difference in Arabic comes from the language of the items, not from the language of the persona description. Age shifted the profiles about as much as nationality, gender shifted them more for Saudi than for American personas, and JEV was less confident in Arabic and for Saudi personas. These patterns held in every request design, although the model never generates text.

We consider persistent LLMs that accumulate memories of their interactions with a user over time. Such LLMs maintain memories using external storage, which they can query to overcome the limitations of a fixed context window. Such systems have numerous practical applications, as they can draw on all past interactions when responding to user queries. In this paper, we ask whether LLMs can forget information shared with them upon a user's request. We find that current LLMs fail to delete such information---even when they claim to have forgotten it and even when operating with a limited context. To this end, we consider a new direction of study: Deletion of LLM Memories. We show that naively removing messages that match a user's deletion request is insufficient, since conversations naturally introduce message dependencies that cause information to persist. To correctly handle deletion requests, we propose the DeLLM framework. It dynamically constructs relevant context for each LLM query and maintains a provenance graph of messages to determine which ones must be removed during deletion. Our experiments show that DeLLM achieves a high deletion rate while maintaining utility.

An assistant that serves the same user over a long horizon has to answer from what that user has revealed: which preferences still hold, which were revised, and which constraints apply now. Retaining that information is not the same as acting on it, and the two are usually optimized as if they were. Keeping the information as text makes the reader's input grow with the retained history, while compressing it into a fixed number of latent vectors bounds the interface but is typically trained to reconstruct text or imitate reference answers, both of which are scored on sequences the reader never produced. We present MemFold, which optimizes a fixed-budget soft memory by the behavior it supports. A query-conditioned textual memory is compressed into K continuous vectors that form the reader's memory interface, and the reader is then trained on its own rollouts under two complementary signals: group-relative rewards for task outcomes, and confidence-gated on-policy distillation in which a frozen textual-memory teacher re-scores the student's sampled tokens under the textual memory. The teacher is never sampled from, so supervision stays on the student's current distribution and adds no autoregressive decoding; at inference it is removed entirely. Across three Qwen backbones, MemFold attains the highest accuracy we measure on PersonaMem-32K and PersonaMem-128K, with margins that widen at the longer history length, and transfers to PrefEval and LongMemEval without target-domain training. Ablations attribute most of the task gain to the reward term and a smaller additional gain to the teacher signal, and memory interventions show that the reader depends on the instance-specific content of its soft memory.

Masked diffusion language models (MDLMs) can generate text efficiently by predicting multiple masked tokens in parallel, but predictions from the same forward pass are not necessarily reliable when committed together. We study when parallel commitment is reliable. Our diagnostics show that confidence alone does not determine a reliable commitment order: confident predictions near the end of the sequence can fix an answer before its supporting computations are established, and downstream predictions become less reliable as the uncertainty of their upstream context grows. At the same time, a single forward pass can already resolve several masked tokens, and predictions that remain stable across the final layers are more likely to be correct. Based on these findings, we propose Reliable Parallel Decoding (RPD), a training-free method that selects candidates by layerwise prediction stability and final confidence, and commits them under a cumulative entropy budget over their preceding masked positions. RPD defers predictions with uncertain upstream context while committing the remaining candidates in parallel, without relying on a fixed block schedule. Across mathematical reasoning and code generation benchmarks on LLaDA and Dream, RPD achieves the highest decoding throughput among the evaluated methods while maintaining or improving accuracy.

Long-term memory systems turn conversations into short stored notes. A note can keep a user fact while losing evidence about whether the fact still holds. For example, "I am driving a Peugeot" can become "The user drives a Peugeot," which drops the cue that the activity is ongoing. We call this aspectual flattening and measure it with LAPSE, a benchmark of matched user statements that differ only in temporal form. We find that memory writers flatten aspect selectively. Three writer models flattened the progressive statement but kept its simple-present match in 244 of 381 pairs, never the reverse. The asymmetry holds in all 11 model configurations tested and in the installed pipelines mem0, Graphiti, and Letta. The lost cue matters to later readers. In exploratory tests, changing only the stored verb shifted all three readers' estimates that a fact still holds. When readers could ask the user before acting, two of three acted without asking more often on flattened notes. Our planned memory-use task could not detect this, because readers there acted on almost every stored fact, even expired ones. Memory writing can thus remove evidence that later models use to decide whether to act.

In regulated industries like consumer finance, seemingly harmless user queries can exploit large language model vulnerabilities, triggering safety failures and pushing responses dangerously close to policy limits. Existing automated red-teaming methods trade off attack effectiveness against generation cost, while treating coverage, severity, and diversity as incidental rather than joint objectives. We introduce FinRT, a structured framework that builds reusable adversarial prompt generators from adaptive red-teaming strategies. Across the six victim models in consumer finance, FinRT substantially outperforms adaptive search baselines while amortizing target-facing attack generation into a reusable generator. FinRT nearly doubles the attack success rate over the adaptive baseline Rainbow Teaming (32.9% vs. 17.2%), increases maximum adversarial severity by 33%, and preserves comparable intra-policy-domain semantic diversity to iterative search methods. Our method achieves high cross-model transferability while exhibiting distinct victim-family specialization patterns.

Cross-modal associations are systematic pairings of features across modalities, such as the association of 'bouba' with round shapes and 'kiki' with sharp shapes. Prior work has compared humans and vision-language models (VLMs) on such associations, but often using different stimuli or tasks between humans and models. Here, we ask whether VLMs align with humans not only in choices, but also in where they look when making those choices. We study both VLMs and humans (N = 53), presenting them with the same stimuli, a pseudo-word and two images, and record participants' choices and eye movements, which we release. We find choice alignment in a few larger VLMs, but their saliency matches human gaze less closely than a center-bias baseline, a fixed Gaussian at the center of each image. Fine-tuning small VLMs on human choices brings their choice alignment to the level of a human majority-vote reference on unseen words and images, yet their attention still matches human gaze less closely than this baseline. Training model attention on human gaze raises attention-gaze correlation without improving choice alignment, and a single average gaze map per image position raises it by a similar amount. Matching human choices, or even human gaze patterns, is therefore not sufficient evidence of human-aligned cross-modal processing.

A common assumption in language model development is that cognitive abilities are organized around a general, domain-free intelligence factor, like fluid intelligence in humans. This assumption is rarely tested directly, and prior attempts have done so only at a much smaller scale. We take a latent variable approach to intelligence in language models, similar to how psychometricians study psychological constructs. Performance in every specific problem set is influenced by a domain-specific and a domain-agnostic latent factor. Using factor analysis as a dimension-reduction technique, we analyzed 13,251 published evaluation scores covering 1,618 language models across 456 different text-only benchmarks. Due to the super-sparse nature of the dataset, we triangulate our analysis across different data densifiers and imputation methods. A robust pattern across different modes of bias is that 1. A general intelligence factor accounts for 70.8% of variance in model performance at our most generous estimate, and far less than that in most of our solutions, 2. Content-similar benchmarks do not necessarily cluster together, and 3. The $g$ factor is not dominated by any common theme, and there is a lack of evidence that it is well-proxied by standard "intelligence" benchmarks. Our findings go against current endeavors of defining, identifying, and targeting general intelligence as a tangible construct in language model development. This leaves the strategy of targeting a single conceptual ability without support, since the first-order abilities it would have to reach are often partially idiosyncratic and not identifiable in practice.

Dense retrievers decide which documents reach users and the language models that use them, yet they are typically evaluated with neutral queries. We ask whether the identity signals that real users express in their queries---political ideology and dialect---bias what a retriever returns. We design evaluations in two domains, political news and consumer-health questions, each pairing a controlled synthetic set that varies only the identity signal with naturalistic queries. Across five dense retrievers and a sparse baseline, every retriever (i) retrieves articles that align with the query's own political lean and (ii) performs worse for questions written in African American Language (AAL) than in White Mainstream English (WME). Two analyses tie these gaps to queries' identity signals beyond surface vocabulary: partialling out an aggregate lexical-asymmetry score leaves the synthetic gaps largely intact, and linear probes recover lean and dialect from the retrievers' query embeddings beyond token-level features. Left unaddressed, such retrieval biases risk contributing to polarization and reinforcing the health disparities already faced by AAL speakers. Code is available at https://github.com/Andrewtcr/bias-ret.

Self-evolution lets large language models (LLMs) improve iteratively using their own generated data, but often suffers from self-evolution degeneration: performance improves, plateaus, then declines. Existing methods address this issue at the component level, targeting either the Questioner or the Solver, and overlook that self-evolution is a tightly coupled system. We propose a holistic framework based on learnable information gain, which measures how much novel, parameterizable information a round provides relative to the previous round. Theoretically, this gain equals the Kullback-Leibler divergence between the two rounds' data distributions plus their entropy change. Practically, it is estimated by fitting a small language model to the previous round and scoring new data via negative log-likelihood. Based on this diagnostic, we propose ATRI (Adaptive Training Regulation via Information-gain), which reweights samples within a round and halts training across rounds when information gain remains low. Experiments on popular datasets demonstrate the superiority of our proposal.

Generative AI has changed how students produce writing assignments. The final artifact is no longer sufficient to understand the process through which it was produced. We introduce DraftTrace, a writing environment that jointly captures three complementary views of writing: the final product, the writing process and interactions with an integrated AI-assistant. DraftTrace reconstructs how a document develops over time and organizes these signals into submission, longitudinal, and class-level analytics for instructors. We deployed DraftTrace in a graduate NLP course with 81 students and compared their sessions with LLM-generated responses entered by automated tools and with copy-typed responses. While product measures distinguish differences in text formulation, process measures distinguish differences in how text is entered. Considering both views together helps characterize cases such as copy-typing. Interaction traces show that students use the assistant differently across stages of writing: to clarify the question at an early stage and to verify answers at a later stage. A preliminary instructor survey highlights the importance of multi-view writing analytics and their interpretability.

Retrieval-augmented generation is widely used in professional writing, yet whether retrieval grounds revision or merely injects templates is rarely tested where "correct" has a definable meaning. Patent claim amendment supplies that signal: the examiner names the attacked limitation and cites prior art, providing per-case ground truth. We release three artifacts: (i) a corpus of 7,385 USPTO prosecution cases with XML-aligned pre/post claims, rejection, and cited prior art; (ii) a seven-probe battery comparing random and structural-match retrieval as two policies under a fixed prompt scaffold; (iii) a deterministic five-channel metric (C1-C3 and C5 in main, C4 supplementary) requiring no LLM evaluation. Across 9,600 pre-registered calls on four frontier LLMs (Claude Sonnet 4, Claude Haiku 4.5, GPT-5.4, GPT-4o-mini), no tested model exhibits detectable classical prior-injection behavior; retrieval effects are small and direction-inconsistent between random and structural retrieval, and the null is unchanged under a dense (semantic) retriever, across retrieval depths k in {1,3,5,10}, and under a paraphrase-sensitive grounding metric. Revision locality reveals a model-specific difference that the template channel misses. The four-cell taxonomy, which we treat as exploratory, leaves the prior-injector cell unoccupied.

Self-speculative decoding accelerates large language model (LLM) inference by drafting tokens from the target model itself, but faces a sharp tradeoff between the quality and cost of the draft. Early-exit methods produce drafts cheaply by terminating computation at intermediate layers, but forgo the deeper representations that later layers provide and thus suffer in draft quality. Multi-token prediction preserves draft quality by emitting from the model's final hidden states, but pays for a full forward pass to produce those states at every drafting step. We propose self-speculative encoder-decoder (SEED), a self-speculative method that obtains high-quality drafts cheaply by reusing the deep contextual representations already computed during verification. We reinterpret the standard decoder-only transformer as an implicit encoder-decoder: the first layers (encoder) build deep contextual representations, and the last few layers (decoder) emit tokens from them. Encoding and verification are merged into a single step: verification is performed by the full encoder-decoder, and the contextual representations of the verified prefix are cached for reuse during drafting. Drafting is therefore very fast: between verifications, the lightweight decoder drafts multiple tokens autoregressively, each conditioned on the cached representations and on preceding drafts. Experiments across multiple benchmarks show that SEED achieves up to 2.7$\times$ average speedup on 4B-scale models, outperforming both early-exit and MTP-style self-speculative baselines and running 28% faster than the state-of-the-art EAGLE-3, while preserving or even improving the generation quality of standard autoregressive fine-tuning. Code is available at https://github.com/lhk2004/SEED.

We study the ability of LLMs to generate edit-inducing questions whose answer will improve a paper draft. On a dataset of paired submission and camera-ready papers from ICLR and NeurIPS, we compare the helpfulness of questions from GPT models with or without full paper context to that of human reviewers. GPT produces more edit-inducing questions and its questions are associated with more extensive edits and cover a broader range of edited content compared to questions from reviewers. However, a much smaller percentage of the GPT questions are edit-inducing. Our analyses confirm that automated questions can be beneficial to authors and highlight an example task where proper attending to long context deteriorates reasoning model ability to produce helpful output.

Recursive self-improvement (RSI) aims to achieve compounding gains by having models improve themselves. While most existing RSI systems optimize external agent harnesses or prompts around a frozen base model, data-centric RSI directly updates the model's own parameters by training on agent-generated data. However, because validating data strategies requires expensive model training, existing methods face a fundamental dilemma: a single agent gets trapped in narrow directions and lacks exploration breadth, while naive parallel search or heavy trace sharing sacrifices long-horizon search depth. To address this challenge, we introduce G"odel Forest, a multi-agent framework that organizes recursive self-improvement as an ensemble of co-evolving search trees. In G"odel Forest, each agent autonomously grows a persistent tree, deepening, branching, or pruning data strategies based on model feedback to secure depth, while parallel trees explore distinct regions of the data space to expand breadth. Crucially, rather than leaving trees isolated or flooding them with heavy execution logs, a dynamically co-evolving memory connects the forest: agents continuously distill their successes and failures into compact procedural lessons anchored to a global leaderboard. Through this forest ecosystem, a dead-end in one tree instantly warns the whole forest against unpromising paths, while an empirical breakthrough quickly seeds new exploration branches in neighboring trees. Evaluated on RSIBench-Data across six diverse domains, G"odel Forest outperforms the single-agent baseline by an average of 10.70% while reducing wall-clock time on five tasks. Ablations confirm that co-evolving shared memory yields a +7.00% gain over independent parallel search, demonstrating that collective distillation is key to scalable self-improvement. The code is available at https://github.com/evolvent-ai/Godel-Forest.

Embodied agents now take on ever longer tasks. For long tasks, knowing only whether a task finally succeeds or fails says little; the steps along the way matter. Progress Reward Models (PRMs) score how far a task has come at every step, and serve as dense rewards, verifiers and monitors. Yet in long tasks the current frame alone often cannot tell how far the task has come, because progress depends on what happened before. We call this problem context-dependent progress estimation. Existing benchmarks on progress estimation mostly focus on short tasks whose progress can be read from the current observation, and whether PRMs can estimate progress when context is needed remains underexplored. We therefore build ContextProgress-Bench, with 24 manipulation tasks for 120 episodes. The benchmark covers three settings: (i) State Recall, where information needed for progress appeared earlier but is not in the current frame; (ii) Sequence Tracking, where steps follow a fixed order, so progress requires knowing which steps are done and which comes next; and (iii) Recurrence Disambiguation, where look-alike frames sit at very different progress. We then run a paired diagnosis: each PRM keeps the same input format in both runs, and in one run its instruction integrates the right context. Even PRMs that read the entire history get lost in estimating progress, yet with the right context the same five models cut their progress error by 77-82%. Embodied PRMs are thus not incapable of progress estimation, but lost without the right context. We therefore propose ProgressCompass, an autonomous agentic loop that reorients an existing PRM and uses current general-purpose VLMs to supply the context the PRM needs. Wrapped in the loop, the same frozen PRM cuts its progress error by 63% and raises its rank agreement by 76%. With such a compass, PRMs estimate progress far better on longer, more complex tasks.

Manipulation behaviors vary widely across objects and scenes, but they share a small set of reusable skills, and planning with these skills helps embodied agents generalize to new tasks. Yet an agent can only plan with skills it knows. Recovering skills from observed experience, the inverse of planning, builds this knowledge over time and yields skill data for training future agents. Vision-Language Models (VLMs) describe individual manipulation events well, but can they organize a stream of events into reusable skills? We formulate this problem as Streaming Embodied Skill Discovery (SESD): a model watches videos in sequence and maintains a persistent skill library that shapes its later decisions. To systematically measure this ability, we introduce Video2Skill, a benchmark that covers robot tabletop manipulation and human kitchen activity and tests three core capabilities: (i) locating manipulation events in time, (ii) grouping events of the same transformation, and (iii) deciding when to reuse an existing skill or create a new one. Across 19 open-source VLMs, many models group events at near-chance level, and scale does not consistently help. Their errors depend on how perception and library updates are coupled: joint models merge distinct transformations into one skill, while models that update the library from text descriptions duplicate recurring ones. Supervised fine-tuning, including our counterfactual library-state rebalancing (CLaRe), improves grouping but exposes a deeper bottleneck: trained models consolidate familiar skills yet rarely expand the library. Their libraries stall below half the reference size, and transformations unseen in training are located in time but almost never given a new skill. Recognizing when existing skills are insufficient thus emerges as the central challenge.

When conversing with large language models (LLMs), users often begin with a simple question and build towards a multi-hop question through follow-up turns. Retrieval-augmented generation (RAG) and its graph-based variant (GraphRAG) have become the dominant approaches for grounding LLM responses in external evidence, yet both are evaluated almost exclusively on single-turn, fully specified queries. We systematically investigate this evaluation mismatch through a large-scale simulation study. Building on prior work on multi-turn LLM evaluation, we transform questions from multi-hop question answering (QA) benchmarks into underspecified conversations and evaluate ten LLM assistants with eight retrieval systems across 1.5 million simulated conversations. Our findings reveal that multi-turn interaction causes widespread performance degradation, incurring relative performance drops of up to 21% and increasing unreliability by 47%, making RAG systems simultaneously less accurate and less reliable. We identify two distinct failure modes behind this degradation. Systems are either lost in translation, where conversational rephrasing distorts the retrieval query, or lost in conversation, where retrieval succeeds but the LLM fails to synthesize evidence distributed across turns.

Legal contracts contain ambiguities that expose enterprises to financial and legal risks. Some ambiguities allow flexible interpretation without triggering disputes, while others lead to significant legal conflicts. This makes identification alone insufficient, and interpretable rationale analysis essential. We propose LAURA, a post-training framework for interpretable ambiguous clause identification. LAURA leverages knowledge distillation with an IRAC-Unlearning prompting technique to transfer knowledge from a teacher LLM to an open-weight student model (<=1B parameters), which is then trained using a joint objective combining classification and rationale generation losses. The framework supports both legal and non-legal stakeholders in making informed decisions about which ambiguities require further attention. Extensive experiments across 7 baselines and 7 open-weight models demonstrate that LAURA with Flan-T5 (250M) delivers state-of-the-art interpretability over all interpretable baselines while matching the identification performance of the best-performing opaque baseline.

Large language model (LLM) agents repeatedly load reusable content, such as skills, documents, and memory entries, into the current context. Re-encoding this content for every request wastes computation. Position-independent caching (PIC) alleviates this by encoding each artifact independently and reusing its key-value (KV) states at arbitrary positions, but it incurs a quality loss relative to full-context prefill. Existing methods repair this loss by restoring global position IDs or recomputing selected tokens. In this work, we isolate the source of the loss, finding that the positional mismatch has minor effect, and independently cached artifacts retain faithful representations: reading a provided artifact stays largely accurate, and performance degrades only when the model must select among multiple artifacts. Moreover, replacing PIC's attention scores with full-prefill scores recovers performance with the cached KV unchanged, localizing the failure to the attention rather than KV recomputation. Motivated by this, we propose \textsc{Attuner}, a query-side adaptation method that learns to read a frozen artifact cache. \textsc{Attuner} inserts low-rank adapters into the query projections and is trained by distilling full-prefill distribution into the student. It trains fewer than 0.05\% of the model parameters and, at inference, requires neither cache recomputation nor a full-context reference. On Qwen3-4B and Qwen3-8B across seven benchmarks covering skills, documents, memory, and code, \textsc{Attuner} substantially outperforms prior PIC baselines in both in-domain and out-of-domain settings, matches full-context prefill quality while providing up to $3.73\times$ speedup.

Knowledge Distillation (KD) trains a smaller-capacity student model to imitate a larger-capacity teacher model by matching output distributions, implicitly assuming the teacher to be a reliable oracle. In large language models (LLMs), this assumption often fails: teacher predictions can exhibit high entropy and hallucinations, causing standard KD to degrade well-calibrated student priors. We propose CaRE-KD, a confidence-gated distillation framework that replaces static objectives with uncertainty-adaptive optimization. CaRE-KD has two components: a token-level loss (CaRE-Divergence) that adaptively switches between Forward and Reverse KL divergence based on teacher--student confidence, and a batch-level epistemic rejection mechanism (Revival) that suppresses updates when the teacher is more uncertain than the student. We provide a gradient-level analysis showing how this dual-granularity design induces a conditional calibration mechanism that prior static divergences cannot reproduce. Empirically, across eight teacher--student pairs and eleven benchmarks spanning instruction following, chat alignment, code generation, and mathematical reasoning, CaRE-KD delivers consistent gains over strong baselines (Skewed-KL, $\alpha$--$\beta$ divergence). Highlights include up to $+3.2$ average ROUGE-L on instruction-following tasks, $+2.1$ pass@1 on MBPP, $+1.7$ accuracy on GSM8k, and $+1.8$ accuracy on CollegeMath over the strongest baseline, with consistent gains in LLM-as-a-judge factuality (up to $+2.5$ per task over Skewed-RKL). Revival further acts as a principled, loss-agnostic plug-in that systematically strengthens existing distillation objectives by filtering epistemically unreliable teacher supervision.

Chinese Semantic Error Correction (CSEC) targets semantic errors in Chinese text, which are typically more subtle and complex than spelling and grammatical errors but remain relatively underexplored. Existing LLM-based approaches face two recurring obstacles in this task: over-correction, and unclear interaction between Chain-of-Thought (CoT) reasoning and self-consistency decoding, such that the benefits brought by CoT cannot be reliably transferred to final corrections. We propose Vote-guided Advantage Allocation for CSEC (VAA-CSEC), a multi-stage framework that combines CoT distillation, Supervised Fine-Tuning (SFT), Reinforcement Learning (RL) and self-consistency decoding. During RL, we design a task-specific reward function that directly aligned with the minimal-editing principle of CSEC. We further introduce Group-Level Relative Policy Optimization (GLPO), which reallocates GRPO advantages according to the margin between individual rollout rewards and the vote-aggregated group reward, aligning the RL training objective with the self-consistency objective used at inference time. Experiments on CSED-C and NaSGEC-Exam show that VAA-CSEC outperforms all LLM-based baselines on CSED-C with an F0.5 of 47.72%, achieves the highest recall of 42.15% among all methods, and establishes a new state of the art of 41.55% F0.5 on NaSGEC-Exam.

Generative content is increasingly entering the production and dissemination of news, transforming fake news from manually fabricated or simply manipulated material into complex forms in which native and generated content jointly participate. Existing multimodal fake news detection research primarily focuses on veracity assessment and rarely characterizes how generativity differences affect the reliability of evidence. In contrast, AIGC detection primarily determines whether content is generated or modified by generative models, but it does not by itself establish whether the underlying news event is true. To bridge the separation between these tasks in data and evaluation, we construct Weibo26, a multimodal fake news detection dataset for generative-content scenarios. On this basis, we propose the Generativity-Aware Hierarchical Reasoning (GAHR) framework, which combines global judgment with local correction so that generativity information participates in news-veracity reasoning. Experiments on multiple existing fake news detection benchmarks and Weibo26 show that GAHR achieves competitive veracity-detection performance while effectively identifying generative content.

Detailed image captioning requires accurate and comprehensive descriptions of fine-grained visual content, yet caption quality spans factual accuracy, information coverage, and clarity. Compared with conventional methods that rely mainly on high-quality supervision or holistic rewards, rubric-based reinforcement learning decomposes these requirements into explicit criteria and provides targeted, structured feedback. However, existing methods often use separate models for caption generation, rubric construction, and judging, which may lead to inconsistent interpretations across roles. Some dynamic rubric methods alternate updates between the caption policy and rubric generator while keeping the judge fixed, but staged optimization may still leave rubric construction and judging out of step with policy optimization. We propose MoCo Rubric, a two-stage framework that coordinates these roles. First, role-conditioned, shared-parameter multi-task supervised fine-tuning equips a single vision--language model to serve as the Caption Policy, Rubric Generator, and Rubric Judge. Then, the Generator constructs rubrics online from captions sampled by the current Policy, reference captions, and image evidence. The Judge provides rubric-based rewards, and only the Policy receives GRPO updates. As Policy updates change the candidates being evaluated, we use an exponential moving average of the Policy parameters to update one momentum model shared by the Generator and Judge. This gradual transfer lets both rubric roles track Policy updates without separate RL optimization while smoothing parameter changes that could disrupt their rubric capabilities under direct synchronization. Across five captioning benchmarks, MoCo Rubric achieves an average pairwise win rate of 72.83\%, the best mean rank in blind ranking, and the highest average score in caption-based question answering.

Existing multimodal fake news detection methods often introduce external information to assist detection. However, most of them rely on entity-level retrieval and are therefore prone to introducing event-irrelevant noise. Meanwhile, existing methods mainly focus on improving overall performance and do not account for differences in the degree of harm posed by different instances of fake news. To address these limitations, we design an Event-Level Evidence Retrieval Framework (ELERF) and propose a Relation-Aware Evidence Graph Network (RAEGNet). ELERF retrieves external evidence based on the complete event semantics of a news item. RAEGNet constructs a directed graph that incorporates news-evidence stance relations and evidence-evidence interaction relations, and introduces a conditional-harm branch to jointly model authenticity and potential harm. Experimental results demonstrate that RAEGNet outperforms multiple baseline methods across all evaluated metrics on Weibo-21, Fakeddit, and our self-constructed SSS dataset.

End-to-end full-duplex speech models have brought open-source machine conversation closer to human-like interaction, yet existing systems remain limited in two intertwined dimensions: long-context robustness and multi-party interaction. Real-world scenarios such as meetings, group lessons, and social-robot reception require a single model to track, contextualize, and respond to multiple speakers over extended durations. Progress is constrained by both data and evaluation: open multi-party speech corpora remain small and are not designed for codec-frame-level full-duplex modeling, while existing long-audio benchmarks focus on passive listening and speech-to-speech benchmarks are mostly short and dyadic. We extend the Moshi paradigm jointly along the long-horizon and multi-party axes in English and Chinese. First, we release 57.6k hours of synthetic training data ($\href{https://huggingface.co/datasets/MultiTalk/MultiTalkPT}{MultiTalkPT}$ and $\href{https://huggingface.co/datasets/MultiTalk/MultiTalkFT}{MultiTalkFT}$) for long-form, multi-party, English-Chinese full-duplex dialogue, with controllable length, participant count, turn-taking, overlap, backchannels, interruptions, addressee shifts, and long-range coreference. Second, we introduce $\href{https://huggingface.co/datasets/MultiTalk/MultiTalkBench}{MultiTalkBench}$, built from real human recordings, for evaluating long-form, multi-party, bilingual full-duplex dialogue. Conversations average 32.6 minutes and include probes for long-range entity tracking, topic coherence, and addressee selection. Third, we train a bilingual Moshi-style model that sustains coherent multi-party English-Chinese conversations over extended durations and substantially outperforms open-source baselines including Moshi, MiniCPM-o-4.5, and Qwen3-Omni-30B-A3B-Instruct on MultiTalkBench.

Transcribing domain-specific entities and rare proper nouns remains a major challenge in automatic speech recognition (ASR). In this paper, we propose BaLEEN (Biasing with Latent Encoded Entities), a lightweight, hypernetwork-based framework for dynamic contextual adaptation without fine-tuning the underlying ASR model. BaLEEN encodes variable-length contextual keywords using a pretrained language model, compresses them into a fixed sequence of latent vectors via a Perceiver bottleneck, and injects context-dependent bias vectors directly into the intermediate encoder representations of the ASR model. Because both the language model and the backbone ASR model remain entirely frozen during training, BaLEEN operates as a plug-and-play adapter that incurs zero computational overhead at inference time when context biases are precomputed. We evaluate our method on a CTC-based ASR model using a Wikipedia-derived corpus with annotated named entities and synthetic speech. Experimental results demonstrate that BaLEEN reduces keyword miss rate by 8.7% on the test set relative to the unbiased baseline while simultaneously improving overall word error rate by 21% and character error rate by 28%.

Post-training has been shown to significantly improve language models' performance on tasks with verifiable outcomes, including mathematical reasoning, software engineering, and computer use. However, whether the same approach can improve forecasting in financial markets is much less clear. Compared with tasks with verifiable outcomes, not only are realized returns noisy, but even what constitutes a relevant information set for making effective predictions is not obvious a priori: the model must decide which observations to gather and then commit to a numerical judgment before the outcome is known. We study this question in a chronological stock-price sandbox, where a language model gathers price, volume, relative-performance, and market-context evidence and predicts a future return. We post-train Qwen3-4B with supervised fine-tuning (SFT) on tool-use demonstrations, then proximal policy optimization (PPO) with a terminal reward given by the forecast score against the realized return. The resulting AURA-4B more than doubles the starting direction--magnitude score, from 20.94 to 43.31, and is comparable to frontier language models on this benchmark. Conditional magnitude agreement rises from 33.3 to 66.2, while directional accuracy changes from 62.9 to 65.4. SFT expands tool use, and PPO further increases the share of ranking and market-context queries. These results show that post-training can substantially improve financial forecasting performance, together with changes in how the model investigates the market, on this outcome-selected benchmark.

Comprehensive evaluations of automatic speech recognition (ASR) for Iberian languages remain limited, and low-resource languages, biases, and efficiency trade-offs are underexplored. We benchmark eleven systems, ten open-weight models and one commercial API, across five Iberian languages (Basque, Catalan, Galician, Portuguese, Spanish), with German and Turkish as controls. Evaluation uses an 85-hour dataset covering read speech, broadcast media, and audiobooks, assessing accuracy and efficiency via word error rate (WER) and real-time factors (RTF/RTFx). Results show no single model dominates: accuracy, efficiency, and language coverage present clear trade-offs. Low-resource languages, especially Basque, degrade significantly, highlighting the role of training coverage. We observe consistent sex disparities across most systems, highlighting fairness challenges in multilingual ASR. Overall, the benchmark provides practical guidance for real-world model selection.

Reproducibility is essential for scientific research, yet prior work shows that LLM outputs vary with hardware and batching. We identify an overlooked factor: the hidden injection of the current date into system prompts, which users cannot control and which changes every day. Across 9 recent LLMs and 6 datasets spanning multiple-choice QA (MCQA), math reasoning, code generation, and machine translation, performance varies solely with the current date, with deltas of up to 6% on MCQA, 14% on math reasoning, 7% on code generation, and 2.84 BLEU on machine translation. Model rankings also shift, affecting leaderboards. This date effect exceeds other sources of non-determinism, such as batch size and numerical precision. Standard prompting techniques -- chain-of-thought and few-shot prompting -- do not reduce the sensitivity; chain-of-thought even amplifies it. Our findings underscore the need for careful evaluation protocols to ensure reproducibility and fair comparisons in LLM research.

Large language models (LLMs) excel at token-level generation but may learn undesirable abstract semantics and lack comprehensive perception. LLM-JEPA mitigates this by aligning different views of the same underlying knowledge via a joint-embedding predictive architecture (JEPA). However, strong alignment does not necessarily lead to accurate, stable predictions. To address this, we propose ER-JEPA, which adds an episodic replay path to LLM-JEPA. ER-JEPA stores training pairs in a memory. At each step, it stores and retrieves relevant data to provide additional supervision. This enables learning from both the current batch and stored training pairs, providing additional supervision for token prediction and representation alignment. Experiments across multiple datasets (NL-RX, GSM8K, Spider, and NQ-Open) demonstrate that ER-JEPA consistently outperforms LLM-JEPA.

System One models such as Jev offer an efficient alternative to generative language models for tasks that require decisions rather than open-ended responses. However, existing Jev models exhibit limited Chinese-language decision accuracy, restricting their utility in both general and specialized settings. In this paper, we introduce Chinese-Jev, a System One model that addresses this gap through a unified data processing and training pipeline. Our data processing protocol converts heterogeneous Chinese-language annotations into probability targets over candidate options, enabling a shared training formulation across domains and question formats. To enable efficient inference, Chinese-Jev adopts a lightweight encoder-only backbone for text encoding and learns to score candidate answers through decision-oriented training. To address the misalignment between the pre-training distribution and downstream Chinese-language scenarios, we first train the model on a general-purpose corpus of 10 million examples, then fine-tune it separately for the medical, legal, and financial domains. To evaluate decision accuracy and calibration in both general and domain-specific Chinese-language settings, we introduce Chinese-Jev Bench (CJ-Bench). After first-stage pre-training, Chinese-Jev exceeds the accuracy of the closed-source Jev model by 1.24% on general-domain tasks while achieving a 20.3x speedup. Subsequent domain-specific fine-tuning yields a 4.0% accuracy improvement over Jev in medicine and achieves 92% of Jev's average accuracy across specialized domains, with a 17x speedup and an average latency of only 15 ms per example. We further demonstrate on-device deployment of an INT8-quantized model on mobile devices, achieving an inference latency of approximately 1.0 second per decision. The project is available at https://gulucaptain.github.io/Chinese-Jev/.

Text-to-speech models loop, truncate and lose count on text that repeats a phrase many times. We show that repetition itself is what breaks them, not the length that comes with it. Every repeated sentence in our test set is paired with a control of matched sentence and word count in which no word ever repeats back-to-back. Six models from three architectures render the controls almost perfectly and fail the repeated twins: 94.3% against 18.2% exactly right at k >= 6. The gap survives greedy decoding, repetition-penalty sweeps, four independent speech recognisers and 420 analysis specifications without once reversing sign; a held-out fourth architecture lands within a point of its predicted gap, and one of two non-autoregressive baselines shows the same failure. Varying the period of the text shows the failure grows smoothly with periodicity, half of it surviving when no word is adjacent to itself.

Large language models (LLMs) have become the foundation of personalized assistants, but maintaining persistent user memory across long-term interactions remains challenging. Existing memory systems often focus on storage, retrieval, or consolidation, while memory writing remains less controlled: transient requests, duplicate statements, and outdated user states may enter memory and later be retrieved for personalization. In this paper, we present AMU: Admission and Memory Update for Personalized Conversations, an SLM-guided (Small language model guided) structured framework for writing-time memory control. AMU uses structured memory filtering to decide what should enter memory and SLM-guided storage management to determine whether an admitted record should be stored separately, discarded as a duplicate, or fused as an update. We evaluate AMU in a controlled memory writing and retrieval setting. Experimental results show that AMU maintains cleaner and more retrievable personalized memories.

Knowledge concept tagging aims to assign specific concept or topic labels to educational content, which is essential for both educators and learners in traditional and online teaching practices. Recent work has explored large language models (LLMs) for this task, achieving promising performance. However, LLMs still struggle to select the correct concept from a large-scale candidate set due to the high dimensionality of the decision space. In this paper, we propose a novel three-stage Select-Reason-Judge (SRJudge) framework, which empowers LLMs with selective reasoning capability for fine-grained knowledge concept tagging. Specifically, the Selector in Stage 1 first narrows the candidate concepts to a top-K shortlist by fine-tuning a small language model (SLM), e.g., BERT, since the top-$K$ predictions hit the correct concept in most cases, thereby reducing the decision space of correct candidates. Next, the Stage 2 Reasoner employs a lightweight LLM for refined reasoning over the shortlisted candidates. It further integrates an improved reinforcement learning strategy with a dynamic task-specific reward function and a pruning mechanism to better align with human reasoning preferences. Finally, a larger LLM acts as a judger that evaluates the overall rationality of the reasoning process and its explanations to determine the final output. In addition, we construct two high-quality datasets for further validation, i.e., the biology dataset S_Bio and the physics dataset S_Phy. Experimental results demonstrate that our method consistently outperforms state-of-the-art baselines across benchmark datasets, verifying its effectiveness and superiority. Resources are available at: https://github.com/Nicozwy/SRJudge.

Graph databases are increasingly queried through natural language, yet every existing benchmark evaluates isolated single-turn queries rather than the multi-turn sessions through which analysts actually work. We introduce CypherTurn, the first benchmark for conversational Text-to-Cypher evaluation, comprising 721 sessions and 5,927 turns across 7 knowledge graphs and 13 conversational phenomena. We evaluate 15 models under a guided oracle protocol and a fully autonomous agentic protocol, yielding four findings. First, the best model reaches only 64.7% execution accuracy, and session-level correctness remains below 5%. Second, despite strong overall rank correlation, frontier models exhibit a consequential reordering of the top of the leaderboard under autonomous operation, a phenomenon we term the Autonomy Divergence, which reveals error-management as a partially independent capability from raw generation skill. Third, scaling action budgets from x3 to x10 fails to close the autonomy gap, as the strongest frontier models self-limit to approximately two actions per turn regardless of available budget. Fourth, single-turn Cypher fine-tuning degrades multi-turn instruction following, while architecture-appropriate specialization outperforms several frontier models. These results establish CypherTurn as an open challenge for conversational graph database reasoning. Code and data are available at https://github.com/BarryQ/CypherTurn.

LLM-based multi-agent systems (MAS) increasingly use latent collaboration to avoid the information loss and repeated encoding-decoding overhead of natural-language communication. However, directly forwarding all sender latents makes the receiver-side context scale with both the number of agents and the reasoning length, increasing computation, memory usage, and collaboration latency. A natural solution is latent compression. But we find that cross-agent redundancy remains unresolved in existing latent compression approaches, which typically compress each sender independently and then concatenate the results. We propose LatCom, a cross-agent latent compression framework for efficient multi-agent latent collaboration. LatCom maps multiple sender latents into a fixed number of receiver-readable and task-relevant slots. Rather than reconstructing all sender hidden states, it optimizes the compressed latents for receiver-side task utility. LatCom trains the compressor in two stages: single-sender readability learning first establishes a latent interface interpretable by the frozen receiver, and multi-sender fusion learning then trains the compressor to fuse complementary evidence and remove redundancy across agents. Experiments on multiple benchmarks with Qwen3-4B show that LatCom achieves an average 2.46x inference speed-up over LatentMAS and reduces output token usage by 70.3% while maintaining comparable average accuracy.

Natural language autoencoders translate a language model's internal activations into readable explanations. Explaining every token position is costly. Which positions should an auditor inspect to understand a potential threat? We study this question across $4.7$ million explanations on prompt injection and concealment. We compare signals from model computation with a ranker trained only on chat structure. Chat structure usually selects more relevant explanations than the computational signals, without requiring a model forward pass for position selection. On three of four datasets, explaining just $5\%$ of positions retains nearly all of the success rate from explaining every position, where success means obtaining an explanation about the threat. The benefit varies with the audit task. We also show that pretrained verbalizers recover words that models have learned to conceal through fine-tuning, without additional verbalizer training. These results identify where auditors can concentrate explanation generation and show that useful explanations can extend beyond the model a verbalizer was trained to describe.

Explicit intermediate reasoning gives large language models (LLMs) a stronger problem-solving mode. We study learning from this think-mode advantage via on-policy distillation (OPD). OPD preserves student-generated trajectories and provides dense token-level teacher targets at student-visited prefixes. Privileged reasoning is used during distillation rather than student inference. Uniform ThinkOPD, a natural think-enabled OPD baseline, conditions a fixed teacher on one shared think trace and uniformly distills every sibling student response. Although its prefixes are on-policy, the trace need not follow a route compatible with every complete response: the same privileged trace can induce different teacher-student discrepancies even when responses reach the same outcome. We summarize this interaction with trace-response divergence (TRD) and introduce ThinkOPD, which routes supervision at the response level by combining group-relative reward gain with a TRD-based compatibility proxy. Final response weights are normalized within each rollout group. Across mathematical reasoning and code generation, ThinkOPD outperforms Uniform ThinkOPD in both same-model settings and both cross-model teacher-student pairs, and it exceeds representative rationale and self-distillation baselines in a controlled comparison. Controlled interventions show that outcome benefit and the TRD-based proxy provide complementary routing signals in this setting. Think-enabled OPD provides a controlled setting for studying how teacher advantage becomes transferable along student responses.

Long-horizon information-seeking agents often accumulate noisy or misleading context, causing early mistakes to persist and making recovery increasingly difficult. We introduce an autonomous search harness in which the agent manages its own search process through three states: Rubric, Answer, and Verify. The agent first defines criteria for a valid answer, searches under these criteria, and then independently verifies the result before deciding whether to terminate or continue searching. It is further equipped with a Seal Memory tool that enables active context management. Training this behavior with reinforcement learning, however, can induce Seal Collapse, resulting in unstable training and preventing the agent from reliably learning when and how to use its memory tools. We solve this with a simple strategy that trains only the final segment after context management. Our 35B model achieves 72.83 on BrowseComp, outperforming comparable open-source systems, and consistently improves over the base model across BrowseComp-ZH, xbench, DeepSearchQA, WideSearch, financial investigation, and product search. Ablations show that autonomous compression outperforms automatic compaction and validate our RL design.

Open-source reasoning models provide unequal access to reasoning capability across languages. When a model can solve a problem but cannot deliver a complete solution in the user's language, language becomes an access barrier rather than merely a source of performance variation. We audit Qwen3 checkpoints on competition-mathematics tasks in eleven languages. Across the ten non-English languages, only 15.4-17.9% of problems receive a correct, terminating solution with visible reasoning in the requested language in any of 16 samples, compared with 92.9% in English. To identify the source of this disparity, we evaluate thirteen endpoints from one model family, spanning released checkpoints, multilingual supervised fine-tuning (SFT) at two scales, controlled SFT ablations, and three reinforcement-learning (RL) reward formulations. We jointly track correctness, language adherence, termination, and delivery efficiency. The dominant bottleneck shifts across post-training stages. Released models often reason in English. Multilingual SFT restores target-language reasoning, but accuracy declines across multilingual, English-only, and single-language SFT runs, showing that this cost is not specific to multilingual mixing; non-English reasoning traces additionally become prone to non-terminating loops. RL restores termination in both arms at no cost in accuracy, but only the arm whose reward includes a language term delivers: rewarding correctness alone returns the model to English. Together, these stages establish a constructive post-training path from English-pivoted capability to multilingual reasoning that is reliably delivered.

Cross-linguistic effects are a central topic in bilingual first-language acquisition. Artificial learners can help investigate L1-L2 interactions by enabling controlled comparisons across language combinations and learning conditions. Recent work explores this direction by training bilingual language models under developmentally plausible constraints. However, human and model learners still diverge in fundamental ways, with one major difference being input modality: children learn primarily from spoken input, whereas language models are typically trained on orthographic text. To reduce this gap, researchers have trained models on phonemic representations of speech. In this work, we combine these research directions to train bilingual BabyLMs with phonemic input. We keep English fixed as the L2 and vary the L1 across German, Swedish, Persian, and Basque, selected to represent contrasting combinations of syntactic and phoneme-inventory distance from English. Our results show stronger L1-related variation in grammatical learning trajectories under phonemic than orthographic input, while early lexical differences align with phoneme-inventory similarity.

Establishing unbranding as a critical practice to prevent visual logos from acquiring negative connotations is standard in image generation. Large Language Models (LLMs) now face a parallel and emerging challenge. These models frequently generate brand descriptions within diverse contexts. This frequency introduces significant risks, such as trademark dilution, false attribution, and brand defamation. In response, we formally define the novel task of LLM Unbranding. We specifically address the complex challenge of managing trade dress within textual outputs. This involves neutralizing characteristic language, slogans, and stylistic markers that define brand identity. Crucially, these elements are less evident than explicit visual logos. To benchmark this task, we introduce a comprehensive evaluation dataset incorporating prominent brands from multiple commercial domains. We rigorously evaluate existing state-of-the-art machine unlearning models using this benchmark. This evaluation identifies their limitations in selective textual unbranding. Finally, we propose MUTE, a novel inference-time method that effectively neutralizes textual trade dress while preserving the LLM's general capabilities and utility. By leveraging an iterative refinement loop, MUTE systematically optimizes system instructions to safely eliminate brand leakage without requiring fragile parameter updates. Code and dataset: The evaluation dataset and code for LLM Unbranding are available at https://github.com/KajetanOzog/LLM_unbranding. The implementation of MUTE is available at https://github.com/KajetanOzog/MUTE.

Retrieval-Augmented Generation has established itself as a fundamental framework in natural language processing, seamlessly integrating information retrieval with the generative capabilities of large language models. However, this process is fundamentally constrained by a critical challenge: semantic space mismatch between queries and retrieved contexts. We propose Knowledge-Aware Semantic Bridging (KASB), a novel framework that improves passage selection quality through semantic space alignment between queries and retrieved documents through intelligent knowledge fusion. Our approach leverages the complementary strengths of generative and retrieval-based knowledge through a multistage process that enhances both relevance and accuracy. We evaluate KASB on three popular open-domain Question Answering datasets to demonstrate the effectiveness of our approach.

Large language models (LLMs) have provided a unified interface for end-to-end combinatorial optimization (CO), but textual serialization alone may obscure spatial and relational structures that are important for generating effective CO solutions. This paper presents a general-purpose vision-language solver that augments textual instance descriptions with input-derived visual representations. A single vision-language model (VLM) is applied across different CO tasks and trained using supervised fine-tuning followed by verifier-guided reinforcement learning. While the visual inputs contain no gold solutions or solution-derived information, our experiments show that the VLM generally improves solution quality over its text-only counterpart, with particularly clear gains on more complex CO problems such as CVRP and JSSP. The advantage of visual information is more pronounced at large problem scales.

Credit-risk prediction is important in banking, but a prediction alone does not explain why an applicant is risky or how it should be combined with other evidence. This paper presents CredWise, a decision-support framework that integrates credit-risk prediction, probability calibration, explainable artificial intelligence, policy retrieval, SQL analytics, and controlled agent-based workflows. An XGBoost model is trained on Lending Club data (1,345,310 loans, 18 features) using a temporal split: 2007--2016 for training, 2017 for validation, and 2018 for testing. On the 2018 test set, the calibrated model achieved a ROC-AUC of 0.7109, PR-AUC of 0.2993, F1-score of 0.3714, and accuracy of 65.44\%. Calibration reduced the Brier score from 0.2157 to 0.1273 and the expected calibration error from 0.2862 to 0.0585. SHAP explanations were temporally stable, with a Spearman correlation of 0.9959 between 2017 and 2018 feature rankings. On 28 labeled queries covering nine policy sections, FAISS achieved the best Hit@1 (0.929) and MRR (0.964), while all three retrieval methods reached Hit@5 = 1.0. Agent routing achieved 95.6\% accuracy (43 of 45 cases), and the SQL benchmark scored 1.0 on exact-match, execution-success, and result-match across six cases. These results show that CredWise can combine predictions, explanations, policy evidence, and structured analytics in one controlled workflow. It is an academic research prototype, and final decisions remain with a human reviewer.

Answering questions and completing tasks over large document collections often requires connecting evidence spread across multiple documents, such as a project's approval recorded in one, its requirements in another, and its latest status in a third. Recent LLM agents approach this by iteratively searching the full corpus rather than reading only a fixed set of top-ranked documents. However, when the corpus is exposed only as a flat collection of files, a relevant document gives no indication of how it relates to others, so the agent must rediscover these relationships for every query, often missing complementary evidence while simultaneously consuming substantial additional tokens. To address this, we introduce CorpusMap, a navigation layer that organizes the corpus around its recurring entities, which are identifiable from the documents themselves and can link a single document to many others across sources. Specifically, CorpusMap represents each recurring entity as an Entity Page that aggregates information about it and links to every document that refers to it, forming a graph between entities and documents that the agent can traverse to gather otherwise disconnected evidence. Moreover, since CorpusMap is constructed offline by resolving mentions of the same entity across documents, its links are shared across queries rather than rediscovered repeatedly at inference time. Using 7 different models with 3 benchmark datasets, we show that CorpusMap improves both evidence discovery and answer quality over raw-corpus agentic search while using fewer tokens on average, and further outperforms 4 alternative navigation layers, suggesting that entities serve as effective anchors for navigating large document collections.

Operations research supports decision-making in domains such as energy, economics, and healthcare. Solving operations research problems typically begins with optimization modeling, which translates a natural-language problem description into executable solver code. LLMs offer a promising way to automate this process, but they remain prone to errors. In practice, these errors can be divided into two categories: syntactic errors refer to solver code that fails to run successfully or is judged infeasible by the solver; semantic errors refer to solver code that successfully returns an objective value but violates the intent of the original problem. Since semantic errors do not trigger runtime failures, they are difficult to detect and rectify. To address this problem, we introduce SemOPT, a semantic-guided framework for correcting LLM-based optimization models. SemOPT combines a semantic reward model that distinguishes faithful math models from plausible but incorrect ones with an adaptive correction system that applies hierarchical reward-guided search over the modeling space. Experiments on seven optimization modeling benchmarks show that SemOPT establishes a new state of the art and achieves an average 7.6% accuracy improvement over the strongest baseline on complex datasets.

Model adaptation is typically governed by a fixed recipe, even though different update programs can produce substantially different behavioral outcomes. We introduce adaptation compilation, which reframes where, how, and to what extent a model should adapt as a joint prediction and decision problem. Rather than searching over candidate programs anew for each learning episode, a compiler learns from prior adaptations to predict a vector-valued counterfactual response surface over candidate programs---their expected effects on acquisition, transfer, boundedness, and preservation---and selects a program before adaptation begins. Because this predicted geometry captures multiple behavioral consequences rather than a single winner or scalar score, it can be reused under different downstream priorities without retraining. Across five learning types, preferred programs vary meaningfully across episodes, and this variation is predictable from pre-adaptation information. On Llama-3.1-8B, compiler-selected programs approach exhaustive search while outperforming global and objective-specific defaults. Replication on Gemma-2-9B preserves program heterogeneity and selection headroom, but shows that exploiting this headroom requires accounting for uncertainty when departing from strong defaults. Together, these results show that adaptation search can be amortized across related learning problems, turning prior adaptation experience into a basis for deciding how future learning should occur.

Narrative extraction allows us to identify online hate narratives, supporting the construction of rigorous detection systems. Existing computational approaches, however, are limited in precision as they rely on semantic representations, which tend to capture only surface-level meaning. To detect more precise and interpretable narratives, we present an extraction pipeline that represents narratives as entity-evaluation pairs. Narratives are extracted using a Large Language Model (LLM) reasoning process that extends Aspect-Based Sentiment Analysis, identifying the aspect, classifying its judgement type as the basis for evaluation, and deriving the evaluation accordingly. Extracted narratives are then clustered using Leiden, following which clusters are resolved to an intended level of granularity through an LLM-guided refinement process. We illustrate this narrative pipeline with English Reddit comments from 2024 that criticize Taylor Swift, analyzing a representative cluster that exhibits hate speech patterns to demonstrate its interpretive value.

Long-term memory enables language models to use past interactions in future conversations. However, evidence needed to answer a question may be scattered across distant turns, while the question itself omits clues needed to locate it. Retrieved facts can reveal these clues, motivating retrieval decisions conditioned on evidence already found. We introduce MERA (Missing-Evidence Retrieval Augmentation), which separates globally searchable memory from a question-specific evidence state. Verified evidence guides subsequent retrieval without restricting access to the global memory. We train a lightweight planner through reinforcement learning, rewarding queries that recover previously missing evidence. MERA achieves strong answer accuracy across Qwen3-30B and GPT-4o-mini backbones. With Qwen3-30B for evidence processing and answer generation, the trained 0.6B planner achieves 77.40% accuracy on LoCoMo and 71.29% on LongMemEval-S, exceeding a 30B planner without retrieval-grounded training by 4.10% and 3.96%, respectively. On LoCoMo, later retrieval rounds increase cumulative evidence recall from 55.5% to 80.5%.

Retrieval-augmented generation (RAG) grounds large language models in external sources, but retrieved passages often name the right entities without providing the facts needed to answer. Even when instructed to abstain, 12 generators answer 40.0-99.3% of insufficient-evidence questions. Training generators to abstain ties the decision to model weights, may reward answers recalled from parametric knowledge, and still requires a full generator call. Can sufficiency be judged from the question and evidence alone, before any answer exists? We identify pitfalls in constructing insufficient-evidence tests: removing relevant evidence or pairing evidence with unrelated questions can reveal labels through lexical overlap or evidence position. We build a paired benchmark using substitution, deletion, and question-swap constructions that vary answer support while controlling selected surface features, such as word use. Sufficiency can be judged without generating an answer, but no single signal works across all datasets. We introduce RINSE (Relevance Is Not Sufficient Evidence), which combines three signals: whether every part of the question is covered, whether any passage offers an answer, and whether a small language model reading the passages together judges them sufficient. Across six datasets, RINSE ranks sufficient above insufficient evidence with a score of 0.837 (chance 0.5), exceeding the best of 10 prior methods (0.746) and a frontier model queried through an API (0.784). Its weakest dataset scores higher than any other method's weakest (0.684 vs. 0.676). RINSE runs locally before generation, taking 36.5 ms per question on a single GPU.

Retrieval-augmented language models are expected to answer from the retrieved evidence, but in practice they often keep answering when that evidence is missing. We trace this behavior to the training signal: answer-focused fine-tuning assigns no target to unsupported contexts, so it cannot distinguish a reader that abstains from one that guesses, and unsupported answering stays near 100% even as supported accuracy improves. We introduce Regime Boundary Alignment (RBA), which trains a single reader on matched variants of the same question and gold answer. The reader is trained to produce the gold answer when the context supports it, including when conflicting evidence is also present, and to abstain when the correct support is removed; inference is ordinary decoding, with no verifier, threshold, or regime label. On three multi-hop QA datasets across three seeds, RBA reduces the unsupported-answer rate by more than sixty percentage points relative to conflict-focused training while matching its supported accuracy. On a held-out TriviaQA retrieval-miss slice, the same reader reduces unsupported answering from 100% to below 1% while also improving supported accuracy. These results indicate that evidence-gated answering must be learned on both sides of the support boundary.

Multiple-choice benchmarks are cheap to grade and are running out of room, and the standard remedy, writing harder items, is slow and repeated for every benchmark. A saturated benchmark still holds a harder task. Each question's wrong options are written for that question alone, so a model can score by eliminating a few options. We propose AnswerPool: take $N$ questions that share a context, pool all their options into one list, and ask the model to assign every question its answer. No item is written and no label changes. The chance of guessing a group right falls from $10^{-3}$ to $5\times10^{-7}$ for five four-option questions, and a model that recognizes its answers keeps its multiple-choice score, so the accuracy lost to pooling measures the credit the format gave for elimination. Deleting answers from the pool makes questions unanswerable with exact ground truth, so abstention is scored in the same pass. Across eight text, image, and video benchmarks and eighteen models, pooling is harder for every model, the elimination credit is largest for the weakest models, and seven of eight open-weight models answer 87 to 100% of unanswerable questions.

Modern transformer models excel at capturing semantic relationships through sentence embeddings, yet their ability to perform pragmatic reasoning remains understudied. This paper investigates whether encoder-based transformers such as DeBERTa employ the M-Heuristic (the neo-Gricean principle that marked linguistic forms implicate marked meanings). We test this hypothesis by contrasting lexical causatives (e.g., ``Larry stopped the car'') with periphrastic causatives (e.g., ``Larry caused the car to stop'') using a Natural Language Inference framework. Our experiments across 188 conditions with 15 ambitransitive verbs reveal that DeBERTa, RoBERTa, and BART show no evidence of capturing the pragmatic distinction between these forms, with DeBERTa predicting ``Neutral'' for 100% of cases. Probing analysis initially suggested a representation-use dissociation, but control experiments reveal the probe was tracking syntactic complexity, not causative pragmatics. Semantic similarity over 30 triplets places periphrastic causatives closer to unmediated manner descriptions in 29/30 cases, opposite to M-Heuristic predictions in the embedding space. Under explicit metalinguistic framing, Gemini Flash-Lite reaches 100% with item-specific traces, so the principle is available under instruction yet unused in default NLI.

Wikipedia aims to provide a unified, neutral record of history, yet its independent language editions often function as distinct epistemic communities, creating divergent narratives around contested events. This paper investigates cross-lingual historiographical bias by analyzing Wikipedia articles across five languages (Romanian, Hungarian, Russian, Turkish, and English) focusing on three contentious events in Romanian history: the Battle of Posada (1330), the Soviet occupation of Bessarabia (1940), and the Night Attack at Targoviste (1462). Using human annotators and Large Language Models (LLMs) to classify citation stance and quantify narrative evolution from 2005 to 2024, we identify a phenomenon of "citation isolation". In the case of the Battle of Posada, only 2 out of 119 citations were shared between language editions, with the Romanian edition exhibiting a 91% pro-national bias compared to the balanced Hungarian edition. Longitudinal analysis reveals that these narratives are volatile and responsive to contemporary geopolitics, evidenced by a significant shift in the Russian framing of Bessarabia in 2024. Finally, we propose a "Peace-Maker" pipeline to automate conflict reconciliation. We demonstrate that while standard prompting leads models to hallucinate consensus, "adversarial" prompting, which explicitly instructs the model to preserve and attribute disagreement, achieves near-perfect neutrality scores.

Historical archives are an under-used source for extending the instrumental climate record backward in time, and LLMs offer a way to extract the indices climatologists derive by hand. Beyond measuring how well systems extract this signal, we check whether their errors are safe to use for cross-century comparison, since a good correlation score does not rule out systematic, era-linked bias. Comparing lexical baselines, fine-tuned historical transformers, and LLM prompting on the Pfister temperature index across five centuries of German text, lexical methods beat every fine-tuned transformer we test, including one pretrained from scratch on historical German (r=-0.016). All six LLMs we test (Gemini 2.5 Flash, GPT-5-mini, DeepSeek v4 Flash, Claude Sonnet 4.6, Qwen3.7-Plus, Kimi-K2.6-Fast) show a warm bias that grows with calendar year, with the same sign in every model (slopes +0.13 to +0.34/century, p<0.01). The effect is modest in size (r-squared approx equal to 0.01 to 0.05) but consistent across six independently developed models. The best-correlated of the six, Gemini 2.5 Flash, matches the best lexical correlation (r=0.32) at double the error. An ablation stripping explicit dates and calendar-era markers from the quotes leaves this trend essentially unchanged, favoring an anachronistic present-day prior over the model correctly inferring the quote's era. Correlation alone is thus insufficient for vetting an LLM as a historical-climate-index oracle.

We propose RegLLM, a diagnostic harness for bounded autonomy in regulated agentic workflows. It instruments six trustworthiness signals: citation validity, source grounding, schema compliance, escalation correctness, constitutional alignment, and unsafe-action rate. Signals are distinguished by their source of supervision: programmatic verifiers, task-level escalation labels, or AI-judge scores. A deterministic runtime supervisor blocks ungrounded answers and forces escalation, logging interventions. The same domain constitution informs evaluation, training rewards, and serving guardrails. Task-level should-escalate labels make the act-versus-defer decision a measurable training signal. We demonstrate the harness at smoke scale. An offline reference run (n=12) lifts escalation recall from 0 to 0.67 and reduces unsafe-action rate from 0.33 to 0.08 when governance is enabled. Two single-GPU Qwen2.5-3B LoRA/DPO pilots (n=8, same seed and evaluation split) expose substantial variation: nominally identical RL-base configurations yield task success of 0.25 versus 0.12 and escalation recall of 1.0 versus 0.5. An answer-quality adapter changes recall from 1.0 to 0.5 in Run A, but from 0.5 to 1.0 in Run B. An escalation-aware variant produces no measurable change in Run B. These small pilots do not establish reliable adapter effects or production readiness. Their contribution is diagnostic: configuration variance can overwhelm apparent tuning effects on bounded-autonomy metrics, motivating larger evaluation sets and repeated runs.

On-policy distillation (OPD) trains a student on its own reasoning trajectories using feedback from a stronger teacher. Teacher interventions can improve these trajectories, but also change the distribution on which the student learns. Our controlled studies show that rollout quality alone is an incomplete criterion for allocating teacher guidance. Deeper intervention yields diminishing gains in rollout accuracy while increasing off-policy load. In a training probe with a restricted rollout horizon, peak student accuracy and performance retention favor different intervention strengths. The preferred intervention depth and placement also vary across benchmarks. These findings motivate MAESTRO, which uses local policy disagreement to jointly adapt when the teacher takes over and how long it generates. Its {policy disagreement score} combines teacher-weighted candidate coverage with local distribution similarity and is aggregated within reasoning paragraphs. Across eight mathematical reasoning benchmarks, MAESTRO achieves the highest macro-average accuracy among the compared methods for both 0.6B and 1.7B Qwen3 students, with the 1.7B student leading on every benchmark. MAESTRO also reduces average training response length by 67.3\% relative to standard OPD. The code is available at https://github.com/yhao-wang/MAESTRO.

Masked diffusion models (MDMs) generate sequences by progressively unmasking several tokens per denoising step, but their reverse process is typically factorized over positions, limiting sample quality in the few-step regime where diffusion's speed advantage over autoregressive decoding matters most. A recent line of work introduces a continuous Gaussian latent, trained as a variational autoencoder, to capture correlations across positions, but such approaches are prone to posterior collapse, where the latent is silently ignored. We propose Enhanced Mixture-of-Experts (E-MoE), which builds the reverse process as a mixture of factorized distributions over a discrete shared latent given by the expert-routing decisions of a Mixture-of-Experts (MoE) backbone, without increasing active parameters over the factorized baseline. Across synthetic multi-modal benchmarks, binarized MNIST, and LM1B, E-MoE improves few-step generation over factorized baselines.

Cross-lingual zero-shot transfer and multilingual fine-tuning are promising approaches for NLP tasks such as Named Entity Recognition (NER) in low-resource languages, but in the absence of target language benchmarks, it is unclear which auxiliary language selection strategy leads to the best transfer. We introduce RunyaNER, the first publicly available NER benchmark for the East African language Runyankore, and use it to investigate the choice of which languages to use for transfer. Created with a semi-automated pipeline and fully manually verified, RunyaNER contains over 237k annotated words across 30k sentences. We benchmark pretrained models on RunyaNER, establishing that our dataset is of sufficient quality and size to produce effective Runyankore NER models. We then use RunyaNER to investigate auxiliary language selection in cross-lingual zero-shot and multilingual fine-tuning settings. Our experiments show that while transfer performance is highly sensitive to auxiliary language selection, embedding-based measures computed from labelled training spans correlate more strongly with downstream transfer performance than traditional linguistic features based on metadata or typology. By releasing RunyaNER and providing a systematic analysis of auxiliary language selection strategies, this work contributes both a new benchmark resource and practical insights for multilingual transfer in low-resource settings.

Merging pretrained models has emerged as an effective approach for consolidating diverse capabilities into a single unified model. However, prevailing merging methods typically treat each task vector as an indivisible merging unit, overlooking the heterogeneous geometric changes encoded within it. This treatment can induce cross-component coupling: when merging decisions are derived from statistics of the complete task vector, the geometric characteristics of one component may influence how another is selected, weighted, or combined, potentially degrading the quality of the merged model. To address this issue, we propose DiGA, a Disentangled Geometry-Aware model merging framework. Using the pretrained weights as a shared geometric reference, DiGA orthogonally decomposes each task vector into components corresponding to distinct geometric attributes. Rather than merging the task vectors as a whole, DiGA aggregates corresponding components independently within their respective subspaces and subsequently recombines them into a unified update. This component-wise formulation preserves the geometric identity of each component and prevents the characteristics of one component from interfering with the aggregation of another. Furthermore, DiGA can be incorporated into a broad range of existing model merging methods. Extensive experiments across diverse models, tasks, and merging methods demonstrate that DiGA improves merged-model performance and reduces capability degradation. Our repository is on https://github.com/wzj1718/DiGA.

Audio-visual large language models (AVLLMs) have made remarkable progress in multimodal understanding and reasoning through interactions among visual, auditory, and linguistic information. However, recent studies show that AVLLMs face a critical challenge: $\textbf{source-confused grounding hallucination}$, where cues from the unused modality induce responses that the required modality does not support, undermining reliability in real-world applications. Existing methods have made progress in mitigating this failure, yet how it arises from internal cross-modal interactions remains insufficiently understood. To address this gap, we conduct path-intervention and representation analyses, revealing a $\textbf{question-relay}$ mechanism: question states carry interfering cues alongside required-source evidence, undermining grounding in required-modality evidence. Cutting pathways from interfering modality to question states yields greater correct-answer logit recovery than cutting those to the generation position. Motivated by these findings, we propose $\textbf{SECRET}$ ($\textbf{S}$ourc$\textbf{E}$-$\textbf{C}$onditioned $\textbf{RE}$lay s$\textbf{T}$eering), a training-free method that mitigates cross-modal interference at the question relay. Using contrasting question representations elicited through different modality-pathway interventions, SECRET steers the original question states toward required-source evidence. Experiments on two widely adopted benchmarks CMM and AVHBench across three AVLLMs show that SECRET consistently outperforms prior training-free methods, substantially mitigating source-confused grounding hallucinations (e.g., up to +18.0 and +7.1 percentage points over base models). Modality-specific captioning further demonstrates its generalizability to open-ended generation.

Large Language Model judges are widely used to rank texts and text-generating systems through pairwise comparison, and their reliability is typically assessed via three proxies: position bias, transitivity, and pairwise agreement (self- or human-labeled). Because these proxies drive judge selection and benchmarking, a substantial literature reporting that judges perform poorly on them risks steering practitioners away from otherwise capable evaluators. We argue this assessment is misleading. Under the Bradley--Terry geometry underlying pairwise aggregation, each proxy is dominated by close-rank-gap pairs, where inconsistency is information-theoretically expected and individual verdicts contribute little to the aggregate ranking; far-gap pairs carry the ranking signal but barely move the proxies. We formalize this argument and validate it in a controlled simulation and on two human-rated corpora: the proxies correlate only weakly with ranking accuracy against gold, and their predictive component concentrates in the far-gap regime. Judges should therefore be assessed on rank-gap-conditional metrics, ideally against human rankings. Code at https://github.com/brunobrocai/PairDifficulty.

Fine-tuning a language model on a narrow set of harmful demonstrations, such as bad medical advice, can make it broadly misaligned on unrelated questions, a phenomenon known as emergent misalignment (EM). The usual defense is to find the offending rows and delete them, but a row locator failed our held-out test and deleting rows helps less than expected. We ask a different question: given a fixed set of poisoned rows, is it better to correct them than to remove them? We fine-tune Qwen2.5-14B-Instruct on a mixture of bad medical advice and benign chat data, select a quarter of the poison rows in advance, and either delete them or replace each with a corrected answer to the same prompt, keeping everything else the same. Replacing the rows cuts the EM rate by about a third and improves answers on held-out medical questions, while deleting the same rows has little measurable effect. The advantage is larger when half the poison rows are corrected, and it holds on a second base model and a second misaligned model organism. The content of the replacement appears to matter: paraphrasing the rows while keeping their bad advice shows no clear benefit, and the correct answers distributed with the dataset appear to do about as well as our rewriter's. Realigning an already-poisoned model with further fine-tuning is known to work, but which data does the work has not been compared directly. We find that a short round of training on corrections beats the same amount of training on generic chat data, that corrections on other medical prompts do roughly as well as corrections of the poisoned prompts themselves, and that instructing the correction writer to model a careful, harm-avoiding assistant adds no measurable benefit over plain corrections. In the settings we tested, correcting harmful training data reduces EM more than deleting it.

LLMs have been studied in recent linguistics as potential models of humans' linguistic abilities. Here we discuss an entirely different use of AI, namely as a co-scientist, to help construct and assess linguistic theories (we refer to the result as "Co-Linguistics"). Since the 1960s, linguistics has developed theories that are in principle mathematically formalizable, often in the language of formal language theory or model theory. The AI revolution in mathematics will thus have consequences in linguistics-but with an essential twist: proving new theorems is rarely the linguist's goal. Rather, one seeks to find the best set of axioms to derive empirical statements. AI could accelerate research by making existing theories fully explicit, by comparing competing theories, and more ambitiously, by proposing new theories (in machine learning, this relates to "program induction"). It will also help assess theories by accelerating the identification and test of crucial predictions, thanks to unparalleled access to data (in machine learning, this relates to "active learning"). While the cycle from theory evaluation to theory construction may give rise to recursive and possibly autonomous improvement of linguistic theories, humans remain central: linguists provide scientific directions and evaluate theories conceptually, and experimental participants are needed to assess empirical predictions that are outside the reach of LLMs.

Jev is a commercial System One model from TypeSafe AI that does not generate text: given a state and typed questions, it returns a choice from fixed options, a position on a rubric, or the probability that a statement is true, with probabilities the vendor describes as calibrated. Such models target small decisions in information access pipelines, such as routing queries, checking grounding, moderating content, or rating against a rubric. We evaluate Jev (jev-1.13.0) zero-shot on 37 datasets spanning classification, routing, natural language inference, reading comprehension, commonsense reasoning, moderation, legal clause analysis and rubric scoring, with one frozen template per dataset and full evaluation splits: 346,009 requests for under USD 10. For reference, we score Qwen3.8-27B and Gemma-4-E4B on identical requests via their exact next-token probabilities over the options. Jev reaches 95-99% accuracy on IMDB, SST-2, HellaSwag and ARC and 86.7% on Belebele across 122 languages. It beats Qwen on 27 of 37 datasets, with none of Qwen's nine leads outside the bootstrap intervals, and Gemma on all 37. All three models degrade on low-resource languages, fine-grained or noisy labels, and rubric-based quality judgments. Jev's choice probabilities are well calibrated and support selective prediction. Binary probabilities rank well but are poorly placed relative to a fixed 0.5 threshold; thresholds tuned on training data raise micro-F1 on UNFAIR-ToS from 0.50 to 0.75. Jev answers MMLU's calculation-heavy questions more accurately than other MMLU questions (94% vs. 91%), whereas both open models, and all three on C-Eval, find them harder. Rotating the options leaves Jev's accuracy unchanged and withholding the question drops it to near chance, ruling out shallow memorization but not memorized question-answer pairs. We release the code, harness and all raw responses.

Graph-based retrieval-augmented generation supports multi-hop retrieval by organizing corpus information into graphs. However, existing relation-free graph retrieval methods rely primarily on query-sentence similarity to search for evidence. This can exclude useful bridging evidence with low query similarity and activate incidental entities unrelated to the reasoning chain. In this paper, we propose a simple and effective approach called NexusRAG, which augments the relation-free Tri-Graph with a corpus-level entity neighborhood structure derived from joint entity co-occurrence and semantic similarity. NexusRAG employs this structure to guide two complementary propagation paths: neighborhood-constrained semantic propagation through sentences identifies the query-relevant entity frontier, while direct structural propagation between neighboring entities expands that frontier to structurally related entities. The propagated entity weights also inform neighborhood-aware passage initialization for Personalized PageRank. Experiments on three multi-hop QA benchmarks and a domain-specific subset of GraphRAG-Bench show that NexusRAG consistently outperforms existing approaches. On the GraphRAG-Bench subset, NexusRAG achieves the highest evidence recall in all question categories, exceeding baselines by 4.2-8.1 points. The implementation code is available at https://github.com/Jacob-biu/NexusRAG.

Reading behaviour varies not only with linguistic input, but also with reader proficiency. In this study, we investigate whether the layer-wise relationship between surprisal from large language models (LLMs) and human gaze behaviour differs across readers with different levels of proficiency and across gaze measures. Using eye-tracking data from the MECO L2 corpus, we compare readers with high and low vocabulary proficiency on first-pass gaze duration (FPGD) and total gaze duration (TGD). We quantify the distribution of the predictive power of surprisal across model layers using Predictive Depth. Across 12 tested LLMs, we find that readers with lower vocabulary proficiency tend to show deeper Predictive Depth for FPGD, while this difference is smaller for TGD. Also, TGD itself shows deeper Predictive Depth than FPGD in both proficiency groups. These patterns suggest that where predictive power is concentrated across LLM layers may be related to the timing and breadth of the reading processes captured by different gaze measures, and that this relationship can vary with reader proficiency. Our leave-one-out analysis further shows that the advantage of informative internal layers extends to unseen texts, although the practical improvements in prediction are limited. Overall, our results show that layer-wise LLM surprisal provides a useful perspective on variation in reading behaviour across both reader groups and gaze measures.

Existing product matching benchmarks primarily contain English-language product data and are often dominated by a single product category, such as electronics. This paper introduces Billiger.de Products, a bilingual German and English entity matching benchmark covering thirteen consumer product categories, including difficult-to-handle categories such as clothing and furniture. The benchmark data originates from the German price comparison platform billiger.de. Following the design of WDC Products, the benchmark offers multiple variants that differ in the fraction of corner cases, the size of the development set, and the fraction of entities unseen during training. An aligned English translation of every offer keeps all pairs, splits, and labels fixed, while cross-language test sets combine German and English records within individual pairs. We validate the benchmark using six supervised matchers and zero-shot GPT-5.2 on both language versions and the cross-language test sets. The validation shows the difficulty of the benchmark. The comparison of the results on the English version of the benchmark to the results on the German version shows that most matchers score on average higher on the English version. The difference is largest for RoBERTa and HierGAT, while the zero-shot LLM runs are largely insensitive to the language. Comparing the F1 scores achieved by PLM-based matchers on the English version of Billiger.de Products with their performance on existing English-language benchmarks, such as WDC Products and Abt-Buy, shows that Billiger.de Products is more difficult than these benchmarks.

Purpose: Most Greek papyri remain unpublished and undigitised; a handwritten text recognition (HTR) pipeline that transcribes them automatically would let scholars discover documents and literary works that have so far gone unread. Recognition systems for Ancient Greek papyri are in statu nascendi, and how accurate they must be for a given papyrological task has not been examined. To answer this and set a benchmark for Greek papyrus HTR, we test a range of character error rates (CER) against four papyrological tasks, using published editions as ground truth. Methods: From 63,846 current editions of Greek texts in papyri.info, we imitate a letters-only "perfect HTR" output by removing the editorial layer, then degrade it with a seeded algorithm to exact CERs of 1 - 50%, with lost lines and four error-shape variants. On these data we train small models (TF-IDF, fastText, a character CNN, ByT5-small) for document type, dating and documentary-versus-literary classification, and apply eight keyword search methods. We compare models trained on clean text with models retrained at a specific CER level, and evaluate across CERs. Results: Tolerance differs by task. With clean-trained models, documentary-versus-literary classification retains 90% of its metric up to 20% CER; document type up to 7.5%; subtypes and search up to 5%; dating only up to 3%. Retraining on text containing character errors largely eliminates the sharp degradation that otherwise sets in above 15% CER. Models generally tolerate concentrated damage in a long document better than small errors spread across a short text. Conclusion: The study provides a CER target for each of the four tasks and shows that models trained on noisy text make current, imperfect text recognition useful for them.

Large language models (LLMs) have shown strong potential as training-free text encoders for long-context embeddings. Existing approaches primarily improve information flow under causal attention and typically construct embeddings by uniformly averaging all token representations. However, for long documents, such mean pooling can dilute salient semantic information with abundant redundant or weakly informative content. To this end, we propose SCSP, a training-free framework that leverages semantic compression for informative token selection in long-context embedding. Specifically, SCSP first partitions a document into sentence-aware chunks and appends a semantic compression prompt to each chunk. A prompt-isolated attention mask preserves information flow among document tokens while restricting each prompt to its corresponding local context. We then use the attention patterns elicited by these prompts to estimate token importance, select informative tokens, and aggregate their intermediate-layer representations into the final embedding. Extensive experiments on long-context embedding benchmarks demonstrate that SCSP can be integrated into both zero-shot and fine-tuned models in a plug-and-play manner, consistently improving their performance.

We propose a rubric for assessing expressed clinical reasoning in model responses, drawing on three bodies of work: medical education assessment frameworks (ART, SCT, Key Feature Problems and OSCE); clinical LLM benchmarks (MedR-Bench, HealthBench, TIMER-Bench, DR.BENCH, PrIME-LLM and PatientSafeBench); and general LLM reasoning evaluation research, including the Factuality-Validity-Coherence-Utility taxonomy, FaithCoT-Bench and C2-Faith. We use groundedness as a clinically oriented adaptation of the taxonomy's factuality category. The rubric brings these concepts together in a multidimensional framework for scoring free-text responses to gold-standard clinical vignettes. It includes provisional behavioural anchors, applicability rules and a separate flag for case-specific safety-critical errors. General-domain frameworks inform its design but are not treated as validated clinical instruments. The rubric does not replace case-specific reference criteria or the task-specific metrics of existing benchmarks. It has not yet been tested for inter-rater reliability, construct validity or clinical utility. Its immediate purpose is to make evaluation decisions explicit and open to scrutiny before empirical testing.

Studying how fine-tuning shapes refusal and noncompliance behaviour requires identifying training examples that refuse, evade or otherwise fail to fulfil the requested task. But existing annotation covers evaluation sets of a few thousand prompts at most. We present CompOrca, a compliance labelling over the entirety of the 4,233,923-example OpenOrca corpus. Every example was classified as compliant or noncompliant by five independent passes of an open-weight LLM judge (LongCat-2.0, 1.6T parameters), and the corpus is released as unanimous compliance (94.75%), unanimous noncompliance (1.28%), and nonunanimous rows (3.97%) along with the raw vote counts. A single pass flags 2.7-3.2% of the corpus as noncompliant, while only 1.28% is flagged by all five, allowing for filtering the most ambiguous samples. Against 450 human-annotated examples, 150 of them annotated twice (human-human $\kappa = 0.93$), the unanimous compliance and noncompliance labels are 97.3% and 86.7% precise, the latter a high-precision subset, not a complete enumeration, of noncompliance. Published refusal-detection methods recall only between 0.4% and 94.1% of the noncompliance class. We release the full corpus with its per-row labels and vote counts at https://huggingface.co/datasets/cemiu/CompOrca

Empathetic spoken dialogue requires models to use both what is said and how it is said to decide how to respond. Explicit CoT can improve paralinguistic perception and make acoustic cues more explicit in replies, yet does not ensure their effective use in response planning. We call this mismatch the perception-reasoning gap. In addition, CoT may not fully capture acoustic cues in words, and generating it adds inference latency. To address these limitations, we introduce LoopSLM, which builds on looped Transformers for latent reasoning, reusing a decoder block to refine hidden states with acoustic grounding at every pass. Its two-stage training further narrows the perception-reasoning gap by separating learning to reason from learning to respond, enabling direct inference without CoT. On EchoMind, LoopSLM improves paralinguistic understanding, reasoning, and reply quality over Qwen2.5-Omni-7B. Against the CoT-SFT baseline, LoopSLM gains over 20 points in reasoning accuracy while generating 64.5% fewer tokens at half the latency. It also outperforms Qwen3-Omni-Thinking on most empathetic reply metrics with 34x lower latency. Despite training only on dialogue data, LoopSLM improves accuracy on general audio benchmarks.

Transformer language models build predictions through successive residual updates, but how their representations become specific to an eventual outcome remains unclear. We study this process by comparing intermediate residual states with their own final states and an empirical bank of final states from other contexts. Across six pretrained language models, the own endpoint becomes preferable to the average alternative early, while many individual endpoints remain closer. These competing sets generally shrink with depth, but their membership changes and their surviving endpoints need not become more similar to one another. Directional alignment and endpoint rank can therefore improve while Euclidean distance to the final state changes little. We develop a simple high-dimensional model that separates the roles of norm, alignment, and endpoint geometry, showing how gradual directional changes can produce sharp reductions in competition. We also prove that a straight path toward the own endpoint cannot introduce new competitors under either Euclidean or cosine distance; observed entries thus establish departures from straight-line convergence. Finally, endpoints associated with lower-ranked output tokens tend to lie farther away in cosine distance across all studied models, connecting residual geometry to output organization. Together, these findings characterize increasing geometric specificity during transformer inference and explain why distance, competitor count, and concentration of the surviving endpoints provide distinct views of that process.

Vision-Language Models (VLMs) remain vulnerable to cross-modal implicit risks: visual and textual inputs that appear benign in isolation can jointly elicit unsafe responses. Existing safety methods often require large preference datasets, costly multi-rollout training, or additional safeguards at inference time. They may also sacrifice helpfulness by directly refusing requests that could be answered safely. In this paper, we propose Intent-Privilege On-Policy Self-Distillation (OPSD), which leverages evidence-grounded intent as privileged supervision during training to help VLMs recognize implicit risks and provide safe, useful responses instead of blanket refusals. OPSD distills a teacher's intent-conditioned preferences over responses into a student using a single rollout per prompt; the student then responds without intent annotations or an additional safety module. With only 1,447 safety-specific examples - 95% fewer than standard preference datasets - OPSD reduces training time by 5x relative to multi-rollout GRPO-style training and average inference length by 7%. It attains the highest ratio for joint safety-helpfulness success, which measures the proportion of responses that are both safe and helpful, across all five evaluation groups. Remarkably, on pooled SIUO+HoliSafe, this success ratio rises from 43.9% to 53.5%. These results show that training-time intent supervision can improve both safety and helpfulness while substantially reducing data, training, and inference costs.

Large language models (LLMs) are increasingly deployed as social agents, yet credible human-like interaction requires more than fluent responses or persona consistency. Agents must autonomously decide whether, when, and how to communicate while adapting to evolving contexts, goals, and relationships. Existing research, however, lacks a unified approach to enabling, evaluating, and improving such capabilities in continuous, open-ended interaction. We introduce AnthroDial, a unified framework for developing anthropomorphic social agents from three complementary aspects: MindFlow, a lightweight interaction harness that enables autonomous, asynchronous, and adaptive communication through a dynamic Mind Buffer; CAPS-Eval, a theory-grounded framework for evaluating cognitive, affective, and behavioral dimensions of anthropomorphic interaction; and a scalable training paradigm that combines SEEDS for environment expansion with DiAPO for adaptive capability optimization. We further construct evaluation datasets covering everyday communication, game interaction, and long-horizon character interaction. Extensive experiments across diverse models and scenarios demonstrate improved interaction autonomy and naturalness, validate the reliability, discriminativeness, and agreement with human rankings of CAPS-Eval, and confirm the effectiveness of our training paradigm. Together, these components provide a unified framework for developing credible human-like social agents in open-ended interaction.

In the Global South, the lower-income countries of Africa, Asia, and Latin America where most of the world's languages are spoken, a deployed text classifier usually runs on ordinary CPUs, serves many languages with a single model, has few labeled examples in any of them, and relies on people to catch its mistakes. Such a system is only useful if it can promise how often it will be wrong: at most a fixed fraction of the labels it assigns on its own may be incorrect, and everything else must go to a person. Split conformal prediction delivers this promise through a single confidence threshold, normally estimated on validation data pooled across languages. We ask whether the promise reaches every language, and it does not. On MasakhaNEWS (16 African languages) and AfriSenti (12 languages plus two never seen in training), a pooled threshold meets the 90% target on average but covers Somali at 77.5%, Tigrinya at 83.7%, and the two unseen languages at 77.5% and 81.2%. Estimating one threshold per language brings every language to between 89.1% and 91.0% without retraining, and it shows how unequal the cost of the promise is: keeping it means sending 43% of Somali news and over 80% of Amharic and Xitsonga tweets to a person, against under 8% of Nigerian Pidgin news. One or two hundred labels per language are enough and the models train in minutes on one CPU core, so the fix is affordable: calibrate, report, and budget human review one language at a time.

Vision-language models (VLMs) are increasingly used in place of human annotators, making it important that substitutability tests reflect the model rather than incidental evaluation conditions. We introduce MIST, the Misleading-Image Stress Test: 200 English sentences, each built around a phrase readable either figuratively or literally and shown with an aligned image depicting its reading, a misleading image depicting the opposite, or no image at all. The guidelines require the label to be decided from the sentence alone, so no image should change any answer. We expected each image to pull a judge's labels toward the sense it depicts, and neither kind did. Across thirteen VLM judges, an aligned image changed 20.5% of labels and a misleading one 19.4%, close for every judge and both above the 11.6% produced by deleting the ignore-the-image instruction with the image left in place. Yet only 37% of the labels that differ between the two images moved toward the sense shown, and agreement with our human annotators is unchanged whether the image is absent, aligned or misleading. The effect is smaller in the seven judges that pass the alt-test than in the six that never do, but present in all of them: what moves a judge is that an image is there, not which of the two it is, so a substitutability verdict describes a configuration as much as a model.

How many tokens from its context does a language model actually use, and what determines that number? We study this question through self-attention. Without retraining, we retain only the tokens with the highest attention weights at each head, layer, and query, keeping their original weights unchanged. By varying the selected set size and measuring the increase in negative log-likelihood (NLL), we estimate the effective attention set size needed to stay within a chosen loss tolerance. Relatively small selected sets can keep NLL close to the full-attention baseline, although the required size varies across models. Attention-based selection substantially outperforms random selection. Selected sets exhibit geometric structure, although geometric separation alone does not establish that model loss is preserved. Extending context while evaluating the same prediction targets increases the required set size, while its fraction of context decreases over the tested range. Experiments with a fixed supporting fact show that additional background pushes its tokens down the attention ranking and reduces their attention mass. Renormalizing the retained weights can substantially reduce the required set size, showing that it also depends on how selected representations are combined. Conditional theoretical models explain how competition and attention-mass retention can produce growing set sizes without more distinct information to retrieve. These results provide a way to measure effective attention set size in language models and investigate its dependence on context, competition, and aggregation.

Every text classifier for an African language begins with a budgeting question: how many labelled examples are needed, and can labels from other African languages stand in for them? We answer both questions empirically for 28 language-task pairs, news topic classification in 16 languages (MasakhaNEWS) and tweet sentiment in 12 languages (AfriSenti), using a character n-gram linear model that trains in seconds on two CPU cores with no pretrained weights and no accelerator. Monolingual learning curves at budgets from 25 to several thousand labels show that topic classification reaches 90\% of its full-data macro-F1 with about 400 labels in the median language, while sentiment is still improving at the full training size in 11 of 12 languages and needs thousands of labels. Pooling the full training data of the other languages in the benchmark is worth a great deal at small budgets and nothing at large ones: at 25 target labels it adds 0.20 macro-F1 on average for news (up to 0.43 for Lingala) and 0.08 for sentiment, the gain decays to zero by 800 labels, and at full size pooling hurts in 9 of 16 and 8 of 12 languages. Twenty-five target labels plus pooled data match what 100 to 400 monolingual labels achieve for most news languages. A complete zero-shot transfer matrix shows that transfer without any target labels recovers a median of only 13\% (news) and 4\% (sentiment) of the gap between a majority-class predictor and the in-language model, with the exceptions explained by shared script (Amharic and Tigrinya), shared lexicon (English and Nigerian Pidgin, the Arabic dialects), or a shared label prior rather than by language family. We release code that regenerates every number from the public benchmark files and translate the results into concrete annotation guidance for teams building African-language classifiers without GPUs.

Pre-trained language models (PLMs) with encoder-based architectures have shown impressive capabilities in zero-shot cross-lingual transfer for various language understanding tasks. However, applying this technique to dependency parsing remains a significant challenge due to its syntactic nature. To boost model generalizability across linguistic typologies, we propose a cross-lingual unsupervised bootstrapping method to improve syntactic knowledge within the PLM. We show that our method achieves a significant improvement in zero-shot parsing performance in low-resource languages. Analysis of these bootstrapped models uncovers increased robustness in recognizing syntactic structures, evidenced by higher scores in parameter-free tree probing tests.

Current pre-training datasets are derived from web crawls, with all their issues, and were not designed to support mid- and post-training pipelines--for instance, they contain little explicit reasoning. Thus, many frontier labs have begun to develop their own internal datasets, starting from state-of-the-art models, to augment their pre-training data mix, eg, with reasoning traces to address cold-start problems. While demonstratively effective, none of these datasets are public, and the effect of this so-called synthetic data on knowledge and skill acquisition of language models, including small ones, remains poorly understood. We present SYNTH, the first open-source synthetic corpus derived from 58,698 Wikipedia articles that collapses pre-, mid-, and post-training into a single training stage via structured amplification of curated encyclopedic seeds. We evaluate SYNTH by training a suite of models: a 56M tiny model (Monad), 0.3B-0.6B dense models (Baguettotron), and a 13B / 1B-active MoE. At iso-compute, SYNTH outperforms filtered web data, and our models remain competitive with similarly-sized open-weight baselines. Because SYNTH is back-translated from grounded passages, SYNTH-trained models achieve high factual precision despite 10-140x fewer training tokens, with memorization targeted by the seed corpus. These results show that synthetic datasets, including our SYNTH dataset, are capable of producing competitive generalist models from a fraction of the training data, enabling rapid iteration as the frontier advances. These findings open up possibilities for both generalist models with significantly increased data efficiency, as well as domain-specific models where no instruction or conversational data is available. Finally, we publicly release our SYNTH dataset and the suite of Baguettotron models under a permissive license, thus supporting open-source language model development.

Fine-tuning large language models on narrow, misaligned tasks can undo their post-training alignment and induce novel misaligned behaviors -- a phenomenon known as \emph{emergent misalignment} (EM). EM has been linked to persona-like representations, where fine-tuning might reduce loss by amplifying a harmful or 'evil' persona. It remains unclear which properties of the training data drive this effect: whether all harmful examples contribute approximately equally to misalignment and whether different models are equally affected by the same fine-tuning examples. In this work, we use training data attribution to quantitatively estimate how much each harmful example contributes to EM. We benchmark the quality of the attribution via retraining -- a sound attribution score should enable us to enhance or attenuate EM by filtering data on that score. Score-based filtering can substantially enhance or attenuate EM; we find that both data-attribution scores and a black-box harmfulness score can identify consequential examples. All models we test become misaligned when trained on the same dataset, and influence scores perform best when filtering data from the same model that computed them. We find cross-model generalization of influence scores from scores derived from the three model families we tested, but this generalization does not recover same model filtering performance.

Persistent textual memory allows language models to carry information across long interactions, but learning what to remember is fundamentally a credit-assignment problem. A memory rewrite may only become useful many steps later, while much of the observed utility may be inherited from information already stored before the rewrite. We introduce Memory Gain Policy Optimization (MGPO), which isolates the incremental value of each memory rewrite by crediting it for its marginal contribution to current and future downstream utility. This turns delayed memory utility into a direct learning signal for optimizing what information should persist. We study MGPO on document-level information extraction, where structured supervision makes the effects of individual memory updates directly measurable. MGPO improves extraction while reducing average memory length by nearly 80% relative to the initial memory policy before optimization. The learned memory policy also supports reuse and transfer across domains, downstream models without further training. These results show that effective memory learning depends not only on preserving useful information, but on identifying which memory updates create lasting incremental value.

The BITEM team entered both subtasks of the NTCIR-19 R2C2 task with a single agentic pipeline, in which a model searches, reads and records evidence over a movie corpus while an orchestrator holds the record and rules on what may be submitted. A claim is admitted only once an entailment cascade has checked it against the passage it cites, and an answer is released only once enough checked evidence stands behind it. Each question is run three or four times, every pass retrieving from a corpus stripped of what the earlier passes have already seen. The confidence filed with each answer is computed by the orchestrator from what the run leaves behind and is never asked of the model, which is offered no way to rate itself. The two retrieval runs placed 4th and 5th of 22, pooling the passes was worth 0.0709 nDCG@20, and the gain was largest on the multi-hop and post-processing-heavy questions, where the organisers rank the pooled run top of the field. Sixteen of the 25 answer runs were built on passages these two runs supplied, 12 of them filed by other teams. HMR rewards a system whose confidence is high where it answers right and low where it answers wrong. The pipeline reached an accuracy of 0.9219, 6th of 25, while the confidence filed with those answers gave an HMR of 0.4915, 13th. A few rules crafted over those same recorded signals, with no further model call and no further retrieval, raise that to an accuracy of 0.9375, 5th, and an HMR of 0.6985, 9th. Ranking on HMR alone can reward a system for answering wrongly with low confidence, so we propose accHMR, the accuracy multiplied by HMR, which reports the reward in proportion to the accuracy, and on which the revised rules would have scored 0.6549, 5th. For future work, fitting a model on the numbers the pipeline already produces, rather than writing such rules by hand, would be a real step forward.

We evaluate an auditable long-term memory system on LongMemEval-S. Its retrieval chain uses hybrid candidate retrieval, cross-encoder reranking, coverage-first packet compilation, and deterministic reasoning scaffolds; an LLM is used only as a replaceable final reader. The chain places all gold sessions in the candidate pool for 468/470 answerable questions and produces gold-complete packets for 462/470. With a Claude Opus reader called through an unpinned CLI alias, two 500-question passes score 479/500 and 475/500 under GPT-4o. The 72 answerable knowledge-update rows used a substantively modified scoring prompt whose effect under the official text has not been measured. The pair straddles Chronos High's published 478/500; differences in reader generation, scoring prompt, and possibly data version, plus within-system variance, establish neither superiority nor equivalence. A grok-4.6-high reader on the same packets scores 476/474, while a maximum-reasoning-effort agentic variant regresses to 461/465. The headline passes differ on eight verdict-flip rows. A second judge agrees with the headline judge on 493/500 rows (98.6%) in each pass and scores both passes 472/500; the official judge also flips three verdicts when re-scoring byte-identical pass-1 answers. Negative controls rejected a verifier that repaired three wrong drafts but broke eleven correct drafts. All components were developed on the same 500 questions, with no held-out evaluation or independent human adjudication; retrieval and scaffold method sources and transcript-derived audits are held; and the headline reader received extra operator context, its complete requests were not retained, and MCP tool availability is unresolved. We release materialized packets, scaffolds, reader outputs, judge verdicts, and controls for inspection and re-scoring.

In recent years, the performance of large language models (LLMs) on reasoning tasks has been remarkable, even surpassing human capabilities on various benchmarks. However, there remains a lack of clear understanding in the academic community regarding how the structure and internal parameters of LLMs progressively solve complex reasoning problems. In this study, we investigate the inference process of LLMs on cross-linguistic materials and propose the hypothesis that LLM layers exhibit a structured division of labor across conceptualization, reasoning, and textualization. Based on this hypothesis, we introduce a bottleneck identification mechanism using sensitivity analysis to pinpoint the most critical functional stage for a specific task. Leveraging this insight, we propose a novel approach, Layer-Informed Fine-Tuning (LIFT), which achieves efficient and effective fine-tuning by selectively updating only these functionally critical layers. We then conduct extensive experiments to show that the LIFT method not only accelerates the training process but also significantly improves model performance.

Understanding gender biases in large language models (LLMs) is increasingly important as these systems become embedded in decision-support tools with real consequences. Prior research has focused only on a small set of models, leaving open the extent to which gender biases are common and heterogeneous across LLMs. We address this gap across ten models released between April 2025 and June 2026, spanning nine vendors, using two paradigms: gender attribution to stereotyped phrases (Study 1) and moral judgment of abuse or torture against a woman or a man to prevent a catastrophic outcome (Study 2). In Study 1, two of ten models attributed masculine-stereotyped phrases to female writers more often than the reverse, while three models showed the opposite pattern. In Study 2, several models converged on a male-disadvantaging asymmetry that was directionally consistent with a documented human tendency to protect female targets from harm, though the specific conditions under which this asymmetry emerged varied by model; three other models, by contrast, showed no variation across conditions. These results indicate that gender-related biases are common in LLMs. Their direction and magnitude, however, are highly heterogeneous, to the point that some models behave in diametrically opposite ways to others. Bias auditing should therefore be treated as an ongoing, multi-vendor process, rather than a one-time assessment.

Chain-of-thought traces are widely read as records of how models reach their answers, informing debugging, agent auditing, and claims about reasoning. Testing this interpretation is difficult because natural-language thinking traces are rarely mechanically verifiable. We revisit it in iGSM, a synthetic grade-school mathematics benchmark designed to study thinking traces and used to support claims of learned reasoning and planning. Crucially, iGSM exposes the exact quantities and dependencies that a correct solution should use, allowing generated traces to be checked programmatically step by step and enabling us to test whether correct answers are reliably accompanied by valid traces. We first evaluate models trained exclusively on valid, minimal traces. Answer correctness and trace validity nearly coincide in distribution but decouple out of distribution: on the hardest instances, 31.6% of correct answers have invalid traces, over half of which pass all syntactic and arithmetic checks but fail semantic dependency checks. We then intervene on trace supervision. Non-minimal training traces induce non-minimal outputs, while re-asking the same problem with a different query reveals computations inherited from the original query, weakening minimality as evidence of selective planning. Shuffling tokens in 10% of training trace sentences preserves near-clean accuracy even out of distribution despite no trace passing verification. Swapped training traces likewise retain high in-distribution accuracy. We discuss the implications of these findings for chain-of-thought monitoring and interpretation in the context of AI safety.

The attention operation is naively position invariant. However, positional information is fundamental to natural language, and therefore a variety of explicit position encodings have been developed in transformer-based models, such as rotary position encoding (RoPE). Although explicit position encodings have long been assumed to be required, recent methods that interleave local mixing layers, such as sliding window attention (SWA) and gated linear attention, while not encoding position (NoPE) in global attention layers has recently been shown to be successful at scale. How and why this approach works is not well-understood. In this paper, we develop an explanation of how hybrid models of this sort can implicitly encode position at global NoPE layers. Supported by both theoretical and empirical evidence, our central argument is that SWA and gated linear attention induce a recency bias in the residual stream that propagates to, and is selected by, the global attention logits. Moreover, in contrast to the implicit position encodings found in models with only global NoPE attention, in which positional information arises solely from the causal mask, the recency bias in hybrid models can be maintained across long sequences. In addition to deepening our understanding of how hybrid models encode position, these findings may provide insights for how to encode position in a way that can extrapolate to longer sequence lengths indefinitely.

Vision-language models (VLMs) may accept false visual premises, answering questions about a target object's color, count, location, or state even when it is absent. We call this reliability-critical behavior a target-absence grounding failure. Existing visual-grounding detectors primarily rely on generated responses, hidden states, or uncertainty measures. We present the first framework to leverage internal routing decisions in Mixture-of-Experts (MoE) VLMs to detect target absence before generation and guide selective correction. We extract target-token routing probabilities from Qwen3-VL-30B-A3B-Instruct and Gemma-4-26B-A4B-it, train a separate L2-regularized linear detector for each model, and use its predictions to selectively invoke a target-aware review prompt. Using routing alone, the Qwen and Gemma detectors achieve ROC-AUCs of 0.9988 and 0.9956 on GQA-Inpaint and retain 0.8095 and 0.7781 on the external OBER dataset, respectively. The resulting routing-gated policy improves end-to-end accuracy on GQA-Inpaint and OBER by +22.25% and +12.17% for Qwen, and by +13.42% and +1.39% for Gemma, without modifying model weights. Further analysis shows that the signal is localized to the target-object token, emerges in early MoE layers, and is distributed across partially substitutable experts. Although cross-dataset threshold shifts require recalibration, false-positive review causes limited harm overall, suggesting that intervention risk can be controlled through joint selection of the detector threshold and review prompt. Overall, we show that routing probabilities alone preserve actionable information about visual perception, allowing computation already produced by an MoE VLM to support low-cost detection and selective visual regrounding.

Language-model (LM) harnesses enable LMs to operate effectively over long contexts using additional compute. However, existing long-context evaluations are insufficient for distinguishing modern harnesses, reflected by saturated accuracy across harnesses and largely similar evaluation costs. In this paper, we introduce a benchmark for evaluating both the effectiveness and efficiency of long-context harnesses. Our tasks require diverse retrieval strategies, including lexical search and semantic matching, together with strategic and adaptive reasoning over global and local context. Much of the context is semantically relevant but only a small subset is useful at each step, creating both a challenging search problem and different accuracy--cost tradeoffs across processing strategies. For example, one task requires identifying every person satisfying several conditions using evidence scattered across documents; strategically checking the most selective condition first can narrow the search before verifying the remaining conditions. We evaluate multiple families of frontier language models with four state-of-the-art harnesses. Our benchmarks remain challenging even for strong model--harness combinations: the best reaches 68\% macro-average accuracy across four evaluation suites. More importantly, we find that the same underlying model can exhibit markedly different efficiency under different harnesses. Our results establish efficiency as an important axis for long-context evaluation and provide a testbed for developing harnesses that process context strategically rather than exhaustively.

Transformer language models (LMs) are feed-forward: deep-layer representations are never fed back to shallower layers, and the only pathway for information to flow downward across generation steps is the decoded token. This narrow channel forces models to recompute intermediate results and to discard alternative continuations. In this work, we remove this bottleneck during pretraining, introducing the LIFT (Latent Information Feedback Transformer) architecture and training method which enable LMs to propagate state across generation. We achieve this by turning recurrent-state learning into a teacher-forced prediction problem: each input token is paired with an information-dense state, derived from the next-token distribution of an off-the-shelf pretrained LM. The model, extended with a small number of additional parameters, is then trained to predict both the next token and the next state. As the input states are precomputed, pretraining remains fully parallel across positions. At inference, the model's own predicted states are fed back, with a minor computational overhead that decreases with model size. Experiments with pretrained models ranging from 135M to 1B parameters show that LIFT consistently outperforms standard Transformers and baselines on language modeling, downstream reasoning tasks, and procedural tasks under token-matched budget, while being on par with or ahead of compute-matched Transformers. Moreover, a controlled study on a state-tracking task shows that a tiny LIFT outperforms same-size Transformers trained on 8x more data, even when trained with the states of a Transformer that fails the task. Overall, we show that LMs can learn to exploit deep-to-shallow feedback during pretraining via scalable teacher supervision.

Linear attention replaces growing KV caches with fixed-size recurrent states, yet these persistent states can become a substantial memory bottleneck under concurrent serving. Directly quantizing recurrent states to low precision often leads to severe accuracy degradation, as quantization errors propagate through successive state updates. We discover that the impact of these errors depends on two complementary dimensions: temporally, errors in long-lived memory can persist across many decoding steps; spatially, errors in different key rows affect model outputs differently, while state magnitudes vary substantially along both rows and columns. Motivated by these observations, we propose STEPQuant, a spatial-temporal post-training quantization framework for Delta-rule recurrent states. STEPQuant allocates precision according to error magnitude and memory lifetime, and jointly fits key-row and value-column scales based on state distributions and key-row impact on output error. Experiments on Qwen3.8-27B and Kimi-Linear-48B-A3B-Instruct across both long- and short-generation benchmarks show that STEPQuant closely matches FP32-state accuracy under a nominal 6-bit budget and outperforms uniform INT8 in its 4-bit configuration. Integrated into SGLang with optimized GPU kernels, 6-bit STEPQuant achieves over 5x recurrent-state compression and reduces total serving memory by up to 68.7%. Our code is available at https://github.com/Dreamer-Toby/STEPQuant.

Understanding how individual traits influence team performance is valuable, but these traits are not always directly observable. Prior research has inferred traits like trust from behavioral data. We analyze conversational data to identify team traits and their correlation with teaming outcomes. Using transcripts from a Minecraft-based search-and-rescue experiment, we apply topic modeling and clustering to uncover key interaction patterns. Our findings show that variations in teaming outcomes can be explained through these inferences, with different levels of predictive power derived from individual traits and team dynamics.

Successful intercultural communication requires more than grammatical competence. It demands sensitivity to culturally embedded social norms whose violation triggers subtle but meaningful nonverbal responses. For German learners of Mandarin Chinese, acquiring this sensitivity is critical yet poorly supported by existing language-learning agents. We present a Wizard-of-Oz (WoZ) study design and supporting real-time system for collecting multimodal behavioral data from native Chinese speakers reacting to social norm violations by German learners. The system features a photorealistic MetaHuman avatar driven by Live Link face capture and MediaPipe upper-body tracking, a wizard console for real-time behavior selection, and synchronized multimodal logging across agent and learner streams. A layered annotation framework, based on psychological theory and covering non-observable socioemotional reactions, norm interpretation, verbal, and observable behavior thereof, and future supervision targets enables the corpus to support training of future automated cultural interpretation and behavior generation models. Four ecologically valid interaction scenarios, developed with cultural and pedagogical experts, provide the methodological and technical foundation for a culturally adapted conversational agent for Chinese language learning.

Given recent achievements of large language models (LLMs), frontier models are expected to perform well on Bayesian reasoning tasks, at least as well as humans. Furthermore, there is no reason to expect that LLMs will condemn others who offer those very same Bayesian judgments, a fallibility observed in human decision-making (Cao, et al., 2019). In 5 experiments with 48 experimental conditions employing over 5,000 trials, GPT-4o and Claude 3.7 Sonnet were tested on two variations of a Bayesian reasoning task. We also assessed LLM evaluation of the competence and morality of a hypothetical person who had offered the same reasoning task as them. LLMs hovered near human performance on the Bayesian task, though their reasoning was more rule-based and rigid. Surprisingly, like humans but to a greater extent, LLMs also demonstrated the same hypocrisy in condemning others who, like them, had deployed Bayes' rule. In demonstrating Bayesian hypocrisy, LLMs highlight a humanlike error of a dissociation between self-performance and other-judgment, and caution against their use in domains where statistical fidelity and fairness norms collide.

While neural theorem provers have achieved impressive milestones in formal mathematics, they largely operate on the assumption that faithful Lean 4 formal statements are already provided. Translating informal natural language into a formal language is a critical data bottleneck plagued by an "illusion of rigor": standard type-checkers accept statements that compile but drop hypotheses, introduce vacuous truths, or subtly alter mathematical bounds. To resolve this, we introduce Sage (Semantic Agent-Guided Formalization Engine), an agentic framework that replaces monolithic translation with a four-stage decomposed generation pipeline coupled with a dual-signal semantic correction loop. By pairing Lean 4 compiler diagnostics with multi-dimensional semantic feedback, our correction loop enforces mathematical fidelity alongside syntactic validity. By explicitly accounting for the gap between open-ended queries and declarative formal targets, our pipeline prevents models from achieving high formalization rates by guessing unverified answers (exhibiting a 70.9% answer leakage rate in monolithic baselines). Consequently, Sage suppresses leakage to 2.7% while achieving 73.3% pass@4 joint compilation and semantic fidelity on the Omni-MATH without proofs (compared to 42.0% for a fine-tuned Goedel-Formalizer-V2 baseline). Finally, on IMO-Unformalized, a novel frontier of 175 unformalized International Mathematical Olympiad problems, Sage demonstrates effective zero-shot generalization with 87.4% pass@4 verified fidelity compared to just 19.4% for the baseline, winning over 79% of blind pairwise evaluations.

Compact surrogates could reduce the cost of simulating large language model societies, but must reproduce collective behavior. We compare individual predictions and collective forecasts using 9,455 published trajectories and new experiments on opinion dynamics. Neighbor information improves individual prediction in all 16 public-data settings and pooled collective forecasts on held-out questions, although collective gains depend on transfer conditions. Tests on 24 new statements do not confirm earlier contrasting history effects in forecasts from the initial state. Qwen benefits from history after three observed rounds. These findings motivate direct collective validation, explicit limits on available observations, and comparisons with simple baselines.

Running a language model on edge hardware provides private and low-latency reasoning without a network connection, and yet the small models that fit on such devices are unreliable on the tasks computers are expected to handle well, such as arithmetic, algebra, and formal logic problems. We argue that much of this unreliability is avoidable. Many queries appearing to demand reasoning are in fact structurally deterministic and permit fast and exact symbolic solutions. Therefore, forcing a probabilistic model to approximate them sacrifices accuracy and energy for little benefit. We present a neurosymbolic router that classifies each incoming query and dispatches it to the cheapest correct solver, sending structured tasks to deterministic engines and reserving the small language model (SLM) for open-ended word problems. Instead of hand-coding the routing logic, we learn a deterministic finite automaton (DFA) with the L* grammatical inference algorithm, using the SLM as a membership oracle and labeled data as an equivalence oracle. On a Raspberry Pi 4B (8 GB RAM, no GPU), evaluated on 100 untested prompts from DeepMind Mathematics, GSM8K, and RuleTaker, learned routing attains 100% routing accuracy and 98.3% overall accuracy with a 512-token reasoning budget (93.3% on word problems), compared with 72.0% for the strongest agent baseline, Program-of-Thought, and 58.7% for a tool-calling agent given the same solvers. Since formatted queries never reach the model, the router answers them in 1-11 ms and, in its 30-token configuration, runs 8.8x faster and 2.8x more energy-efficient than Program-of-Thought.

Pre-tokenisation restricts which text fragments can become prediction units, but its compression cost is obscured when tokenisers are compared only under the same boundaries. We measure this cost by bounding the minimum token count from both sides, with and without a regular-expression boundary rule. Nonnegative prices on token occurrences yield a lower bound through shortest paths and vocabulary-budget selection; maximising over all prices recovers the linear programming relaxation, and an independent integer checker certifies the reported values. On English Wikipedia, boundaries increase the optimal token count by 28.3--36.8\%. Byte pair encoding lies 2.1\% above the constrained lower bound, but 10.9\% above the unrestricted bound. Compression and prediction favour different dictionaries: at 85M non-embedding parameters and matched training-token budgets, unrestricted fitting yields higher mean held-out bits per byte under a common unrestricted decoder in all 12 languages in the paired study and 11 of 12 under independent tuning and evaluation. To study intermediate boundary policies, we introduce boundary licences, which limit the vocabulary entries permitted to cross cuts and admit the same form of certificate. On separate English and Chinese fitting corpora, licensing 10\% of the vocabulary budget recovers 85.2\% and 100.0\% of the achieved token-count reduction from removing all cuts. These results quantify the compression cost of boundaries while separating it from the prediction quality of the resulting token units.

Real-world embodied agents often pursue independent objectives within a shared physical environment, where their actions can alter the conditions faced by others. Existing benchmarks, however, typically assume shared goals or explicitly prescribed interaction protocols, leaving such emergent physical coupling underexplored. We introduce VehicleArena, a 3D urban-driving benchmark for studying independently operating agents in a dynamic shared world. In VehicleArena, LLM-controlled agents must fulfill evolving passenger requests while navigating complex traffic, and each agent's driving decisions can reshape traffic flow, delays, risks, and subsequent observations for surrounding agents. The benchmark provides 112 evaluation tasks spanning single-agent and multi-agent driving. Across nine evaluated models, the highest arrival rates reach only 65.0% on single-agent tasks and 65.6% on multi-agent tasks, while strong passenger-request or cabin scores do not reliably translate into successful trip completion. Moreover, in matched multi-agent runs, every tested focal policy reduces the arrival rate of surrounding vehicles relative to the simulator's native traffic controller, revealing measurable externalities beyond the focal vehicle itself.

When a doctor, a judge, or an engineer must decide whether to trust an AI model's output, they cannot avoid asking what the model understands. Purely mathematical or statistical descriptions struggle to distinguish trustworthy from untrustworthy outputs without reintroducing the question of AI understanding in all but name. Yet the question is ill-framed as it stands, because the inherited concept operates within a monophonic paradigm: the idea that a cognitive system's understanding of something must be localised to a single mechanism underpinning all the capacities conferred by such understanding. Drawing on a wide range of mechanistic evidence, we show that LLMs are pervasively polyphonic: outputs emerge from coalitions of parallel mechanisms of uneven reliability, which variously complement, duplicate, or drown out one another, with several coalitions sufficing for a task without any one being indispensable. Polyphony not only complicates attributions of understanding, but renders monophonic inference patterns hazardous. In response, we develop a conception of understanding fit for polyphonic AI. It centres on sound circuitry that is reliably and correctly recruited and in control of outputs. Attributions of understanding thereby become tractable claims about internal organisation, and can do the work of guiding trust in AI.

We introduce GeoOutageBench, a benchmark for assessing LLM-based geospatiotemporal KGQA for multimodal outage and resilience analysis. Unlike existing KGQA benchmarks for Web knowledge, GeoOutageBench considers a spatiotemporal KG that integrates visual, textual, and structured data from outage records, remote sensing, weather observations, storm and power events, geographic entities, and domain ontologies. It provides a competency query taxonomy at different difficulty levels from spatiotemporal containment and proximity, spatiotemporal co-occurrence analysis, multimodal evidence, to hypothetical evaluation. Over multimodal KG and query classes, GeoOutageBench provides user-configurable evaluation of three important, highly coherent yet less studied tasks: (1) LLMs' understanding for ambiguous geospatiotemporal questions in terms of NL to SPARQL interpretation, (2) query-driven assessment of ontology utility, and (3) answer accuracy of multimodal KGQA retrieval. GeoOutageBench provides a design principle and foundation for assessing LLM-KG systems that support real-world infrastructure resilience analysis. Our benchmark, source code, data, results, and other documentation are available at https://github.com/UCF-SAGE/GeoOutageBench.

Using predictive models to explain cognition requires more than accurate behavioral predictions. Input ablations offer an appealing route: remove information from a model and interpret the resulting performance change as evidence of its importance for behavior. Yet this inference assumes that the model's dependence on information reflects the dependence of the process generating the behavior. We test it in two synthetic sequential bandit tasks with known generating policies, where past choices can remain informative when feedback is unavailable to a predictor. We compare GRUs and Transformers trained from scratch, a fine-tuned LLaMA model, and cognitive models across systematically varied reward contributions. Our analyses distinguish prediction after training without reward observations from the response of a fixed predictor to donor-reward replacement. Three findings emerge. First, in the restless task, neural models trained without rewards predict held-out choices better than four simple training-fitted behavioral baselines. Second, under matched donor replacement, accurate predictors can respond much less than the known generator. Third, at some reward weights, neural networks predict better than a pooled reinforcement-learning model but have less faithful changes in choice probabilities; the model ordering differs between the two tasks. These independent-test results separate information sufficient for prediction from response fidelity under a specified ablation in sequential choice. They motivate validating model-ablation responses independently of predictive performance before using them to infer how the observed behavior was generated.

Large language models can reason in continuous space instead of decoded text, by recurring on their own hidden states or by passing those states between agents, while training supervises only the Cross-Entropy (CE) of the final decoded answer and does not constrain the thought. Theoretical and empirical analyses establish and confirm four failures of CE-only training that lead to a lower probability of the correct answer such as collapsing thoughts across distinct questions and retaining irrelevant information. We introduce REST (REpresentation-Supervised Thoughts), a training objective that turns four properties of a valid thought representation (causality, minimality, separability, and stability) into differentiable losses added to CE. We instantiate it in latent single-agent and multi-agent systems, without architectural changes or added parameters at inference. Across 7 benchmarks spanning mathematics, science, medicine, and code generation, with the same training data, compute, and latent budget, REST increases accuracy over CE-only training across agent settings and model sizes by up to 7.5 percentage points and convergence on a final answer by 30\%. Furthermore, REST thoughts encode more of what is required to achieve the correct answer, and decoding them better recovers the intended output of the agent, which makes latent communication easier to interpret. Project Website: https://fard-lab.github.io/REST

Reliable evaluation of large language models (LLMs) is essential for their development and deployment, yet is often costly, risky, and difficult to perform safely online. We study off-policy evaluation for LLMs, where limited human-labeled data from a behavior model are used to evaluate a newer target LLM. This setting is challenging because labels are scarce, behavior--target distribution shift is common, and response likelihoods are often unavailable for black-box LLMs. We propose the Optimal Transport-based Robust Off-Policy Evaluation (OTROPE), a likelihood-free evaluation that performs distributional correction in a semantic space via optimal transport to align labeled behavior-policy samples with unlabeled target-policy samples. OTROPE combines corrected human-labeled residuals with proxy predictors, yielding a doubly robust-style evaluation without behavior-policy modeling or density-ratio estimation. We theoretically characterize why baseline evaluators fail under LLM distribution shift, and establish consistency and convergence rates for OTROPE when either the reweighted behavior distribution or the proxy predictor converges. Experiments on synthetic and real LLM evaluation tasks show that OTROPE consistently outperforms baselines while enabling ensembles of weaker LLM evaluators to approach and sometimes surpass stronger evaluators. Code is available at https://github.com/LinerXiang/OTROPE.

Large language models can learn abstract rules from just a few in-context examples, but how their internal mechanisms activate as examples accumulate is not well understood. We trace a three-stage symbolic reasoning circuit (abstraction, induction, retrieval) across shot counts in three model families and find it is detectable and functional well before the model achieves high accuracy. Per-head causal contribution grows up to 8x from 1- to 10-shot, and cross-shot activation patching raises accuracy from 1% to 56% at 0-shot and 17% to 88% at 1-shot. Function vectors scaled and injected at 0-shot rescue accuracy up to 86%, largely substituting for the induction stage but depending critically on an intact downstream retrieval stage. The infrastructure for abstract rule-following is present in the weights before any demonstrations; in-context examples, function vectors, and related interventions appear to supply input to the same latent circuit.

We investigate the extent to which the Hamas attacks on Israel of October 7, 2023, have affected German social media debates about Judaism and Israel. For this, we develop an approach to detect 26 anti-Semitic categories in user postings via large language models (LLMs). The approach is applied to Facebook and Telegram posts (N=125,718) from three months before and after the event. Methodically, we test different open-weight models in two setups---with and without user information as additional context to the post text. The best setup achieves up to 83 % F1-score for binary anti-Semitism detection on our manually coded validation set. User context provides valuable information for most LLMs and drastically reduces false positives, for example, when (critically) reporting on anti-Semitic incidents. Concerning our topic, we find that anti-Semitism is surging significantly on both platforms, while being about ten times more prevalent on Telegram compared to Facebook. Facebook users express anti-Semitic views most likely in posts about an alleged genocide in Gaza carried out by the Israeli army, whereas classic anti-Semitic stereotypes related to power and conspiracy theories are dominant on Telegram. After the attack, the discourse patterns on both platforms show signs of convergence, as classic anti-Semitism increases on Facebook, whereas Israel-related categories surge on Telegram.

Sparse autoencoders (SAEs) expose interpretable features in large language model activations, yet existing structured SAEs impose single-parent trees or forests, while post-hoc graphs permit multiple parents but neither guide feature learning nor ensure reliable relation recovery. We introduce the Adaptive Graph Sparse Autoencoder (AG-SAE), a structure-guided training paradigm that treats each feature's complete parent set as an atomic structural hypothesis and lets evidence select zero, one, or multiple parents. By competing complete parent sets against null, subset, and alternative explanations, AG-SAE identifies jointly necessary multi-parent relations while rejecting redundant or spurious alternatives and verifying that each child contributes beyond its parents. The induced topology over SAE features then defines a differentiable structural loss that guides SAE training, while topology-guided refinement mitigates feature absorption and uses persistent reconstruction gaps exposed by the learned structure to initialize new features. The entire graph is then induced again from the revised dictionary by reassessing every feature's complete parent set, closing the dictionary-graph self-consistency cycle. Experiments demonstrate exact mixed-topology recovery in a controlled toy model, greater relational reliability and semantic validity than structured and post-hoc baselines on real LLM activations, and stronger feature-level causal interventions than conventional SAE features. AG-SAE thereby turns recovered mixed-topology feature structure into an unsupervised training signal that improves the dictionary, enables reliable feature organization beyond the topological limitations of trees, and exhibits stronger causal control beyond reconstruction.

Mixture-of-Experts (MoE) layers are central to frontier language models, and recent architectures push toward more and smaller experts. In this regime, the standard linear router becomes a bottleneck: with $M$ experts and hidden dimension $h$, its per-token cost $\Theta(Mh)$ dominates the MoE layer once $M$ is large. We introduce MoRE (Mixture of Rank-reduced-routed Experts), which factorizes the router weight matrix at rank $r$ and reduces the routing cost to $O((h + M)r)$. We prove that rank logarithmic in $M$ suffices for routing expressivity when the number of active experts is fixed, and is necessary up to precision factors. We also prove that logarithmic rank preserves load balance in a Gaussian memorization model, and training on a synthetic phonebook task shows that low rank does not hurt memorization. At matched active FLOPs, the factorization allows a factor of $\Theta(h/r)$ more experts. To realize this gain in wall-clock time, we design a fused Triton kernel at inference that avoids expensive memory operations on HBM. Empirically, MoRE improves memorization on the phonebook task and performance on knowledge-intensive Q\&A benchmarks after pretraining, while matching reasoning ability. Code available at https://github.com/Matheart/MoRE_code.

State Space Models (SSMs) compress sequence history into a bounded recurrent state, making the resulting memory law a central architectural choice for long-context performance. Most modern SSMs rely on ODE-based dynamics that lead to exponential forgetting, limiting their ability to retain information over broad temporal ranges. We introduce FRAC, a selective SSM architecture derived from fractional dynamics that replaces this exponential decay with power-law long memory. To make fractional dynamics practical, FRAC approximates the heavy-tailed target kernel with a finite-state, log-spaced sum of exponential modes. This construction turns fractional memory into an efficient recurrent module with parallel training and prefill, while retaining bounded-state autoregressive decoding. Extensive experiments, including 1.3B-parameter language modeling, demonstrate that FRAC consistently improves long-context performance over state-of-the-art SSM baselines while staying competitive on short-context. These results show that fractional dynamics provide a practical and effective prior for long-context SSMs.

Distortion-free watermarking enables reliable attribution of machine-generated text while preserving output distribution. However, existing methods operate independently on each generated token, making their detection capability fundamentally constrained by the entropy of the next-token distribution. We present Tandem Token WaterMark (TTMARK), a general pairwise watermarking framework that extends distortion-free watermarking from individual tokens to adjacent token pairs. By watermarking the joint distribution of consecutive tokens, TTMARK enlarges the effective watermarking alphabet from V to $V^2$, allowing the detector to exploit both token entropy and conditional entropy while preserving distortion-freeness over the joint distribution. We further introduce a branch-isolating concatenated tandem generation algorithm that efficiently constructs the joint distribution in a single forward pass. Theoretically, we show that pairwise watermarking achieves better expected detection strength in low-entropy regimes. Extensive experiments across multiple language models, datasets, and three representative distortion-free watermarking schemes demonstrate that TTMARK consistently improves detectability without degrading generation quality, while also improving robustness to edits and substantially enhancing localized watermark detection.

Large language models often preserve meaning despite substantial changes in wording, style, and syntax, while small semantic edits can systematically alter their hidden representations. This suggests that semantic variation may be organized along recurring local directions. We propose the Invariant Atom Hypothesis: local semantic motion admits preferred sparse coordinates along directions that remain stable under meaning-preserving transformations. We learn a shared semantic frame and sparse coordinates that reconstruct semantic displacements while suppressing nuisance variation, with anchor-dependent diagonal modulation adjusting atom strengths without sample-specific rotations. Empirically, the atoms exhibit strong semantic--nuisance separation, sparse reconstruction, reproducible directions, and causal effects on model predictions. The learned geometry generalizes to unseen semantic neighborhoods and nuisance families, while local reweighting improves semantic selectivity and preserves a consistent global-to-local structure. Atom signatures also remain stable under model modification. These findings support reusable invariant directions as a sparse coordinate system for local semantic geometry in language models.

Semantic-preserving transformations can induce substantial motion in learned representations, while small changes may strongly affect model predictions, raising a basic question: what local metric best captures semantically consequential variation? We propose Fisher-induced invariant representation geometry (Fisher-IRG), which measures local representation directions through their predictive sensitivity. Around each representation, we construct semantic-preserving and semantic-changing neighborhoods, aggregate their local Fisher information, and recover invariant directions through a contrastive generalized eigenvalue problem. Controlled displacement analyses first show that comparable Euclidean motion can have substantially different predictive consequences, supporting the need for a predictive geometry. Across language and vision models, Fisher-IRG yields stronger semantic-versus-nuisance predictive selectivity and generally more reproducible subspaces than covariance-based geometry, while recovering systematically distinct local directions. Representation interventions further localize semantic effects to the Fisher-derived subspace, and held-out separation and retrieval show that the recovered geometry generalizes beyond the discovery neighborhoods. These results support Fisher-IRG as a principled framework for characterizing local invariant representation geometry.

Long-horizon agents require context compression to manage growing interaction histories. Compression quality, however, is ultimately determined by downstream execution. Existing prompt-adaptation methods infer compression errors by comparing full-context and compressed trajectories. Such comparisons cannot isolate individual compressions and are confounded by agent stochasticity. We first find that compression degrades reliability before solvability. Using matched counterfactual continuations that compare execution from the same agent state with versus without compression, we further show that severe degradation concentrates at isolated compression events. Motivated by this finding, we propose PAIR (Prompt Adaptation using Interventional Rollouts) for adapting structured compression prompts. PAIR identifies individual compressions that degrade subsequent execution, diagnoses their effects, and revises the relevant sections of a fixed compression template. PAIR achieves the strongest cross-run reliability among compressed methods in every main benchmark-scope combination, consistently exceeding the competing prompt-adaptation baseline. Without modifying the downstream agent, PAIR brings compressed execution close to the no-compression baseline and sometimes numerically exceeds it.

Recurrent neural networks (RNNs) compress the historical context into a memory state of fixed size, thus allowing for constant-time inference. The memory state size is a crucial factor in their performance, as exemplified by the strong performance and resurgence of linear attention, which extends the vector-valued hidden states of ordinary RNNs to matrix-valued hidden states. Crucially, linear attention does so in a parameter-efficient way, in particular by using an outer product of the key and value vectors to write to the matrix-valued hidden state. We generalize this construction and propose triadic linear attention, which writes the triadic outer product of a key, a second key, and a value, into a third-order (i.e., 3D) tensor state, and reads from it by contracting both key axes with two queries. An $E$-dimensional second key thus yields an $E$-fold increase in state size while adding only two projections. Triadic linear attention is compatible with data-dependent forgetting, the delta rule, and chunkwise-parallel training. Applied to Gated DeltaNet and scalar-gated linear attention, triadic linear attention substantially improves long-context language modeling and recall, outperforming alternatives that enlarge the state.

Audio large language models (ALLMs) can reason about the content of audio recordings to perform complex tasks. However, these capabilities usually collapse in real-world environments when background noise and competing sources mix the target sound. Inspired by long-term memory in human listening, we propose Long-Term Memory-Guided Audio Enhancement (LTM-AE) to improve selective target perception by refining the audio representations of ALLMs without training. LTM-AE extracts representations in hidden states from separate clean reference recordings as long-term memory for each category, guiding enhancement toward a user-specified listening target. We reconstruct incoming audio tokens in the selected category long-term memory and interpolate the reconstructions with the original tokens before language backbone decoding. This interpolation controls the influence of stored auditory experience while keeping all ALLM parameters fixed. Diagnostic readouts across twenty sound categories and three ALLMs show that LTM-AE strengthens responses to a specified target amid three interfering sources. Averaged over constrained and free-form classification, accuracy gains over raw mixtures range from 29.53 to 46.15 percentage points across multiple open source models. For speech content recovery, LTM-AE with an additional learned token-level gate reduces Qwen2-Audio's word error rate from 23.07% to 14.77%. This work takes an initial step toward using principles of human long-term memory to enhance ALLMs for real-world listening. Our code is available at https://github.com/aynlp/ltm-audio-code

Pretrained transformers use little of their depth to follow references in context. Thirteen base models reliably follow only 1.4-3.6 lines, and extra pretrained loops add little. A task-trained rank-8 LoRA at one early layer extends this computation with all model weights frozen. Qwen3-8B improves from 15.5% to 99% exact accuracy on 24-line chains; a longer-trained LoRA reaches 50 lines. Ouro-1.4B reaches 60 lines after four loops and at least 160 after eight. The LoRA starts a relay: program lines pass on their chain identity through a short range of middle layers. Frozen heads read progressively further up the chain, and removing parent-line attention stops the relay. A frozen-model measurement locates the last useful intervention layer within tolerance in three of four held-out models. Task-specific LoRAs also improve MuSiQue. Default answers therefore understate the computation accessible through a tiny edit. Code and an interactive demo are available at https://lunamos.github.io/stop-thinking-too-early/

On-policy distillation (OPD) trains multi-turn language agents with dense teacher supervision on student-generated responses. However, standard think-then-act rollouts require lengthy reasoning before each short action, delaying environment transitions and experience collection. Generating actions directly reduces this delay but can degrade rollout quality. To address this, we propose ActFirst-OPD, an act-first, reason-later training framework that decouples environment interaction from full-response generation. The student infers and executes actions through reference-conditioned inverse dynamics using its current interaction context and a reference next observation, and switches to autonomous next-action prediction when the resulting transition deviates from the reference trajectory. From the collected interaction contexts, the student asynchronously generates full think-then-act responses for token-level teacher supervision. Experiments across 0.6B-, 1.7B-, and 4B-parameter Qwen3 students show that ActFirst-OPD achieves average wall-clock training speedups of $2.3\times$ on ALFWorld, $1.8\times$ on WebShop, and $4.9\times$ on ScienceWorld over Vanilla OPD. It matches or exceeds all compared OPD baselines in mean task success rate across eight of nine benchmark-model settings. These results demonstrate that reasoning need not block acting during multi-turn agent distillation.

Looped language models (LoopLMs) increase computational depth through parameter sharing, offering a path to scale inference computation without adding parameters. However, it remains unclear when additional recurrence is beneficial and how architectural choices affect its effectiveness. Through controlled experiments, we systematically examine (1) when recurrence helps, (2) where it should be applied, and (3) how its conditioning affects performance. Our evaluation covers inference budgets below, within, and beyond the training horizon under knowledge and reasoning tasks. (1) We find that recurrence can improve reasoning beyond the training horizon while degrading knowledge performance, but harder reasoning instances do not consistently benefit more. (2) Performance also depends on how distinct layers and recurrent iterations are allocated, showing that effective depth alone is insufficient to predict behavior. Non-recurrent output layers improve robustness to under-unrolling, while the preferred placement of input and output layers varies with inference budget. (3) Finally, we find that conventional initial-state injection offers limited robustness to varying recurrence depth. We therefore propose history-state injection as an alternative, and show that channel-wise history-state injection combined with timestep conditioning offers a low-cost and more effective design, better preserving knowledge under extended unrolling while improving robustness across inference budgets. Overall, our results clarify when recurrent computation helps, where it fails, and offer practical guidelines for designing LoopLMs across variable inference budgets.

Large language models make weight storage and memory traffic major inference costs, motivating low-precision formats that represent each weight with only a few bits. Such formats use a scale to map floating-point values into a small codebook; NVFP4 improves local range utilization by letting every 16 E2M1 weights share an E4M3 block scale. Choosing that scale is difficult in GPTQ because quantizing one column updates those that follow, so evaluating a block independently can misestimate its final reconstruction error. Large models pose a second challenge: full-precision weights, calibration activations, and second-order state cannot all remain on one accelerator, while assigning complete layers to devices leaves each time-consuming layer solve serial. We introduce \emph{Schur Replay}, a scale-selection algorithm that reproduces the GPTQ updates caused by each block scale and scores the resulting block error after accounting for compensation from unquantized columns. Separately, our execution infrastructure keeps only the active layer resident, tiers activations across device, host, and disk, retires full-precision layers after export, and distributes independent output rows across tensor-parallel ranks. Together, the algorithm and infrastructure attain $99.35\%$ and $100.84\%$ question-weighted recovery from BF16 across seven benchmarks on Qwen3.5-397B-A17B and Llama-3.3-70B-Instruct. On the 397B model, the infrastructure reduces measured per-layer time by $15.17\times$ over ModelOpt and $23.14\times$ over LLM Compressor, with lower memory used per GPU.

Molecular optimization is inherently iterative: a candidate is proposed, evaluated against several objectives, and revised while preserving a relationship to the source molecule. Most instruction-following models instead emit one edited molecule, forcing validity, property improvement, and similarity control into a single response. We introduce MARCO, an evaluator-grounded reinforcement-learning framework that trains molecular editors on bounded proposal--feedback--revision trajectories. MARCO aggregates shaped turn rewards into an undiscounted trajectory return for group-relative policy optimization. We evaluate two consequences of this training: Same-1 tests the trained policy under a one-response budget, while Same-5 tests whether the same policy can use verifier feedback when up to five responses are available. Across the three-objective MuMOInstruct benchmark, three Qwen backbones, and seen/unseen instruction splits, SFT-initialized MARCO obtains the highest product of property success rate and similarity in every reported primary setting. Same-5 further improves the observed score under the tested budget, while four-objective and public-checkpoint experiments test transfer across constraint sets and initialization regimes.

Large language models have achieved significant progress in event forecasting, yet their probability outputs exhibit systematic calibration bias that varies heterogeneously across different domains and question types, undermining the trustworthiness of probabilistic outputs for decision-making under uncertainty. However, existing calibration methods typically correct probability outputs after prediction is complete, without modeling the structural sources of bias within the prediction process itself. To address this challenge, we decompose probabilistic prediction over causal-temporal hypergraphs into three stages, evidence weighting, evidence aggregation, and source fusion, and propose CHAIN, which designs stage-specific mechanisms to mitigate bias at each stage: (i) modulating the temporal decay function by causal topological distance, (ii) aggregating approximately independent causal chains via Noisy-OR after direction-aware deduplication, and (iii) driving adaptive fusion by causal coverage and directional balance. Experimental results on cross-domain forecasting benchmarks show CHAIN outperforms existing methods in expected calibration error, Brier score, and accuracy. Our project is available at https://github.com/QwenQKing/Chain.

Agents increasingly build on code written by other agents, and they reimplement rather than reuse, growing the codebases later agents must work in. To measure how well agents design libraries for other agents, we introduce LibraryDesignBench, a two-phase benchmark in which an agent implements a full-featured library from a specification that defines required capabilities and potential use cases without prescribing the design. We evaluate the library through the correctness and simplicity of programs written by three user agents from different model families. The benchmark spans 242 expert-validated programming problems across 15 library-design tasks in four languages. On eleven of the fifteen tasks, agent designers reproduce the abstractions of the human-written production library. Downstream agents adopt agent- and human-written libraries alike but underuse them, reimplementing capabilities the library already provides. Our failure analysis finds that downstream agents write extra code mainly because agent-written libraries are rigid or hard to use, not because capabilities are missing. We also experiment with giving designers more prescriptive, agent-first guidance and having them test their library with subagents; this improves downstream scores and yields simpler programs. LibraryDesignBench provides both a testbed for evaluating library-design practices for agent users and an initial design baseline that improves downstream reuse.

Speech brain-computer interfaces (BCIs) aim to restore communication by transforming neural activity related to speech, language, or communicative intent into external outputs such as text, synthesized voice, or avatar control. Recent advances in intracortical and electrocorticographic recording, deep sequence models, and language-model-assisted decoding have enabled rapid progress, including high-performance attempted-speech decoding and increasingly naturalistic speech synthesis. Yet these achievements also reveal that speech BCIs are not simply neural-to-text decoders. They are adaptive clinical systems in which neural representations, recording hardware, decoding architectures, language priors, feedback, and user learning interact over time. Here, we synthesize speech BCI research from a system-level perspective. We first examine the neural substrates of speech and language, emphasizing their hierarchical, distributed, temporally structured, and non-stationary organization. We then examine recording and decoding choices, closed-loop adaptation, evaluation, clinical translation, and ethics. Across these domains, we highlight recurring trade-offs between signal resolution and invasiveness, low-level motor and high-level semantic targets, decoder accuracy and user agency, and language-model fluency and faithful neural evidence. We argue the next generation of speech BCIs should be evaluated not only by offline accuracy, but also by robustness across sessions, calibration burden, latency, uncertainty, usability, and safeguards against unintended decoding. By reframing speech BCIs as adaptive, user-centred systems, we outline the interdisciplinary priorities spanning speech neuroscience, neural engineering, machine learning, clinical practice, and neuroethics needed to move from proof-of-concept decoding toward reliable, expressive, and controllable communication neuroprostheses.

The vocal tract is the region of the human body responsible for filtering one's voice to create speech. In this paper, we present a differentiable and GPU accelerated acoustic simulator for the vocal tract. The differentiable simulator synthesizes speech by propagating sound along an acoustic tube model of the vocal tract, and via its gradients, can solve the inverse problem: reconstructing the shape of the vocal tract solely from the sound it produces. Although the inverse mapping between geometry and sound is notoriously non-convex, we discover that gradient descent succeeds with three technical contributions: (1) we design a frequency domain formulation of the vocal tract's fluid dynamics that is 70x more GPU parallelizable than finite differences in time, (2) we integrate a differentiable model for turbulence to synthesize consonants, and (3) similar to prior work in implicit neural representations (INRs) and neural fields, we find that parameterizing the geometry with a neural network accelerates convergence and escapes local minima that trap discrete representations. Because the simulator is differentiable, it is readily integrated with other deep learning pipelines to enable novel linguistics and medical imaging applications. (1) We demonstrate self-supervised autoencoding of vocal tract shapes across 11 languages, and (2) we couple our simulator with a generative model of MRI (magnetic resonance imaging) images to reconstruct one's moving vocal tract from only their speech without paired data.

Optimizer momentum is usually stored as a parameter-sized moving average of past gradients, which makes history costly and fixes each past signal in the coordinates in which it was computed. We introduce Backpropagated Output Momentum (BOM), which instead stores a compact moving average of prediction errors at the model output and reprojects that history through the current network at every step. A batch-level analysis characterizes the information retained and omitted by this relocation, while the implementation preserves the current supervised gradient and can replace the first-moment component of several adaptive optimizers. As a plug-in for momentum-based optimizers, including ones that already compress their state, BOM reduces parameter-shaped optimizer state by 49.7-99.8% in three compositions and, averaged over three language backbones, paired step time by 4.0%. It also improves mean validation performance across language and vision fine-tuning, by 1.42 points in the primary five-task comparison. Language and vision pretraining studies, together with matched mechanism controls, further test the construction across output spaces and model scales.

Reinforcement learning with verifiable rewards (RLVR) has become a standard paradigm for improving large language models (LLMs) on various tasks, yet its sparse outcome rewards lack token-level credit assignment for intermediate steps. To address this, on-policy self-distillation (OPSD) leverages a self-teacher with privileged context to provide additional dense learning signals. However, because the self-teacher is often overconfident and imposes excessive penalties on long reasoning trajectories, OPSD frequently struggles in practice. To mitigate this, we propose self-instructing policy optimization (SIPO) with a contrastive self-teacher to provide dense credit. At each iteration, SIPO samples multiple rollouts per prompt from the current policy, scores them with environment rewards, and constructs two teacher contexts for each rollout by pairing the reference answer with mistakes made within the group. The model then re-evaluates its own responses under both contexts, using the difference between the two teacher log-probabilities as token-level feedback, so that biases shared by both contexts are expected to largely cancel. The resulting objective yields a token-level advantage for every rollout: the reward still sets the main direction of each update while the self-teacher redistributes credit across tokens. Even in groups where every rollout fails and group-relative advantages vanish, SIPO still provides a learning signal. By preserving direct optimization of the task reward while providing dense, token-level feedback, this approach bridges reinforcement learning and on-policy self-distillation. Extensive experiments across multiple reasoning and code-generation benchmarks demonstrate that SIPO outperforms both RLVR and OPSD baselines without an external teacher or additional generation.

Self-rewarding reinforcement learning (RL) enables large language models (LLMs) to self-evolve without human labels. Existing ensemble-based methods construct reward references from rollout groups and assign rewards accordingly. However, a response's reward representation also depends on its randomly sampled group context, i.e., the other responses in its group. Using only one group-context realization may miss desired reward signals and provide unreliable guidance for policy optimization. To address this issue, we propose Group-Marginalized Advantage Estimation (GMAE), which aggregates reward realizations across possible contexts into a response-level distribution and estimates expected advantages. Experiments across eight benchmarks and four base models demonstrate strong performance and cross-domain generalization. GMAE also exhibits stable learning, low extra cost, and good applicability across training datasets and RL backbones.

Explicit prosodic cues may help automatic speech recognition (ASR) of spontaneous speech, but auxiliary representations typically require additional trainable components, making it unclear whether gains come from the auxiliary information or the fusion mechanism. We address this using a frozen HuBERT backbone and a 64-dimensional representation trained to predict log F0, voicing, Delta log F0, log energy, and spectral tilt. We compare a frozen-backbone recognizer (Baseline), trainable fusion with zero auxiliary input (Null), and the same fusion supplied with the learned representation (Learned). Across Buckeye, Switchboard, and AMI IHM, Null reduces WER by 0.71-1.45 points over Baseline, whereas Learned differs from Null by +0.07, -0.09, and +0.00 points, with no significant differences. However, removing or mismatching the representation at inference increases Learned WER. Thus, Learned depends on the representation yet shows no measurable incremental WER benefit over the parameter-matched control.

Multi-Head Latent Attention (MLA) enables expressive multi-head attention with compact caches for its content and decoupled RoPE paths, yet cache memory still scales linearly with context length and batch size. In this work, we establish a systematic model of MLA's dual-path quantization errors, characterizing their distinct effects on attention-output distortion and explaining the pronounced amplification of RoPE-path errors. Guided by this analysis, we introduce QuantMLA, a function-aligned framework for low-bit dual-path quantization. We derive path-specific transformation spaces that preserve full-precision computation while remaining fully fusible into model parameters offline, eliminating online transformation overhead. Within these spaces, QuantMLA learns path-specific transformations with function-aligned objectives: attention-output reconstruction captures the content path's coupled matching and aggregation errors, while positional QK reconstruction preserves the RoPE-induced component of the attention logits and admits a theoretical bound on output distortion. Across four MLA model families, QuantMLA enables, to our knowledge, the first reported joint INT4 caching of the content and RoPE caches with minimal accuracy degradation. Further compressing the content cache to INT2 while retaining the RoPE key cache at INT4 maintains competitive performance on challenging reasoning and code benchmarks. We develop a native low-bit MLA attention kernel that integrates unpacking and dequantization directly into attention computation. The physical cache layout provides 3.59x compression at 128K context, while a cache-pressure serving workload achieves 5.168x higher whole-job output throughput than BF16. The code will be released upon acceptance.

Omni-modal large language models (LLMs) are expected to answer a question using the modality it explicitly refers to. However, existing training paradigms rarely verify whether models actually follow this modality, because multimodal inputs from the same sample often provide redundant evidence for the same answer. In this work, we uncover a pervasive cross-modal shortcut in omni-modal LLMs: when asked an audio-related question, models rely on the image as much as on the audio, and sometimes even more. To systematically diagnose this behavior, we introduce the Factorized Modality Diagnostic, which independently swaps audio and images between samples to isolate each modality's causal contribution. Across two model families in different settings, we find that this shortcut persists throughout supervised fine-tuning and reinforcement learning post-training, while judge-based RL may further amplify such reliance on irrelevant visual information. Based on this finding, we propose DMC-Repair, which trains models on the same kind of cross-modal swapped samples while assigning supervision according to the modality specified by the question. This prevents models from exploiting the spurious correspondence between modalities within the same clip. Experiments demonstrate that DMC-Repair reduces the image-induced share of the answer effect by 59.9%, effectively suppressing the cross-modal shortcut without compromising audio-question answering performance. The reduction in shortcut reliance generalizes across two model families and zero-shot to an unseen dataset and an unseen benchmark, and persists through subsequent post-training. Code is available at https://anonymous.4open.science/r/DMC-Repair.

Group Relative Policy Optimization (GRPO) is widely used to train reasoning language models, where it computes advantages by centering and normalizing rewards across rollouts of the same prompt. For multiple rewards, GRPO sums the reward components and normalizes the total reward by its within-group standard deviation. The corresponding variance equals the sum of all pairwise reward covariances. For a fixed centered reward, larger aggregate covariance produces smaller advantages, and vice versa, allowing update magnitudes to adapt to reward dependence. However, correlated rewards with large scales can dominate this normalization and suppress signals from smaller-scale rewards. We propose Correlation-Normalized GRPO (CorrGRPO), which normalizes pairwise covariances into Pearson correlation coefficients. CorrGRPO keeps the centered total reward unchanged while balancing the influence of differently scaled rewards on the correlation-based normalization. This allows advantage magnitudes to adapt to reward correlations without the normalization being dominated by large-scale reward components. We compare CorrGRPO with GRPO and other variants on code generation, tool calling, and agent security, using models ranging from 0.5B to 8B parameters. These tasks all involve multiple rewards that can improve together or present tradeoffs. Results show improvements across three domains, including code generation, tool calling, and agent security. Our code is available at https://github.com/HKUST-KnowComp/CorrGRPO.

Visual agents solve problems by interleaving reasoning with image operations, and on-policy distillation (OPD) provides guidance from a strong teacher on student-generated interaction trajectories. However, image operations change the evidence available for subsequent reasoning, so local errors in evidence acquisition (Acquire), reading (Read), or answer grounding (Ground) can propagate through the trajectory and lead to incorrect answers. Existing multimodal OPD methods primarily construct or contrast auxiliary views of the original image to strengthen supervision, without explicitly modeling the connections between student actions, resulting observations, and subsequent reasoning. This limits their ability to provide corrections tailored to different failure stages. We introduce Reflection on Visual Evidence (ReVuE), an on-policy distillation method for visual agents. ReVuE compares multiple student-generated trajectories for the same query, summarizes the observed visual evidence, and diagnoses the first failure across the Acquire, Read, and Ground stages. The resulting reflections provide training-time context for the teacher. We group and reweight token-level distillation losses according to how strongly these reflections affect the teacher's predictions. This design translates trajectory-level evidence diagnosis into targeted token-level supervision, guiding students to improve their visual evidence acquisition and reasoning. Across 11 benchmarks spanning the Qwen2.5-VL and InternVL3.5 model families, ReVuE outperforms all evaluated OPD baselines in weighted-average scores for perception, mathematical reasoning, and general tasks. ReVuE also reduces redundancy in reasoning and tool calls while improving tool-call accuracy and task accuracy. Code is available at https://github.com/sylvain-wei/ReVuE

As the capabilities of large language models (LLMs) continue to advance, increasing attention is turning to how to translate their abilities into useful behavior. Personal agents bring this question into everyday settings, where models are expected to serve individual users and continually adapt to their preferences. With the underlying model held fixed, such adaptation relies on harness engineering: designing and evolving the surrounding layer that manages context, memory, tools, and execution. Despite rapid progress, the factors governing effective harness evolution remain insufficiently understood. To narrow this gap, we investigate three central questions concerning harness architecture, harness scale, and self-evolution algorithms through complementary empirical and theoretical analyses. Empirically, we introduce a preference-oriented benchmark and systematically characterize the capabilities and limitations of personal agents associated with these three dimensions. Theoretically, we formulate harness evolution as a learning problem and explain these phenomena through approximation, generalization, and optimization errors. Analyses of reachable policies, capacity under finite interaction evidence, and biased update dynamics provide theoretical accounts of the observed phenomena. Together, these results offer a unified perspective on the limits of personalization through harness evolution and inform future harness design.

Long-context reasoning is essential for complex and long-horizon tasks, yet the performance of large language models (LLMs) degrades as context length increases. Recent approaches address this by processing input chunk by chunk while maintaining a bounded textual memory in model context. However, premature information compression can discard critical details essential for subsequent reasoning. In this paper, we introduce Commit-on-Evidence Memory (CoEM), which learns when to convert source evidence into compact memory facts. Specifically, under a fixed context-memory budget, CoEM preserves potentially useful source excerpts verbatim in a pending set, allowing subsequent context to clarify their relevance before irreversible compression. As new context arrives, a learned policy revisits each pending excerpt and decides whether to promote it to the committed memory, retain it for further consideration, or discard it. A frozen verifier ensures proposed facts are accepted only if supported by retained excerpts and current context. To further guide effective memory management, we train this policy using reinforcement learning by combining fine-grained, step-level evidence rewards with final answer rewards. Extensive experiments demonstrate that CoEM consistently improves long-context reasoning. When evaluated on 6,400 documents long-context input, CoEM outperforms the strongest memory baseline by 10.4-11.4 F1 points on Qwen3.5-9B. Code repository: https://github.com/benmagnifico/CoEM.

Production RL for language models lets the sampler fall behind the learner and repairs the resulting mismatch with a truncated importance weight. We ask how long the sampler can go without a refresh under that correction, and find a cliff: on Qwen2.5-Math-1.5B and GSM8K, importance-corrected GRPO refreshed every 192 updates learns well for 180 steps and then degrades severely in all three data seeds before the refresh arrives. Published remedies for staleness act on the update; we act on the sampler instead. Decoupled cooling draws samples at temperature 0.8 while the learner, the reference model and the importance weights stay at temperature 1, with the behaviour probability recorded from the tempered distribution, so the learner's objective is unchanged. All corresponding cooled runs are stable, and the longer interval keeps what the short one delivered: at the same update budget, a cooled sampler refreshed every 192 steps matches an uncooled sampler refreshed every 96 at the end of training (0.857 for both) and averaged over it (0.79), whereas lowering the learning rate to a safe value ends 3-7 points lower. On Qwen2.5-Math-7B the degradation points at interval 192 predict that an interval of 144 is fatal without cooling and survivable with it; on two data seeds the uncooled runs degrade before their first refresh and the cooled runs pass it and end at 92-93% against 68-81%, with one cooled run degrading transiently late in the second cycle. The benefit has a window: at three times the safe interval and in a high-mismatch MATH setting cooling delays degradation without preventing it, stronger cooling is not better, and cooling without the correction collapses. Sampling temperature is a control on staleness tolerance, and temperature and refresh interval should be chosen together.

Repeated verifier calls are useful only when they contribute conditional information. We introduce VStress, an auditable replay contract, and VStress-CA, a correlation-aware allocation policy that estimates the conditional marginal information of an unqueried verifier on a sealed calibration split, discounts uncertainty, normalizes by call cost, and stops or abstains when the next call is not informative. The controller freezes its decision and cost ledger before joining the clean oracle; a dependence-shift alarm disables channel preference and falls back to exact-stop. The controlled audit gives the mechanism boundary: at 35% symmetric corruption, majority-5 improves balanced accuracy from 0.6578 to 0.7739, whereas at 65% it loses 0.1226 points. In the matched fixed-budget comparison, breadth, redundancy, and adaptive allocation obtain balanced accuracies 0.6048, 0.6375, and 0.6538, with 3.4216 calls per item and an RLVR score of 0.6417 for VStress-CA. Dependence diagnostics also increase from same-model repeats to cross-family channels, with conditional marginal gains of 0.0126, 0.0462, and 0.0913. These measurements turn correlation from a post-hoc warning into an auditable allocation decision.

Language model agents can be improved by updating their model weights or refining the harness that guides task execution. These components are coupled: weight updates change how the model uses the harness, while harness updates change the trajectories used for training. We propose VACE, Validation-Gated Alternating CoEvolution, which alternates agentic reinforcement learning with trajectory-driven harness refinement. After each RL stage, VACE reuses the collected trajectories to propose a harness revision and evaluates the incumbent and candidate with the updated model held fixed. The candidate guides subsequent training only if it improves validation performance. With Qwen3.5-9B, VACE achieves 45.26% test accuracy on OfficeQA and a mean partial-credit score of 75.19% on AutomationBench, exceeding weight-only RL by 6.43 and 9.09 percentage points and ungated alternation by 4.59 and 6.95 points, respectively. Across 44 harness proposals, 17 reduce validation performance at the updated checkpoint and are rejected before subsequent RL training, highlighting the importance of validation gating.

Recent approaches to reinforcement learning (RL) post-training for large language models increasingly remove the critic to reduce training instability and memory overhead. Even where a critic is trained, it is discarded once training ends, although it has learned to predict outcomes. We revisit this trend and show that a pretrained critic's ability to predict future outcomes can make it a valuable asset for efficient long-horizon reasoning. First, we find that instability in critic-based RL for long chain-of-thought reasoning is largely an optimization artifact: keeping policy updates small and low in variance restores stable convergence. Second, a well-pretrained critic estimates the posterior probability of eventual success from later trajectory states and unfinished prefixes. Its predictions provide outcome-derived, dense, per-prefix learning signals that, during policy optimization, require neither completed rollouts, step-level annotations, nor external reward labels. Building on this insight, we introduce Reward-Free Policy Optimization (RFPO), which repurposes a single calibrated, frozen critic as a rollout-level reward, a value baseline for generalized advantage estimation, and a success forecaster for unfinished prefixes. We further show that binarizing the debiased score stops the policy from exploiting the critic's length bias. Binarized, RFPO matches supervised PPO without a single label in the training loop, while cutting compute and memory overhead. This makes RFPO well suited to long-horizon reasoning tasks, where outcomes arrive late and generation dominates cost: because rollouts can be rewarded before they finish, training no longer has to pay for waiting on every trajectory to complete. Our findings challenge the prevailing critic-free paradigm and establish critic-based, reward-free optimization as a scalable and computationally efficient path for LLM post-training.

Modern coding agents can deliver increasingly large repository-level changes, and recent benchmarks reflect this by emphasizing long-horizon tasks with large reference implementations. Many benchmarks evaluate coding agents' implementation capability to produce correct code edits from detailed specifications. However, practical modular development tasks also require the perception capability of grounding user intent and high-level design to derive a specification. We introduce LoLBench to evaluate both capabilities through the entire proposal-to-implementation process on large software systems. It is a multilingual benchmark of 100 tasks across 29 software systems in five domains. Each task provides a human-written enhancement proposal with user intent and high-level design. On average, proposals contain about 5,000 words, software systems contain 2.4 million source lines of code (LoC), and implementation pull requests (PRs) change approximately 5,500 LoC. Across 28 agents we evaluated, the best agent resolves only 14% of tasks and achieves a 52.7% Fail-to-Pass (F2P) pass rate. Failure analysis identifies incomplete code localization as a major bottleneck, while providing reference-derived file trees alongside API specifications improves resolved rates by 16--22 percentage points (2.4--17$\times$), reaching at most 34%. These results show that both perception and implementation remain central challenges for coding agents in practical modular development on large software systems. LoLBench is available at https://huggingface.co/datasets/lolbench26/LoLBench.

Many of the qualities that matter most in how people learn and grow, how someone regulates their emotions, reflects on a setback, or stays aware of others during a difficult conversation, are not directly observable. They have to be inferred from how someone speaks, moves, and sounds over time, and they resist the kind of clean labeling that most machine learning pipelines are built around. We study this challenge through a case that is well grounded in psychological theory but rarely modeled computationally: self-compassion, the tendency to respond to one's own setbacks with patience rather than harsh self-criticism. We examine how it appears during structured reflective interviews in a technology-mediated training setting, where people naturally talk through socio-emotionally demanding situations. Since no existing dataset captures this kind of construct in this kind of setting, we collected and annotated 51 reflective dialog sessions using an independent, temporally overlapping annotation scheme grounded in established theory. We consolidate the underlying six-component psychological model into a three-class supervision space, balancing self-kindness and mindfulness against self-critical or overwhelmed states, and build a reproducible window-based pipeline that aligns video, audio, and text on a shared timeline. Unimodal models trained on each modality separately are compared against a simple probability-level fusion strategy, which yields modest but consistent gains over the best single modality. We close by discussing where each modality succeeds or struggles, what this suggests about how this kind of construct is actually expressed in reflective speech, and what would be needed to model it, and constructs like it, more effectively.

Mid-training equips pretrained large language models with specialized and reasoning capabilities, but the returns of this stage are bounded since additional serial compute yields little further downstream improvement and can even degrade some capabilities, which places a practical ceiling on how much compute mid-training absorbs. We revisit how this compute should be allocated to a single run or multiple similar optimizations. We find that branches forked from a shared checkpoint under various controlled recipe reaches measurably different regions of parameter space, and establish a form of compatible diversity that extending one run cannot supply. Therefore, we introduce Trajectory Soup, which distributes a mid-training budget over several independent branches, and consolidates strongest checkpoints selected on validation through intra- and inter-trajectory averaging into a single model. A local bias and variance analysis separates the two averaging levels, showing that inter-trajectory averaging removes residual error beyond the reach of averaging within a trajectory, while checkpoint selection carries a bias that bounds how many checkpoints are worth merging. Across model scales, learning-rate schedules, token budgets, and trajectory counts, Trajectory Soup improves aggregate downstream performance over the strongest single-trajectory average under matched budgets and keeps improving as budgets expand, with the advantage preserved after an identical post-training pipeline. These results position trajectory allocation and merging as a practical way to extend the compute-scaling frontier of mid-training beyond serial saturation.

An agent that uses tools typically responds to what the user explicitly asks, yet completing the task may require information the user never requested. Work on proactive agents mainly studies whether and when an agent should act on its own, not what information it should pursue. We study a distinct axis of proactivity: its content. Horizontal proactivity pursues unstated information that the current context already identifies, and vertical proactivity pursues needs that only earlier evidence reveals. A need graph, recovered from a benchmark's own decomposition, records which needs depend on which, so both forms, and whether the agent stops at the right time, can be scored from a transcript without a model judge. To learn this behavior, we propose Q&D (questioner and drafter), which trains a questioner to prefer the question whose continuation retrieves more of the required evidence, with no reward model or judge. On held-out splits of three multi-hop question-answering benchmarks, at equal retrieval spend, the trained questioner improves both forms of proactivity over the same model, prompted, and outperforms a prompted model $15\times$ larger in the same role on two of the three, and the gain persists after controlling for question volume and length. Without further training, we place the questioner in an interactive customer-service agent with a simulated customer, where it completes more tasks while asking fewer questions, and in retail it outperforms the $15\times$ larger model with fewer follow-up turns from the customer. These results show that proactivity depends not only on whether an agent acts without being asked, but also on what it chooses to pursue and when it stops.

Can reasoning models trick chain of thought (CoT) monitors and perform hidden computation without revealing it in their thinking traces? We show that the answer depends on the underlying task difficulty and the model size. Simple computations can be performed covertly; however, beyond a threshold depending on model size, successfully solving the task necessarily leaks a near-linear amount of information about the covert task input into the CoT. Therefore, sufficiently complex hidden computation always leaves an information-theoretic footprint. However, concerningly, this leakage need not be readable: Under plausible cryptographic assumptions, even a one-layer Transformer can encrypt its reasoning online so that no polynomial-time monitor can extract information about the hidden computation. Overall, our theoretical and empirical results provide a holistic view of both the opportunities and the limitations of CoT monitoring.

On-policy distillation, where a student learns from a stronger teacher's feedback on its own outputs, is a common way to pass reasoning to smaller models. We analyze what it transfers at small scale, distilling Qwen3-8B into Qwen3 4B, 1.7B and 0.6B students, in thinking mode (reason at length, then end the reasoning and answer) and, for comparison, in non-thinking mode (no separate reasoning phase). Long reasoning needs two abilities, solving a problem and knowing when it is solved, and we find that distillation transfers the first, but in thinking mode not the second. Solving improves at every size, up to two ceilings, which we measure comprehensively across both modes and all student sizes: a student's single attempt never exceeds what it could already reach in many attempts before training, and the smaller the student, the further it stays below the teacher. Stopping is where the modes part. In non-thinking mode every student keeps stopping; in thinking mode students stop ending their reasoning early in training, and the smaller the student, the less of this ability survives: the teacher signals a stop almost only where a student already ends its reasoning, so distillation teaches no new stops; it only keeps the student's existing stops that land on a right answer, and a weak student has few such stops. The smallest students often reach the right value but do not commit to it: they either rarely mark it or mark it and write past it. Together, these results describe how small students behave under on-policy distillation, and a diagnostic that separates answer marking, correctness and stopping.

Test-time compute scaling has emerged as a cornerstone of advanced machine reasoning, yet performing iterative deliberation directly within continuous latent representation spaces reveals a catastrophic pathology: the Deliberation Drift Cliff. While unconstrained recurrent latent models achieve initial reasoning gains at short horizons (K <= 4), their reasoning collapses when extrapolated to deeper thinking steps (K >= 16), dropping by 22% to 62% across standard logical benchmarks. We resolve the trilemma among expressivity, Lyapunov stability, and computational efficiency in test-time latent reasoning through a 22-round empirical and theoretical investigation. We demonstrate that strictly conservative scalar potential gradient flows suppress long-range drift (cliff 3.40%) but bottleneck peak reasoning accuracy at 32.73%, whereas unconstrained rotational flows achieve high symbolic expressivity (82.33%) but suffer a severe 36.87% drift cliff. To resolve this geometric duality, we establish Port-Hamiltonian Latent Deliberation (PH-LD) and propose the Direct-Gradient Pure-Tensor Helmholtz-Hodge Decomposition (DG-HHD). DG-HHD parameterizes the attracting flow as a tangent projection tensor network while orthogonally decoupling non-zero circulation (Hodge machine error 1.65e-17, contraction error 5.55e-17), eliminating runtime autograd dependencies to achieve 1.84x vector field and 2.09x RK45 rollout speedups. In a 15-arm symmetrical Pareto benchmark, DG-HHD achieves 58.67% peak accuracy (+25.94% absolute gain over conservative HHD) and retains 35.27% at K=32. Transferred to small language model (SLM) multi-hop causal reasoning, DG-HHD delivers monotonic compute scaling (49.33% to 51.56%) and suppresses out-of-distribution drift (cliff -0.66%). All 30 Level 0 deterministic invariants are certified.

Frozen multimodal large language models (MLLMs) now solve standard referring expression comprehension with a single prompted call, yet on adversarial benchmarks with same-category distractors and negation, even the largest models are confidently wrong, and resampling repeats the error. Models built from different data and architectures rarely fall for the same confounder, so their agreement is a strong label-free signal of the correct target. We present FORUM, a training-free test-time fusion of frozen MLLMs guided by two fixed geometric rules: agreement-based selection keeps the region supported by the most distinct models, and medoid localization returns an actual member box instead of a coordinate average, so one loose prediction cannot shift the answer. Fusing three open MLLMs, FORUM surpasses the 397B-parameter published reference by a relative 5% in mean accuracy on the adversarial Ref-Adv-s benchmark, and a plain averaging ensemble by 15%. The gains transfer to standard RefCOCO+, and a balanced lineup with no dominant member still surpasses the 397B model by 5%.

Serving an answer from a large language model requires deciding when to abstain, yet a verifier's ranking accuracy alone does not determine the error rate among served answers. We introduce PriceCheck, which builds a compact family of decision rules from label-free checks such as re-solving a problem. Each check has a price: its agreement rates on correct and incorrect answers and its cost per run. Prices fitted on a small, class-enriched labelled set compose into predictions of a schedule's coverage and cost, guiding which checks to run and when to stop. A calibration test then selects a schedule at a stated selective-risk target. In mathematics, the selected schedules serve 76.1% of answers on average and keep held-out selective risk below 1.5% on all 15 splits. Under the shared testing protocol, PriceCheck serves more answers at that target than reward models, a prompted judge, the generator's confidence and a trained correctness classifier. At matched coverage, it keeps the fewest wrong answers among these scorers. Across 118 diagnostic schedules, price-based coverage predictions have a rank correlation of 0.97 with observed coverage. These results show that choosing how checks are combined and stopped matters alongside how well a verifier ranks answers. Code is available at https://github.com/js-lee-AI/PriceCheck.

Thorough evaluation of vision-language models (VLMs) has become prohibitively expensive, as benchmarks span an ever-broader spectrum of capabilities and new models arrive at a relentless pace. Benchmark compression methods that preserve model rankings at a fraction of the cost are well studied for language models, but for VLMs the question remains under-explored. We present PRIMEBench (Pruning Redundant Items for Multimodal Evaluation), a vision-aware hierarchical benchmark compression framework that substantially reduces evaluation cost while preserving model rankings. This hierarchical framework operates in four stages: data cleaning to remove items answerable without the image and all-correct items, category representative selection to pick one benchmark per capability category, item pruning with Vision-Aware Variance (VAW), and category-count pruning. VAW combines inter-model variance with a vision-dependence score computed from multimodal embeddings alone, while encouraging coverage of diverse items within each benchmark. On models held out from item selection, it has the highest mean fidelity at the released 5% retention. The hierarchical design lets practitioners stop at any stage to match their compute budget; the released suite removes over 97% of items while preserving model rankings. Beyond compression, our analyses show how VLM evaluation behaves as model panels grow and evolve, providing guidance for designing future benchmarks that are more efficient, robust to model turnover, and explicit about the limits of evaluation-side pruning.

Large Language Models (LLMs) are increasingly used to enrich user queries in information retrieval (IR) so that a standard retriever such as BM25 can bridge vocabulary gaps with the target corpus. Any single LLM, however, is limited by its training data and architectural biases, and its enrichment behavior depends on hand-crafted prompts that must be re-engineered for each new model -- an expensive and poorly scalable process. We present MERGE (Multi-LLM Ensemble for Retrieval via Generative Enrichment), a two-stage framework: three heterogeneous 7-8B open-source LLMs independently produce candidate expansions, and a larger LLM generatively synthesizes them into a single query. To make prompt engineering scalable across the ensemble, we integrate a task-grounded Automatic Prompt Optimization (APO) loop into both stages. Unlike APO methods that judge candidates with an LLM evaluator, our loop scores each candidate by its downstream retrieval performance and runs a small tournament between the current champion prompt and optimizer-proposed drafts, terminating once the champion survives two consecutive rounds; a history-augmented variant additionally feeds the recent tournament trajectory back to the optimizer. MERGE is retriever-agnostic and issues a single BM25 pass with no rank fusion, no supervised document expansion, and no re-indexing. On five BEIR benchmarks (NQ, SciFact, FiQA, Touche-2020, DBPedia), MERGE improves BM25 nDCG@10 over the original queries by +2.1 to +14.9 points and matches or outperforms strong LLM-based query-expansion baselines despite using only compact open-source models. Ablations confirm that the Stage-2 ensemble beats any single Stage-1 LLM, and that task-grounded APO converts large seed-prompt regressions into consistent gains without hand-tuning.

Users of language-based assistive agents often make ambiguous requests. In response, an assistant can either directly act on its interpretation of the request --- risking misalignment with the user --- or ask a clarifying question. Which option is the most safe and helpful? A common approach is to ask questions that minimize uncertainty about the user's intent until a threshold is reached. However, this neglects the impact of uncertainty reduction on downstream performance, the costs of asking versus acting immediately, and the possibility that users may provide corrections without being asked. To navigate these trade-offs, we introduce Rational Enquiry via Value-of-Information Reasoning (REVOIR). REVOIR makes clarification decisions via inference-time reasoning about the value-of-information of a question, which captures the expected improvement in task reward due to the answer received. In two assistive tasks --- ambiguous question answering (CondAmbigQA) and preference-aligned household task planning (ADAPT) --- we show that REVOIR achieves greater success with fewer questions than approaches based on prompting, chain-of-thought, fine-tuning, or information gain, improving preference satisfaction on ADAPT by 13-15% over a fine-tuned clarification policy while requiring no training and asking five times fewer questions. Furthermore, when the assistant can receive cheap user corrections after acting, REVOIR naturally infers that asking questions is not always efficient, demonstrating the adaptivity of our approach. In contrast, we find that vanilla reasoning agents fail to adaptively clarify user requests, and request fewer clarifications as reasoning effort increases.

LLM agents accumulate interaction histories that grow linearly with task length, causing quadratic inference cost scaling and performance degradation from attention dilution. Existing context-compression methods learn what to discard offline: by contrastively optimizing guidelines, distilling compressors, or training compression policies. This incurs a substantial cost. Further, the compression policy is learned a priori and is not dynamically conditioned on the evolving test-time trajectories. In this paper we ask a complementary question: Which past interactions causally shape the agent's future decisions? We recast context compression as a causal decision preservation problem over discrete interaction units and introduce FOCUS, a training-free context compression framework that operates entirely at test time. Our method requires no offline data collection or fine-tuning, and is architecture-agnostic, attaching to any closed-API frontier model as a modular compression layer. We evaluate FOCUS on diverse agentic benchmarks including API and tool-calling, QA, web domain and multi-turn dialogue. Our method establishes new state of the art performance, cutting peak context by up to 48% and dependency by 73% while improving task success by up to 8.9 percentage points over uncompressed execution.

Language models tend to agree with whatever a user asserts, and post-training increasingly targets this sycophancy so that models evaluate claims on their merits rather than deferring to the user. Yet the same models are far more compliant when a wrong answer is attributed to a verified source, which is how retrieval results, tool outputs, and grounded-search content often present information. We measure this gap across five open-weight families and three closed APIs. A single verified-source note endorsing a wrong answer flips 45-88% of baseline-correct responses in seven of eight models, and compliance rises with how authoritative the note sounds. Source deference and user agreement are not behaviorally interchangeable inside the model: on matched items with the same wrong answer, causal interventions can selectively suppress one without equally affecting the other. In three open-weight families, removing a fitted source direction lowers source compliance by 65-80 percentage points while removing a user or assistant direction has far smaller effects, and removing the user direction shows the reverse preference. A separately fitted intervention derived from source-versus-user cue activations moves compliance in both directions while leaving the prompt text unchanged. An authority direction fitted on trivia also transfers to PIQA and multi-turn SYCON dialogues without refitting, and removing it lowers wrong-source compliance by tens of percentage points in four of five families with no detected change in MMLU-Pro or GSM8K accuracy at our evaluation sizes. Source deference and user agreement therefore need separate evaluation.

The common paradigm of reinforcement learning with verifiable rewards (RLVR) is to let agents make multiple attempts at a task, and optimize towards the successful ones. This becomes problematic in the realms of self-improvement, where tasks are so difficult that the agent has a low or even no chance of success, and where there are no teacher models or example solutions to distill from. In this paper, we introduce RLTL;DR. After each failed attempt, we show the policy the verifier outputs and let it write its own feedback, in the form of a single TL;DR insight. The next rollout is conditioned on all previous insights, and we sequentially sample rollouts until a solution is found. Moreover, we enable backpropagation on the in-context insights to internalize a direct task to insight mapping. On challenging tool-calling and coding datasets (filtered to Pass@128=0), standard GRPO training of a Qwen 3.5 9B Thinking policy stays flat at a Pass@1 of 0% to 1%. RLTL;DR breaks through this learning barrier, achieving a Pass@1 of 14-31% with insights in context during training and, crucially, 12-13% when no insight is in context at eval time. We identify that the key is the task to insight internalization. To study this further, we reduce our approach to SFTL;DR, training only on (task, insight) tuples, without showing or backpropagating on any rollouts. Training on only 4k of these tuples recovers almost the full performance of RLTL;DR and classical SFT on full rollouts. This demonstrates a promising compacted training paradigm of the form "on this sort of task, keep this sort of thing in mind", which we hope to inspire future research on.

Experienced professionals know more than just facts and conclusions. They know which cues matter, why a judgment is reasonable, and which action to take. Routine work records often leave out this tacit knowledge, making it difficult for Large Language Model (LLM) agents to use professional experience effectively. We introduce KUPAS MASTER, an experience engineering platform built around nine-layer cognitive corpus construction. It turns heterogeneous work records and practitioner interviews into traceable, reusable experience corpora for agents. Six case elements preserve the task process: context, cues, judgment, action, boundaries, and outcomes. Nine-layer cognitive corpus construction organizes tacit experience along nine extraction dimensions and stores the resulting assets in six libraries: rules, constraints, best practices, negative examples, corner cases, and skills. Semantic alignment, individual experience distillation, organizational consolidation, and cross-review preserve source evidence, conditions of use, and unresolved disagreements. The platform packages these assets into callable skills with explicit inputs, steps, dependencies, and stopping conditions, connecting experience collection to task execution and evaluation feedback. Using authorized samples from 20 randomly selected practitioners, the platform processed 1,576 source files into 23,024 individual experience records and 13,113 organizational assets. The evaluation spans multiple professional domains. Under common task inputs and scoring criteria, the base model, raw corpus retrieval-augmented generation (RAG), and KUPAS MASTER agent scored 70.63, 79.75, and 89.58, respectively. The KUPAS MASTER agent improved on raw-corpus RAG in all seven scoring dimensions. The platform provides a practical path from individual tacit experience to organizational knowledge and agent capabilities.

One of the current premises of mechanistic interpretability research is that detailed accounts of the geometry of neural network representations can tell us how models perform computations, and how to effectively intervene on them. While low dimensional manifolds have been observed for multiple concepts in the literature (e.g. numbers encoded on helices, days of the week on a circle, ...), with structure believed to reflect properties of data and tasks, the extent to which models rely on them for computation, and how they manipulate them, remains unclear. We characterize precisely the geometry of computation in a number-comparison task, as an abstraction of comparison for decision making, and how models utilize geometry in an elegant fashion to implement it. Specifically, we study the causal geometry of number comparison in Qwen2.5-7B-Instruct, a capable and widely studied open-weight model, and find Qwen largely uses linear representations of numbers despite the presence of curved geometry. To compare two numbers, the model first encodes each number along a vector and adds the two representations using attention and the residual connection, bringing them into a shared space in the residual stream. Then, the model uses MLP neurons to compare the pair of numbers on local regions in this shared space, which correspond to smaller intervals of input numbers, and combines these to obtain the position of the maximum. In fact, this reliance on linear representations for comparison also persists when the model compares three numbers. Our findings demonstrate that the manifold hypothesis can co-exist with linear representations: while concepts that are ordered may have manifold structure in representations, the model may use an underlying linear structure of the concept in certain computations.

Autonomous agents have made rapid progress in general-purpose computer use, but reliable automation of professional industrial engineering remains out of reach, as engineering workflows demand reasoning over geometric and physical constraints and dependencies preserved across software and design stages. We present EngiWorld, the first benchmark structured around the complete design loop: 1,301 expert-curated tasks spanning 6 engineering domains (CAD, CAE, CAM, BIM, EDA, and 3D visualization) and 26 professional software platforms, with both GUI and CLI interfaces and 6 task types ranging from software-selection to open-ended tasks. We further introduce an artifact-centric evaluation methodology built on a unified domain-verifier suite, which programmatically checks the geometric validity, physical feasibility, and rule compliance of final and intermediate artifacts, and scores quantitative design tasks continuously by specification attainment rather than binary success. Evaluation of seven frontier models reveals a substantial capability gap: the strongest model achieves an EngiScore of only 44.3, and just 3.6% of multi-software attempts succeed. EngiWorld provides the first rigorous foundation for measuring progress toward agents that operate professional engineering software end to end.

Decoder-only transformers are trained only through a terminal next-token prediction loss, yet this loss constrains every intermediate hidden state through the fixed downstream computation. We formalize this constraint by studying layerwise loss-to-go functions: the terminal loss obtained by continuing a candidate hidden state through the remaining transformer blocks. Around successful validation trajectories, we show that the local second-order geometry of these functions is governed, up to low-loss residual terms, by a pullback Fisher operator on hidden-state space. Its spectrum identifies output-sensitive directions and approximately prediction-null directions, yielding a local observable subspace of the residual stream. For causal transformers, the same geometry induces a tokenwise curvature score: a Fisher-weighted sensitivity of the target logits to perturbations of each token's hidden state. This score vanishes outside the causal ancestor set of the target and is controlled by downstream Jacobian couplings, making it a loss-aware alternative to attention magnitude. We estimate these quantities using matrix-free Jacobian-vector and vector-Jacobian products and evaluate them across decoder-only language models on WikiText, OpenWebText, and FineWeb. Empirically, the induced geometry predicts perturbation sensitivity, supports nonuniform layerwise rank allocation, yields competitive structured token-pruning signals, and improves low-rank student recovery when added to stronger autoregressive distillation objectives such as reverse KL and skew KL. These results support a predictive-geometric view of transformer computation: near successful trajectories, the terminal loss induces a thin, anisotropic set of output-relevant hidden-state directions that can be measured and exploited for compression and distillation.

We introduce Context Language Models (CLMs), language models that natively manage their own context. We implement this by treating the context as a file and allowing the model to make unrestricted updates to this file. This allows the model to learn what is most important to maintain in context, and naturally extends to multi-agent systems where multiple agent contexts coexist as files. Building CLMs zero-shot with existing models outperforms SOTA context management strategies across a variety of tasks: 11.4% higher accuracy with 21.5% fewer FLOPs on BrowseComp-Plus, 5% higher scores with 59% fewer FLOPs on 12-hour EdgeBench, and 65% greater improvement with the same compute on a 24-hour multi-repository agent-swarm task. Moreover, by shifting context management from external harness control to intrinsic model behavior, CLMs naturally enable both in-context and parametric learning of context-management strategies. We show that CLMs can be steered with natural-language instructions evolved through a standard skill-optimization loop, improving held-out accuracy by up to 35.9 points on a context-management task while reducing compute. We also introduce an online reinforcement learning method for CLMs, improving Qwen3.5-9B performance on BrowseComp-Plus by 47.6% while using 12% fewer FLOPs. Finally, we co-design Suffix Cache Reuse for CLM serving, further reducing server-side compute by 35% relative to standard SGLang at matched performance.

Certo is a small non-generative decision model (Qwen3-4B): it scores candidate actions from their text and returns a probability, instead of generating an answer. The accurate design reads the state, the rules, and each candidate together (a joint scorer), so cost grows with the menu. Independent encoding lets each candidate be encoded once and reused across states (about 5x cheaper at 77 candidates), but separates state from candidate. We ask how much rule-sensitivity survives that move, and whether it can be trained back. Four experiments on Certo: (1) the tested conversion to cacheable scoring loses rule-sensitivity (recall@1 1.00 -> 0.24) while the joint scorer holds 1.00, and a shortlist+rerank rescue fails; (2) targeted counterfactual supervision restores strong performance on held-out synthetic rule tasks (paraphrase, counterfactual, composition; reproducible across seeds), though we do not isolate whether predictions depend on the supplied rule; (3) on real rules the added benefit is not established -- after fixing a truncation confound, the joint scorer wins significantly on the short tier (0.861 vs 0.500) and directionally on the hard tier (0.655 vs 0.483, n=29); (4) a matched cross-domain real-prose mixture did not help and reduced contract accuracy (-9.3, -16.2 points). A cacheable encoder can be made rule-sensitive on its training distribution, but transfer to unseen-source real rules is not established; the joint scorer keeps an edge at the cost of caching.

Many localized large language model (LLM) unlearning methods select a small parameter subset from a localization signal and keep it fixed during optimization. The parameters most associated with a target, however, need not be the best ones to update, and candidate interventions can change value as optimization proceeds. In a controlled experiment, a storage-localization score reaches an area under the receiver operating characteristic curve (AUROC) of 0.981, yet storage identity agrees with the better intervention on only 17/36 targets, while low-rank adaptation (LoRA) wins 35/36. We introduce Intervention Score, which ranks editable groups by the predicted effect of the actual unlearning update while accounting for collateral damage, and use it to form the static intervention-value baseline (Static-IV). We then introduce selective dynamic intervention re-ranking (DIR-R), which revisits that subset only when a calibrated probe justifies the comparison. On the Natural-TOFU dataset, our method has positive descriptive margins in 19/20 comparisons between methods and objectives, although several are near zero. On the LACUNA localization-precision benchmark, our mean terminal utility is higher in all six negative preference optimization (NPO) and SimNPO comparisons: NPO margins range from +0.431 to +0.848, and SimNPO margins range from +0.503 to +0.571. The gradient-difference (GradDiff) objective reveals substantial field dependence. Relative to Static-IV, the primary four-field GradDiff evaluation has six wins, six ties, and no losses, with mean and median paired gains of +0.165 and +0.0025. The evidence supports separating localization, initial intervention selection, and checkpoint-dependent support revision.

Reinforcement Learning with Verifiable Rewards (RLVR) methods such as GRPO rely on successful self-generated trajectories, but finite rollout budgets can produce all-fail groups with no reward-based policy-gradient signal. While additional rollouts improve the chance of success at higher cost, successful trajectories missing from one model's rollouts may already have been discovered by another. Indeed, we observe that heterogeneous models often succeed on complementary prompts, creating opportunities for mutual learning without a designated stronger teacher. To exploit this complementarity, we propose GRAFT (Gated Replacement of Answer-Failed groups with peer Trajectories), an off-policy-aware framework that replaces all-fail groups with informative peer groups. GRAFT transfers both successful and unsuccessful peer responses with peer-computed advantages, while controlling cross-model mismatch through sequence-level compatibility weighting and token-level importance ratio clipping. Across three heterogeneous model pairs and five mathematical reasoning benchmarks, GRAFT consistently improves both models over GRPO with the same per-model rollout budget, gaining 2.1 points on average and up to 4.5 points in model-level average performance. Stored peer trajectories preserve most of the gains, improving over GRPO by 1.8 points on average without simultaneous co-training.

On-policy self-distillation (OPSD) trains a student to match a privileged teacher distribution along its own sampled trajectory. Standard OPSD applies this supervision to unverified student rollouts while conditioning the teacher on privileged context, typically a reference solution. We separate these roles in a factorial analysis and find that scaffold correctness has a stronger effect on downstream accuracy than context correctness. Unverified scaffolds create an imitation gap because the teacher can use information unavailable to the student. This gap shrinks with model scale, yet OPSD continues to supervise mostly unverified trajectories. In contrast, verified scaffolds remain effective even when the teacher is conditioned on the student's own unsuccessful rollout. Based on this finding, we introduce OASIS, which retains the OPSD objective but supervises mostly verified by label on-policy trajectories and replaces written solutions with unverified model-generated attempts as the teacher context. OASIS therefore requires only final-answer labels. Across Qwen3-1.7B, 4B, and 8B on AIME 2024, AIME 2025, and HMMT 2025, OASIS improves over the base model by 3.2--3.8 points on average, while OPSD's gain falls from 3.05 points at 1.7B to 0.14 at 8B. At 8B, OASIS improves over OPSD by 3.05 points, showing that verified on-policy scaffolds preserve the effectiveness of self-distillation as models scale.

Recent work on anchored diffusion language models improves denoising by shaping an intermediate latent space with supervised important-token targets. In this work, we introduce time-based (self-supervised) anchoring, which learns and reuses latent anchors without requiring such targets. Our key observation is that anchors encode persistent properties of the clean sequence, such as its semantic intent, global structure, or intermediate plan. Although their hidden representations become stale as the token canvas evolves, their semantic content remains useful across nearby diffusion times. This is implemented through a two-stage architecture consisting of a relatively expensive anchor network that generates the latent cache state and a lightweight denoising network that intelligently combines the cached latent state with the current state at each reverse step using a fusion module. This gives anchoring a latent-space caching interpretation: the anchor network is evaluated periodically, while its cached representation is reused across multiple reverse steps. We instantiate this framework as TADM:Post-train, which time-anchorizes pretrained DLMs, and TADM:Pretraining, which learns time-based anchors during pretraining. Applied to DiffusionGemma-26B, TADM:Post-train improves throughput by approximately 49% to 79% on several math, code, and STEM benchmarks (GSM8K, AIME26, GPQA-Diamond, LiveCodeBench-v6, HumanEval, MMLU-Pro). TADM:Pretraining reduces Transformer-layer computation by up to 38% relative to a standard single-stage DLM, achieves up to 73% higher measured throughput than ADLM.

Advances in the coding capabilities of LLM agents allow them to inspect and modify their own instructions, tools, and execution procedures. Existing approaches use this ability to search for improved agents through repeated downstream evaluation, which incurs substantial costs and ties the search to the evaluated tasks. We introduce \textbf{SelfSearch}, a reward-free search procedure in which agents modify themselves using records of previous self-improvement episodes. These records capture the reasoning, tool actions, and outcomes of earlier modification attempts, providing concrete experience for improving both task solving and self-modification. Without downstream reward signals during search, SelfSearch improves population-mean success over the initial agent in all six model--benchmark settings, with individual agents gaining up to 11.2 percentage points on Terminal-Bench 2.1. On SWE-bench Multilingual, an agent improves success by \textbf{5.0} percentage points while reducing execution cost by \textbf{38.5}\% on tasks solved by both the initial and evolved agents. SelfSearch achieves competitive task success with evaluation-guided search baselines at lower search cost. With only \textbf{\$4.03} in search cost, it produces a harness that solves \textbf{82.0}\% of Terminal-Bench 2.1 tasks with DeepSeek V4 Flash under the settings of a public nine-harness comparison, matching the top-scoring harness, Codex. These results suggest that experience gained through self-modification can improve agents' downstream capabilities and efficiency.

Masked diffusion models (MDMs) generate text by unmasking several tokens per step, but they are trained and sampled under different conditions. The model is trained on randomly masked sequences, whereas inference follows a trajectory shaped by the model's own predictions. Additionally, each step has no access to what the previous one computed. Recent methods narrow these limitations from separate angles, leaving open how these choices interact. We introduce PUMBA, a unified framework for trajectory-aware training that trains the denoiser on consecutive steps of policy-induced trajectories, passes information between steps, and optimizes them jointly by backpropagation through time. A controlled study of this design space shows that i) exact train--inference alignment fails due to local overfitting, whereas a looser alignment still brings training masks closer to those seen at inference; ii) passing continuous information outperforms discrete gradient estimators through the commitment at each step; and iii) performance improves as backpropagation through time spans more steps, which we support theoretically. Combined, these components match the best checkpoint of a same-size autoregressive model. Building on these findings, we scale PUMBA to supervised fine-tuning of LLaDA-8B, where it improves the trade-off between performance and number of function evaluations (NFEs) in both full-canvas and block diffusion generation. At matched performance, it needs up to 22% fewer NFEs than standard fine-tuning with twice the budget in full-canvas generation, and up to 26% fewer than standard fine-tuning for the same number of steps in block diffusion.

LLMs trained with Chain-of-thought excel in reasoning capability, but often come with excessive token cost. We find that the core of reasoning capacity lies in the Thinking model's weight component within the null space of a projection defined by the corresponding Non-thinking model's dominant singular directions, and removing the subspace component can largely improve reasoning efficiency without hurting the accuracy gained during thinking-mode post-training. Unlike existing efforts that mostly operate within the dominant subspace, we are the first to unveil the critical role of the null space and harness it for model optimization. Motivated by this finding, we propose Spectral Null-Space Swap ($S^3$), a training-free composition of paired Non-thinking and Thinking checkpoints. Our method keeps the Non-thinking model inside its own dominant subspace and takes the Thinking checkpoint outside it, improving reasoning efficiency while maintaining accuracy. We extensively evaluate $S^3$ on 2B-30B dense and mixture-of-experts (MoE) architectures spanning 28 evaluation environments across mathematical, multimodal, and audio reasoning domains. $S^3$ establishes new empirical Pareto Frontiers among training-free model composition strategies: across all settings, it reduces inference token overhead by an average of 27.4% compared to full Thinking models while simultaneously improving overall task accuracy by 1.0 percentage point (e.g., yielding +8.3% accuracy on HMMT25 alongside a 33.0% token speedup). We further use attention entropy for explanation and find that the retained component produces more concentrated attention, and we use a simplified analytical model about optimization to demonstrate why null-space can effectively reduce attention entropy, thereby improving the efficiency of reasoning.

On-policy distillation (OPD) trains a student on its own generated responses using dense, token-level supervision from a stronger teacher. Vanilla OPD treats all teacher signals equally, assuming that the teacher's supervision is equally important for every token. However, teacher signals at different tokens may have very different effects on the student's performance: some correct important reasoning errors, while others have little effect on the final answer. Motivated by this observation, we introduce Dr. OPD (OPD Done Right), which defines the optimal weighted OPD to maximize the student's performance. We formulate Dr. OPD as a bilevel optimization problem in which the student learns from weighted teacher supervision, while the weights are selected to maximize the expected reward of the resulting student. To solve Dr. OPD, we develop an efficient iterative solver that updates the token weights and student policy alternatively. At each round, it updates weights in closed form and then takes one gradient step on the resulting weighted OPD objective. Under regularity conditions, we show that this weighted update achieves a higher expected reward than a vanilla OPD update. Empirically, across strong-to-weak and same-size distillation on math and code, Dr. OPD consistently outperforms all evaluated baselines. In particular, in the strong-to-weak distillation setting, Dr. OPD improves average math performance by $9.7$ points over vanilla OPD, and enables the smaller student to surpass its larger teacher.

Turning decoder-only large language models (LLMs) into strong dense retrievers typically requires some form of retriever training. In this paper, we ask whether LLMs can instead be prompted to produce effective representations for dense retrieval given only a few in-context examples. To answer this, we introduce RICE (Representations from In-Context Examples), a simple "training-free" approach that extracts high-quality dense representations from LLMs. To do so, RICE conditions the LLM on examples that provide a shared context for query and document encoding. Our results demonstrate that RICE embeddings can substantially improve the accuracy of prompt-based LLM embeddings, establishing it as a simple method to build LLM-based dense retrievers that do not require training. We release our code at https://github.com/nourj98/RICE.

Speech-LLMs are expensive to run, making compression important for real-world deployment. However, compressed models are usually selected using aggregate word error rate (WER), which can hide how pruning affects different demographic groups. In this work, we systematically study the effect of audio encoder pruning on SLAM-ASR for different demographic groups. Using the Fair-Speech and Common Voice datasets, we found that the pruning does not affect all demographic groups equally; the gap between best- and worst-performing groups increases in fold. These disparities appear across all three encoder scales, but only the largest model initially hides them behind aggregate WER. LoRA adaptation improves WER for every group, but benefits groups already performing well more strongly and widens for certain groups. On Common Voice English, Danish, and Dutch, accent gaps persist but do not clearly widen, showing that the fairness effects of pruning vary across datasets and must be measured directly. Our findings suggest that for pruned models, deployment decisions should include per-group WER, with the worst-performing group's error rate as an explicit criterion.

A small trainable advisor can steer a frozen language-model executor using natural-language advice. In addition to learning from task rewards, the advisor can use feedback from completed interactions to improve its advice. However, a plausible correction need not change execution, yet learning from such corrections can still affect the advisor's future decisions in other contexts. In a shared-parameter model, we prove that such corrections can limit learning if their targets favor useful advice less strongly than those of other corrections. Keeping them less often than the rest improves the model's eventual performance compared to learning from every correction. Motivated by this, our method, Advisor Self-Distillation (AdviSD), pairs outcome-based reinforcement learning with self-distillation from a feedback-conditioned copy of the advisor selectively. Reflection proposes corrections, and the advisor scores the same recorded executor response with and without its issued advice, using the magnitude of the difference to select decisions for supervision. This approach does not require executor likelihoods or additional executor rollouts. Experiments with Qwen3-8B advisors for Gemini and Claude show that AdviSD outperforms advisor-GRPO by 4.2-6.4 percentage points on BFCL-v3 and by 3.9-5.1 score points on EnvScaler. The trained advisors generalize to out-of-domain tasks and transfer across different executor versions and model families. AdviSD also beats matched-count random selection, supporting the value of its selection rule.

Agent performance depends on both reasoning ability and the environment in which it acts. We study test-time AI-for-AI, asking how a Builder can learn to construct better execution environments for a Target while both models' weights remain fixed. To make the Builder's experience reusable, we introduce Meta-Skill: principles specifying when support is needed and what resources to provide. The Builder learns these principles from Target's execution feedback on the development set, then uses the frozen skill bank to construct harnesses for unseen tasks. Across Harness-Bench and NewtonBench, full-bank meta-skills improve macro-average performance by 8.95 percentage points over no-skill construction, and 12.02 points over direct delivery of the same bank to the Target. These results highlight the value of translating experience into executable support. Gains when the same model serves both roles further suggest a path to system level self-improvement through learning to build better environments.

Answering questions about long videos often requires connecting events involving the same objects across hours or days. Chronological descriptions and text-derived entities can leave physical identity unresolved: different objects may share a description, while observations of the same object remain disconnected across events. Retrieving relevant events therefore does not necessarily recover the "biography" of the particular entity a question concerns. To address this, we introduce Grounded Entity Biographies (GEB), a long-video memory framework that groups visually grounded observations of the same physical instance across clips into retrievable biographies while preserving the context of each moment. During question answering, the biography is retrieved alongside episodic evidence, allowing the model to follow an entity through events using identity links established during memory construction. Evaluations across four benchmarks, including day-long and week-long recordings, demonstrate improvements over prior memory frameworks in both multiple-choice and open-ended question answering. On EgoLifeQA, GEB achieves 72.0% accuracy, 4.4 percentage points above the best published result. Ablations show that grounded identity association and biography reading both contribute to the gains, which additional descriptions alone do not fully recover.

Emotion-conditioned text-to-speech (TTS) models may fail to express the requested emotion reliably, and improving controllability by additional training is costly in both computation and emotion-labeled speech training data. We therefore study vector steering, a training-free approach that modifies the internal representations of a frozen model. CoCoEmo, a conventional vector steering method for emotion TTS, treats each emotion vector as an indivisible direction controlled by a single global strength, limiting adherence to the requested emotion. In this work, we first discover that an emotion vector can be decomposed into a shared component that moves speech away from neutral expression and a residual component that directs generation toward the requested emotion. Building on this finding, we propose Emotion Residual-Enhanced Steering for TTS (EmoRES), a novel method that controls the two components without retraining the backbone. On IEMOCAP, EmoRES outperforms CoCoEmo across all four objective emotion metrics on the IndexTTS-2 and CosyVoice2 backbones. Rank correlation improves by 26.13 and 12.97 percentage points, corresponding to relative gains of 118.8% and 33.1%, while emotion hit rate improves by 12.95 and 6.92 points, corresponding to relative gains of 20.1% and 9.8%. Human evaluation further shows a relative improvement up to 35.0% in the rate at which listeners correctly identified the dominant requested emotion and up to a 17.3% improvement in fidelity, while listeners prefer EmoRES for naturalness in up to 63.8% of pairwise comparisons. Component ablations further demonstrate that effective control benefits from preserving the shared component while strengthening the residual of the emotion steering vectors.

Reasoning about the 3D world from multi-view images remains a fundamental challenge for Multimodal Large Language Models (MLLMs). While modern MLLMs handle single-image inputs effectively, they struggle to integrate evidence across viewpoints into a coherent 3D understanding. A growing body of work attempts to close this gap by injecting 3D awareness into MLLMs, either by boosting fine-grained pixel-level cross-view correspondence or by fusing features from 3D geometry foundation models, yet a substantial gap to human reasoning persists. In this work, we revisit human spatial reasoning, which suggests that rather than relying on fine-grained geometry cues, humans roughly identify common objects across views, infer the relative geometry between viewpoints, and assemble a coarse 3D layout of the scene. Inspired by this process, we introduce Imagine3D-LLM, an MLLM that learns to assemble a similar compact 3D representation of the scene and conditions its answer on this representation. Concretely, we append a small set of learnable summary tokens after the image tokens, decode them into a compact 3D Gaussian Splatting representation supervised by a photometric reconstruction loss, and train jointly with the standard next-token prediction objective. Notably, although only the summary tokens receive direct reconstruction supervision, this objective also induces stronger cross-frame correspondence within the LLM's underlying image features, suggesting that learning to reconstruct propagates 3D-aware signals throughout the model. As a result, Imagine3D-LLM consistently outperforms prior approaches across multiple spatial reasoning and 3D understanding benchmarks, suggesting that imagining the scene can be more effective than being told its pixel-wise geometry.

The advancement of LLMs and their accessibility have triggered renewed interest in multi-agent reinforcement learning as robust and adaptive frameworks for dynamically changing environments. This paper introduces \texttt{RL-Focal}, a two-stage RL agent framework that routes and ensembles LLMs. \textit{First}, we develop the Decider RL-agent, which learns to dynamically select an ensemble of small size ($m_i$) among $N$ LLMs ($m_i \ll N$) for incoming queries from a user-defined downstream task $i$, by maximizing both error-diversity and reasoning-performance of the selected ensemble through iterative updates of task-adaptive rewards and policy. \textit{Second}, to enable effective fusion of dynamically selected LLMs, we develop the stage-2 Fusion RL-agent, which learns to resolve reasoning conflicts from different LLMs and dynamically adapt to different ensemble teams composed by the Decider Agent for different downstream tasks. {\em Third}, we introduce the focal diversity metric to better model the error correlations among multiple LLMs further improving the generalization performance of the Decider Agent, which actively prunes the ensemble combinations. By focal diversity, we enhance performance across tasks by effectively promoting reward-aware and policy-adaptive ensemble selection and inference fusion. Extensive evaluations on five benchmarks show that RL-Focal achieves the performance improvement of 8.48\% with an ensemble of small size compared to the best individual LLM in a pool and offers stronger robustness. Code is available \href{https://github.com/git-disl/RL-Focal}{here}.

Large language models (LLMs) often exhibit sycophantic behaviors -- such as excessive agreement with or flattery of the user -- but it is unclear whether these behaviors arise from a single mechanism or multiple distinct processes. We decompose sycophancy into sycophantic agreement and sycophantic praise, contrasting both with genuine agreement. Using difference-in-means directions, activation additions, and subspace geometry across multiple models and datasets, we show that: (1) the three behaviors are encoded along distinct linear directions in latent space; (2) each behavior can be independently amplified or suppressed without affecting the others; and (3) their representational structure is consistent across model families and scales. These results suggest that sycophantic behaviors correspond to distinct, independently steerable representations.

Recent advancements have endowed Large Language Models with impressive general reasoning capabilities. However, these reasoning models often perform worse than non-reasoning models on personalization tasks. While some methods use outcome-based RL to improve personalization reasoning, they fail to supervise the reasoning process. As a result, models may reach correct answers through flawed reasoning chains, limiting further improvement. To address this, we propose TagPR, a novel framework that adds semantic tags to the reasoning process for step-by-step guidance. TagPR first automatically generates a structured, tagged dataset for Supervised Fine-Tuning. It then employs a multi-stage RL process guided by a composite reward signal, which integrates tag-based process supervision with a novel Personalization Reward Model with User Embeddings to achieve fine-grained alignment with user-specific logic. Extensive experiments on public LaMP, LongLaMP, PGraphRAG, and a self-constructed dataset demonstrate that our approach achieves state-of-the-art results, delivering an average improvement of 32.65% over the base model across all LaMP benchmark tasks. Our work demonstrates that tag-guided process supervision is an effective approach for personalization reasoning.

Recent advances in text-to-image (T2I) generation have achieved impressive results, yet existing models often struggle with simple or underspecified user prompts due to a distributional gap with their descriptive training captions. This frequently leads to suboptimal image-text alignment, aesthetics, and overall visual quality. To bridge this gap, we propose POET (\textbf{P}reference \textbf{O}ptimization for \textbf{E}nhanced \textbf{T}ext-to-Image generation), an automated prompt rewriting framework that leverages large language models (LLMs) to refine user inputs before feeding them into frozen T2I backbones. POET introduces a carefully designed composite reward system and an iterative Direct Preference Optimization (DPO) training pipeline, enabling the rewriter to learn model-preferred prompt structures directly from multimodal feedback without requiring costly high-quality supervised fine-tuning (SFT) data. Extensive evaluations across diverse T2I models and benchmarks show that our prompt rewriter consistently improves image-text alignment, visual quality, and aesthetics, outperforming strong baselines. Furthermore, we demonstrate strong transferability by showing that a rewriter trained on one T2I backbone generalizes effectively to others without needing to be retrained. These findings highlight that POET is an effective, robust, and practical model-agnostic strategy for improving T2I systems.

In-context learning is governed by both temporal and semantic relationships, shaping how Large Language Models (LLMs) retrieve contextual information. Analogous to human episodic memory, where the retrieval of specific events is enabled by separating events that happened at different times, this work probes the ability of various pretrained LLMs, including transformer and state-space models, to differentiate and retrieve temporally separated events. Specifically, we prompted models with sequences containing multiple presentations of the same token, which reappears at the sequence end. By fixing the positions of these repeated tokens and permuting all others, we removed semantic confounds and isolated temporal effects on next-token prediction. Across diverse sequences, models consistently placed the highest probabilities on tokens following a repeated token, but with a notable bias for those nearest the beginning or end of the input. An ablation experiment linked this phenomenon in transformers to induction heads. Extending the analysis to unique semantic contexts with partial overlap further demonstrated that memories embedded in the middle of a prompt are retrieved less reliably. Despite architectural differences, state-space and transformer models showed comparable temporal biases. Our findings deepen the understanding of temporal biases in in-context learning and offer an illustration of how these biases can enable temporal separation and episodic retrieval.

Temporal reasoning over evolving semi-structured tables poses a challenge to current QA systems. We propose an approach that recasts the task as automated knowledge base construction: (1) prompting an LLM to synthesize a 3NF-compliant relational schema from Wikipedia infobox timelines, (2) populating the schema to obtain a queryable database, and (3) generating and executing SQL queries against it, with QA accuracy serving as an extrinsic evaluation of the constructed knowledge base. In a controlled grid of three schema generators crossed with six query models, the schema source accounts for 79.5% of the exact match (EM) variance against 1.6% for the query model: replacing the schema, and the prompt scaffolding derived from it, shifts EM by 14.7 to 20.0 points, whereas replacing the query model under a fixed schema shifts it by 4.4 to 12.1. From this evidence, we distill three candidate schema-design principles: balanced normalization, semantic naming, and consistent temporal anchoring, framed as correlational hypotheses. Our best configuration (Gemini 2.5 Flash schemas + Gemini-2.0-Flash queries) reaches 80.39 EM, 11.5 points above the strongest reported baseline (68.89 EM); an open-weights configuration reaches 79.52.

Trustworthy clinical summarization requires every claim to be traceable to its evidence, yet existing attribution often resolves only to the sentence or document, leaving clinicians to scan surrounding text for the few words that matter. We argue that the unit of attribution should match the unit of verification: the precise phrase the reader's eye must land on. We present MAPLE (Medical Aspect-Based Summarization with Phrase-Level Evidence), a human-annotated benchmark that grounds each summarized claim in both cited sentences and contributory phrases within them. Spanning 152 randomized controlled trial (RCT) abstracts and 16 clinically motivated aspects, MAPLE comprises 1,799 aspect-based summaries with two-level evidence. We further introduce a decoupled evaluation framework that separately scores content, traceability, and locatability, together with a proxy for the amount of source text a clinician must inspect to verify a claim. Benchmarking eleven LLMs shows that sentence-level citation is consistently strong (C-F1 up to 90.9%), while phrase-level grounding remains less stable and the most discriminative axis across models (P-F1 66.1-84.5%). These results suggest that the key challenge is not only producing accurate summaries, but localizing their supporting evidence precisely enough for efficient clinical verification. Data and code are available at https://github.com/chubohao/maple.

Large language models (LLMs) excel at semantic understanding, yet their ability to reconstruct internal structure from scrambled inputs remains underexplored. Sentence-level restoration is difficult to evaluate automatically because scrambled sentences often admit multiple valid reorderings. We introduce OrderProbe, a deterministic benchmark for structural reconstruction using fixed four-character expressions in Chinese, Japanese, and Korean, which have a unique canonical order and thus support exact-match scoring. We further propose a diagnostic framework that evaluates models beyond recovery accuracy, including Semantic Accuracy, Logical Validity, Structural Consistency, Robustness, and Information Density. Experiments on twelve widely used LLMs show that structural reconstruction remains difficult even for frontier systems: zero-shot recovery frequently falls below 35%. We also observe a consistent gap between meaning-oriented generation and exact structural reconstruction, suggesting that structural robustness is not an automatic byproduct of semantic competence.

Large language models have made substantial progress in mathematical reasoning. However, benchmark development for multilingual evaluation has lagged behind English in both difficulty and recency. Recently, GSM-Symbolic showed a strong evidence of high variance when models are evaluated on different instantiations of the same question; however, the evaluation was conducted only in English. In this paper, we introduce MGSM-Pro, an extension of MGSM dataset with GSM-Symbolic approach. Our dataset provides five instantiations per MGSM question by varying names, digits and irrelevant context. Evaluations across nine languages reveal that many low-resource languages suffer large performance drops when tested on digit instantiations different from those in the original test set. We further find that models robustness in HRL setting do not necessarily translate to LRL. Moreover, proprietary models, such as Gemini 2.5 Flash and GPT-4.1 are less robust to digit, whereas Gemini 3.0 Pro is more robust. Among open models, GPT-OSS 120B and DeepSeek v3 show stronger robustness. Based on these findings, we recommend evaluating each problem using at least five digit-varying instantiations to obtain a more robust and realistic assessment of math reasoning.

Text-to-SQL is a key natural language processing task that maps natural language questions to SQL queries, enabling intuitive interaction with web-based databases. Although current methods perform well on benchmarks like BIRD and Spider, they struggle with complex reasoning, domain knowledge, and hypothetical queries, and remain costly in enterprise deployment. To address these issues, we propose a framework named IESR(Information Enhanced Structured Reasoning) for lightweight large language models: (i) leverages LLMs for key information understanding and schema linking, and decoupling mathematical computation and SQL generation, (ii) integrates a multi-path reasoning mechanism based on Monte Carlo Tree Search (MCTS) with majority voting, and (iii) introduces a trajectory consistency verification module with a discriminator model to ensure accuracy and consistency. Experimental results demonstrate that IESR achieves state-of-the-art performance on the complex reasoning benchmark LogicCat (24.28 EX) and the Archer dataset (37.28 EX) using only compact lightweight models without fine-tuning. Furthermore, our analysis reveals that current coder models exhibit notable biases and deficiencies in physical knowledge, mathematical computation, and common-sense reasoning, highlighting important directions for future research. We released code at https://github.com/Ffunkytao/IESR-SLM.

As people turn to LLMs for social advice, understanding their behavior in such contexts becomes essential. In this work, we focus on behavioral dispositions: the underlying tendencies that shape responses in social contexts. We introduce STAR, a framework for studying how closely the dispositions expressed by LLMs align with those of humans. STAR builds on established psychological questionnaires, adapting their items into realistic advice-seeking scenarios, as self-report may not transfer to actual advisory behavior. Using STAR, we construct a dataset of 23k scenarios, each validated by 3 raters and annotated with preferences from 10 participants. Across 25 LLMs, we find that (1) when human consensus is high, frontier models can fail to reflect it in 15-20% of cases, and smaller models fail at substantially higher rates; (2) when humans disagree, LLM recommendations are substantially less diverse than human choices, both within individual models and even across models from different providers, potentially narrowing the range of options users are guided toward; (3) LLMs' self-reported values are poor predictors of their recommendations. To support future research we make our dataset and code publicly available.

Accurate confidence estimation is critical for reliable use of large language models (LLMs). Prior work on LLM calibration largely focuses on response-level confidence, which estimates the correctness of a single generated output. However, this formulation is misaligned with many practical settings where the central question is how likely a model is to solve a query overall. We show that this mismatch results from the stochastic nature of modern LLM decoding, under which single-response correctness fails to reflect underlying model capability. To address this issue, we introduce capability calibration, a new evaluation framework for measuring how well query-level confidence aligns with a model's expected accuracy on individual queries. We formally distinguish capability calibration (CC) from response calibration (RC) and show that the two differ both theoretically and empirically. We further show that CC is better suited than RC to applications like pass@k prediction and inference budget allocation. Finally, we evaluate common confidence estimation methods to understand the practical feasibility of CC.

Heterogeneous multi-agent systems combine models with different capabilities through a common communication interface. Exchanging internal states directly requires translating between model-specific representations and controlling intermediate computation. We introduce the Vision Wormhole, which repurposes the visual input interface of Vision-Language Models (VLMs) for continuous communication between frozen heterogeneous agents. A Universal Visual Codec encodes each sender's latent rollout into a fixed-size message, maps it through a shared reference space, and decodes received messages into the receiver's image-token span. Per-model codecs and affine reference maps form a hub-and-spoke architecture with $O(N)$ components for $N$ models. Each model learns its codec independently through self-distillation on anchor texts, and shared-anchor alignment enables reuse across communication partners. Across four VLM families, six team configurations, and nine reasoning benchmarks, Vision Wormhole improves accuracy by 6.0 percentage points on average over text-mediated MAS and achieves a 1.69$\times$ geometric-mean speedup in batch-normalized end-to-end runtime.

This study investigates the diverse characteristics of nouns, focusing on both semantic (e.g., countable/uncountable) and morphosyntactic (e.g., masculine/feminine) distinctions. We explore inter-word variations for gender markers in noun morphology. Grammatical gender shift is a widespread phenomenon in languages around the world. The aim is to uncover the underlying patterns governing the variation of lexemes. To this end, we propose a new computational component dedicated to pairing items with morphological templates (e.g., the result of a generated item-template pair: (funas, $\{N, +SG, -PL, -M, +F, -COL, +SING\}$), with its spell-out form: $\eth$a-funast 'cow'). This process is formally represented by the Template-Based and Modular Cognitive model. This proposed model, defined by a set-valued set function $h : \mathscr{P}(M) \rightarrow \mathscr{P}(M)$, predicts the nonlinear dynamic mapping of lexical items onto morphological templates. By applying this formalism, we present a unified framework for understanding the complexities of morphological markings across languages. Through empirical observations, we demonstrate how these shifts, as well as non-gender shifts, arise during lexical changes, especially in Riffian. Our model posits that these variant markings emerge due to template shifts occurring during word and meaning formation. This study achieves two primary objectives. First, on the formal side, we prove the model's representational completeness in learning and prediction. Second, on the linguistic side, we challenge and broaden the conventional view of word formation by formally demonstrating that conversion is applicable to noun-to-noun derivation. This data-driven mathematical model not only contributes to a deeper understanding of morphosyntactic variation but also offers potential applications in other fields requiring precise modelling of linguistic patterns.

While different stakeholders are trying to leverage Arabic Language Models (ALMs), safety alignment in ALMs remains largely underexplored, hindering their mainstream adoption. Existing safety benchmarks are predominantly English-centric and evaluate Arabic only in its standardized form, obscuring fine-grained safety vulnerabilities in Arabic NLP systems. This paper introduces SalamahBench, a unified benchmark of 8{,}270 human-verified harmful prompts across ML Commons hazard categories, each rendered in Modern Standard Arabic (MSA) and five regional Arabic varieties, namely Egyptian, Syrian, Saudi, Lebanese, and Moroccan, for a total of 49{,}620 paired instances. To analyze the resulting data, we introduce two complementary metrics, namely Dialect Shift, which measures a model's aggregate change in safety under dialectal reformulation, and Category-Specific Dialect Deviation, which isolates harm categories whose change departs from that aggregate trend. Evaluating models such as Fanar 2, ALLaM 2, and Karnak 1 under multiple safeguard configurations, we find that cross-variety robustness is strongly model dependent, and that aggregate scores can conceal category-level divergence. Our findings highlight the necessity of evaluating Arabic model safety jointly across linguistic varieties and harm domains rather than relying on aggregate scores or MSA alone.

Large language model (LLM)-based judges are widely adopted for automated evaluation and reward modeling, yet their judgments are often affected by judgment biases. Accurately evaluating these biases is essential for ensuring the reliability of LLM-based judges. However, existing studies typically investigate limited biases under a single judge formulation, either generative or discriminative, lacking a comprehensive evaluation. To bridge this gap, we propose JudgeBiasBench, a benchmark for systematically quantifying biases in LLM-based judges. JudgeBiasBench defines a taxonomy of judgment biases across 4 dimensions, and constructs bias-augmented evaluation instances through a controlled bias injection pipeline, covering 12 representative bias types. We conduct extensive experiments across both generative and discriminative judges, revealing that current judges exhibit significant and diverse bias patterns that often compromise the reliability of automated evaluation. To mitigate judgment bias, we propose bias-aware training that explicitly incorporates bias-related attributes into the training process, encouraging judges to disentangle task-relevant quality from bias-correlated cues. By adopting reinforcement learning for generative judges and contrastive learning for discriminative judges, our methods effectively reduce judgment biases while largely preserving general evaluation capability.

SiDiaC-v.2.0 is the largest comprehensive Sinhala Diachronic Corpus to date, covering a period from 1800 CE to 1955 CE in terms of publication dates, and a historical span from the 5th to the 20th century CE in terms of written dates. The corpus consists of 229k words across 185 literary works that underwent thorough filtering, preprocessing, and copyright compliance checks, followed by extensive post-processing. Additionally, a subset of 59 documents totalling 65k words was annotated based on their written dates. Texts from the National Library of Sri Lanka were selected from the SiDiaC-v.1.0 non-filtered list, which was digitised using Google Document AI OCR. This was followed by post-processing to correct formatting issues, address code-mixing, include special tokens, and fix malformed tokens. The construction of SiDiaC-v.2.0 was informed by practices from other corpora, such as FarPaHC, SiDiaC-v.1.0, and CCOHA. This was particularly relevant for syntactic annotation and text normalisation strategies, given the shared characteristics of low-resource language status between Faroese and the similar cleaning strategies utilised in CCOHA. This corpus is categorised into two layers based on genres: primary and secondary. The primary categorisation is binary, assigning each book to either Non-Fiction or Fiction. The secondary categorisation is more detailed, grouping texts under specific genres such as Religious, History, Poetry, Language, and Medical. Despite facing challenges due to limited resources, SiDiaC-v.2.0 serves as a comprehensive resource for Sinhala NLP, building upon the work previously done in SiDiaC-v.1.0.

Understanding and evaluating multi-step reasoning in LLMs at the level of individual steps remains a key challenge. Process reward models (PRMs) provide a solution by scoring each step, enabling fine-grained supervision and improved reliability. However, training them requires costly human annotation or computationally intensive rollout-based labeling. To solve this, we introduce MCNIG, a scalable method for automatically labeling the quality of individual reasoning steps in any verifiable domain. Its step score, net information gain (NetIG), improves upon single-reference information gain (IG) by comparing the most-supported correct answer against the most-supported incorrect one, yielding a robust signal even for long and structured outputs like code and SQL, where IG fails. We show that the signal produced by MCNIG correlates with human judgments of step quality, and we apply MCNIG labels to train PRMs that achieve the best average best-of-K accuracy across eight benchmarks spanning mathematics, code generation, text-to-SQL, and scientific QA. Crucially, MCNIG generates no rollouts, cutting labeling complexity to O(N) and making it up to X times cheaper than rollout-based methods at comparable label quality, which makes large-scale process supervision practical.

Recent developments in reasoning capabilities have enabled large language models to solve increasingly complex mathematical, symbolic, and logical tasks. Interestingly, while reasoning models are often trained to generate monolingual text, these models have also been observed to code-switch (i.e., mix languages). Prior works have either viewed code-switching as an undesirable error, attempted to control code-switching through modifications to input prompts or the output decoding process, or focus on narrow subsets of languages, domains, tasks, and models. We address these gaps by introducing the first linguistically and behaviorally motivated fine-tuning framework for identifying beneficial code-switched reasoning behaviors in large language models and teaching these models to code-switch more effectively for reasoning. We create the Code-Switched Reasoning (CoRe) corpus, consisting of (1) 7k reasoning traces from 15 models, 18 languages, 10 scripts, and diverse reasoning domains, providing insights into potentially helpful code-switching behaviors, and (2) 40 carefully curated datasets for training and evaluating six interventions for improving code-switching in reasoning across three models and seven languages, totaling 120 fine-tuning conditions. Across 80k+ reasoning traces from both language/culture-agnostic and -specific evaluations, English-dominated reasoning, semantically accurate code-switching, and a more even mix of languages are positively associated with correct answers, whereas dense switching is associated with incorrect answers. Moreover, we are able to elicit positive behaviors through fine-tuning tasks that do not directly demonstrate code-switching. Our work suggests that small but well-curated datasets can change how reasoning models code-switch, allowing us to reap the benefits of reasoning in the many languages that lack large-scale reasoning data.

Long-horizon memory systems increasingly improve how evidence is stored and retrieved, yet the generator must still reason over fragments whose cross-session relationships are implicit. We study generation-time memory organization as a distinct design dimension and introduce GRAVITY (Generation-time Relational Anchoring Via Injected Topological MemorY), a host-independent auxiliary memory layer. GRAVITY consolidates raw dialogue into entity profiles, temporal event traces, and cross-session topic summaries, then retrieves and injects query-relevant records through the prompt interface. Across five heterogeneous memory systems on LongMemEval and LoCoMo, it improves every host--benchmark baseline under two distinct LLM configurations. Controlled analyses separate gains from organizing already available evidence and from consolidating information across the full history. Under a matched LightMem pipeline, the entity--event--topic representation reaches 83.9% on LoCoMo, 3.6% above the strongest of six alternative auxiliary representations. These results show that generation-time structure is a portable complement to existing memory retrieval, while its interaction with host evidence depends on the benchmark and host.

A single base LLM now comes with dozens of post-training variants, quantized, LoRA-adapted, or distilled, and each has to be checked before release. Existing evaluations provide only a partial picture: benchmark scores and likelihood screens say that a variant has degraded, similarity scores such as CKA and SVCCA say how its features moved, and nothing connects the two. We connect them with one structural fact and one design choice: the prediction head is linear, so feature geometry reaches the loss, and we compare the two feature sets through an orthogonal map, which leaves the geometry being measured unchanged. From these we derive PRISM, a closed-form upper bound on the cross-entropy risk gap between a target model and a proxy variant, and prove that it splits exactly into three measurable axes: scale, shape, and head. Each axis names a failure mode and where to intervene: low-bit quantization distorts shape and, at the lowest bit-widths, also inflates activation scale; quantizing the output embedding inflates the head term, which dominates the bound at high bit-widths. Because the shape term is differentiable, the same geometry becomes a regularizer that curbs catastrophic forgetting more than experience replay. Across two model families and five benchmarks, PRISM ranks quantized and fine-tuned variants from a single forward pass at mean Spearman above 0.8, and eight reference sequences already recover the ranking that the risk gap itself needs over a hundred to reach.

Multimodal Large Language Models (MLLMs) exacerbate safety risks by introducing vulnerabilities across multiple modalities, such as language and vision. Current MLLM safety evaluation tools, however, suffer from major limitations: 1) English-centric dataset construction, and 2) a focus on generic risks that are not tied to local cultural contexts. This paper introduces KSAFE-MM, a benchmark for Korean multimodal safety evaluation that covers both general safety risks and culture-specific vulnerabilities. KSAFE-MM consists of two complementary parts: KSAFE-MM-G evaluates globally shared risks in Korean contexts through linguistic contextualization, which transforms generic safety queries into contextually grounded multimodal samples. In contrast, KSAFE-MM-C targets safety vulnerabilities that are culture-dependent, using localized visual queries drawn from real-world. It pairs these visual queries with jailbreak-style textual queries to cover multimodal safety risks involving cultural visual cues and malicious textual intent. We evaluate 12 state-of-the-art MLLMs on KSAFE-MM and reveal culturally grounded vulnerabilities that translation-based evaluation fails to capture. Notably, jailbreaking strategies substantially amplify attack success rates, with ProgramExecution yielding up to 74.2% ASR compared to 13.4% for standard queries. Furthermore, we identify a systematic trade-off between safety and over-refusal, where models achieving low ASR tend to exhibit excessive refusal behavior on benign queries. These findings highlight the urgent need for culturally grounded safety evaluation beyond English-centric benchmarks.

Large language models often exhibit a substantial gap between their performance in English and in lower-resourced languages on equivalent knowledge queries---a cross-lingual consistency issue that remains underexplored for Indian languages and their code-mixed counterparts. To study this gap, we introduce IndicKLAR, an Indic extension of the KLAR-CLC benchmark covering 18 of the 22 scheduled Indian languages. For 11 widely used language pairs, we additionally provide code-mixed variants. Both monolingual and code-mixed inputs verified by native speakers. This three-way alignment enables us to examine how knowledge recall consistency varies across English, code-mixed, and native Indian language inputs. Across nine open-weight models, we find that the accuracy gap between native-language and English inputs can reach $\sim$0.50, while code-mixed inputs substantially reduce this gap, bringing performance within $\sim$0.05 of English without any model-level intervention. Motivated by this finding, we evaluate several prompting strategies that differ in how explicitly language conversion is exposed: a two-stage translate-then-answer setup, a one-stage joint translation-and-answer prompt, and Translate-in-Thought (TinT)---a single-step strategy in which the model internally converts the input and outputs only the final answer. Across the native $\rightarrow$ code-mixed $\rightarrow$ English performance trajectory, we observe a consistent flip point---the transition from incorrect to correct prediction---between the native and code-mixed settings. Notably, this pattern holds both when the code-mixed representation is explicitly provided as input or when the model is prompted to convert internally using TinT.

Proactive assistants continuously decide when to intervene and what context should support the intervention. Large language model (LLM) pipelines repeatedly interpret activity histories to make these decisions, paying an inference cost even when the assistant remains silent. We show that a small graph model can handle both decisions and improve the language agents it controls. Our key insight is that user activity has a native graph structure: events involve persistent entities whose recurrence connects interactions over time. Triggering and context selection map directly to predictions on event and entity nodes. We introduce a temporal-graph-learning (TGL) controller that learns these predictions jointly and supplies both outputs in one forward pass. The downstream language agent generates suggestions on triggered events using the activity history and scored entities. At 11.13 ms per event on a GPU server, TGL achieves the highest AUCs among nine trigger architectures and gives approximately $4$--$7\times$ trigger-stage speedups over the two single-forward LLM triggers. A shared TGL model improves F1 across all 14 downstream backbones by a mean of 16.7 points. The controller also runs at 13.99 ms on a consumer laptop with an approximately 220 MiB BF16 resident footprint, bringing effective proactive control to on-device deployment.

Public discourse about artificial intelligence (AI) often uses anthropomorphic language: language that attributes human capabilities and characteristics to AI systems. This practice has been criticized for setting misleading expectations, inflating claims, and fueling hype around AI, which may distort public understanding of AI and impact policy priorities. We study the effects of anthropomorphic framing by comparing changes in participants' perceptions of AI (N=815) when reading passages with and without anthropomorphic language, designed to reflect realistic public-facing AI discourse. We further examine whether these effects differ across two types of AI technologies -- large language models and recommendation systems -- and measure changes in perceptions of AI across several dimensions that are prominent in current public discourse. In a separate condition using a text that explicitly discusses the dangers of AI, we show that individuals' views of AI can shift in response to reading a text; yet in the main conditions of the experiment, where we compare anthropomorphic and non-anthropomorphic descriptions, we find that whether the text uses anthropomorphic language does not substantially affect participants' perceptions of AI. Our results indicate that any immediate effects on opinions of AI are modest, although they leave open the possibility that anthropomorphic language could have an effect in naturalistic settings, or over gradual, continued exposure.

Medical large language models are commonly adapted with a fixed low-rank budget, even though medical questions differ substantially in confidence, clinical coverage, and cross-domain difficulty. We study adaptive rank budgeting for parameter-efficient medical question answering: for each question, the adapter decides whether to activate a small, medium, or large subset of LoRA rank channels. The central challenge is that a naive adaptive budget router can collapse to unstable choices or spend capacity without improving shifted benchmarks. We propose TriageRA-CCF, a source-side teacher for adaptive rank-budgeted LoRA. It combines three signals computed only from source training data: base-model answer confidence, metadata-cell clinical coverage, and a counterfactual close-miss proxy. These signals supervise a straight-through budget router over active ranks {2,4,8}, together with budget-cost, entropy, and rank-balance regularization. Under a matched CMB-source training protocol, TriageRA-CCF achieves the best average accuracy among LoRA, DoRA, and MoELoRA baselines on both Qwen3-8B and Llama3.1-8B. The gains are modest and non-uniform across benchmarks: +0.21 average points over the strongest external baseline on Qwen3-8B and +0.16 on Llama3.1-8B. Component ablations show that confidence, coverage, and counterfactual signals all provide useful budget supervision, but their combination is not monotonically best on every backbone.

Diversity in LLM mathematical reasoning is critical for exploration, but common diversity metrics mostly capture surface-level variation rather than differences in how a problem is solved. We address this gap by introducing approach-level diversity: variation in strategies across correct solutions to the same problem. Using a human-calibrated LLM judge framework, we show that prior diversity measures are unreliable proxies for approach-level diversity, and this mismatch carries over to diversity-aware RLVR, where target metrics are preserved while approach-level diversity declines. Investigating when approach-level diversity helps and whether it can be directly induced, we find that approach-diverse candidate sets improve test-time scaling. However, optimizing an LLM judge diversity reward during training causes the policy to exploit judge-specific preferences rather than broaden its approaches, leaving direct optimization of approach-level diversity as an open problem. Together, our work introduces the notion of approach-level diversity and uncovers a systematic divergence between surface- and approach-level signals, marking a step toward LLMs that reason in genuinely diverse, human-like ways.

Large language models (LLMs) have achieved remarkable performance on high-school and competition-level mathematics, yet their capabilities on advanced mathematics remain poorly understood. Existing benchmarks, however, fall short in both scope and evaluation granularity: they provide limited disciplinary coverage and often rely on final-answer correctness or coarse judgments, leaving the validity of the reasoning process inadequately assessed. To bridge this gap, we introduce AdvancedMathBench, a benchmark suite designed to evaluate the reasoning capabilities of LLMs on advanced mathematical proofs. Its core generation benchmark, ProverBench, contains 245 problems spanning undergraduate (UG) and doctoral qualifying-exam (QE) levels. To reliably evaluate these proofs, we develop a dedicated automatic verification pipeline that is trained on large-scale expert annotations, produces both correctness verdicts and fine-grained analyses, and exhibits strong agreement with human experts on held-out proof trajectories. We further introduce VerifierBench, consisting of 888 model-generated proof trajectories paired with expert ground truth, to evaluate whether models can correctly judge proof validity and provide sound verification rationales. Experiments show that AdvancedMathBench remains challenging for frontier models. On proof generation, the best-performing model, GPT-5.5-xhigh, achieves only 64.5 and 48.9 on the UG and QE splits, respectively. On proof verification, the best model only attains a Balanced F1 of 65.1. Further analysis reveals a notable mismatch between proof generation and verification capabilities across models.

Large language models can write, patch, and search code, but oncall root cause analysis (RCA) demands something different: reasoning over noisy metrics, logs, traces, and source code, starting from ambiguous user-facing reports, often hours after the incident began. We introduce ORCA-bench, a benchmark that puts general-purpose coding agents in a production-fidelity oncall setting. ORCA-bench pairs 1,079 RCA tasks with six days of metrics, logs, and traces collected from an OpenTelemetry-instrumented microservice system under continuous simulated user load. Agents investigate this recorded history through real observability interfaces---Prometheus, Jaeger, and OpenSearch via Grafana---with full access to application source code. Tasks systematically vary report specificity, time-to-detection, and co-occurring fault scenarios. Ground-truth symptoms are curated and signed off by expert SREs, and our LLM-as-judge is independently re-scored by humans (Cohen's $\kappa_w = 0.91$). Across five frontier agents, the best RCA Accuracy is 25.3% on Medium-difficulty tasks (the realistic-input setting) and 10.0% on Hard---a gap that remains even with Claude Fable 5. The weakest model hallucinates an implausible root cause in 40% of incident reports, and removing source-code access reduces RCA accuracy and increases the hallucination rate for every evaluated model. These results come from a curated 50 GB / six-day testbed of standalone tasks on a system whose code and instrumentation are public. Since real production systems are orders of magnitude larger, more dynamic, and more idiosyncratic, the gap we report underscores the engineering work still needed before agents can be entrusted with production reliability. We release the public set at https://hub.harborframework.com/datasets/orca-bench/orca-bench.

Domain adaptation of small language models (SLMs) has emerged as a practical strategy for deploying capable NLP systems in resource-constrained, high-stakes environments including healthcare, legal services, and financial analysis. While performance gains from parameter-efficient fine-tuning are well characterised, the corresponding impact on trustworthiness (factual calibration and adversarial robustness) remains poorly understood. This paper presents the first systematic cross-domain, cross-architecture empirical study quantifying the trustworthiness cost of domain adaptation across three SLM architectures (TinyLlama 1B, Gemma-2 2B, Llama 3.2 1B), three domains (healthcare, legal, finance), two training-data conditions (benign and adversarially perturbed), and four fine-tuning strategies (baseline LoRA, Safety-DPO, Dark Experience Replay, and Task Arithmetic LoRA, TA-LoRA). Trustworthiness is evaluated through TruthfulQA MC2 (factual calibration) and HarmBench ASR (adversarial robustness) across all 216 experimental configurations with three random seeds. Three principal findings emerge. First, baseline QLoRA domain adaptation produces minimal TruthfulQA MC2 change across all model-domain combinations (mean |Delta TQA| < 0.02). Second, adversarially perturbed training data consistently improves domain adaptation quality (Delta loss approximately -0.040) without worsening trustworthiness benchmarks. Third, none of the three safety-preserving strategies reduced adversarial harm susceptibility: Safety-DPO was effectively neutral (mean Delta ASR < 0.001), while Dark ER and TA-LoRA increased mean HarmBench ASR by +0.171 and +0.155 respectively in safety-aligned models (Gemma-2 2B, Llama 3.2 1B), with individual configurations exceeding +0.45. These results challenge the assumption that replay-based and arithmetic-merge strategies transfer alignment to domain-adapted SLMs.

Conflicting objectives are general in RL alignment, and training on them data-efficiently is hard. Training a safety guard with RL means optimizing two objectives that conflict: catch real harm, and do not refuse benign prompts. Our finding is that over-refusal improves 22.4% to 12.8%, while under-refusal on adversarial attacks silently worsens 0.27 to 0.33. We present C-Instrument, a constitution-grid data instrument that generates the RL training data, and C-LIM, a per-cell learnability score that decides each cell's move: prune, densify, amend, expand. C-LIM flags the dead-weight data region before any training budget is spent: 187 untargeted rows had bought zero gain, and our method lifts the same region's learning impact 0.733 to 0.80. Code and the constitution are open-sourced.

AI agents operate in persistent environments where early state changes can influence decisions far into the future. Unlike conventional language-model interactions, agent behavior is mediated through a shared state that is repeatedly modified and reused across long-horizon workflows. Current safety benchmarks often fail to capture these cumulative risks because they focus on short, static tasks. To address these limitations, we introduce OpenART, an open-ended arena for scalable agent red teaming through environment evolution. OpenART provides over 10,000 validated stateful scenarios across 50 domains, drawing from a pool of more than 500,000 tools and skills. These tasks require a median of 97 tool calls and enable unified evaluation across 75 different agent-model configurations. To systematically explore these evolving attack surfaces, we propose the Evolutionary Markov Hypergraph Attack (EMHA). EMHA is a black-box policy that performs feedback-driven environment evolution by coordinating authorized state transitions without requiring parameter updates. Throughout the evaluation, task objectives remain fixed while only the environment state changes. Across all configurations, EMHA achieves a pooled Attack Success Rate (ASR) of 85.0%. Its advantage over instruction-only evolution increases from approximately 2% on simple environments to over 17% on the most complex ones, demonstrating that environment evolution increasingly exposes safety failures as task complexity grows. Furthermore, our analysis shows that the specific runtime implementation of an agent explains a significant portion of safety variation beyond the underlying model's capabilities. These results establish OpenART as a scalable foundation for studying agent safety in complex, evolving environments. Code is avaible: https://github.com/AI45Lab/OpenART#

Parameter-efficient fine-tuning (PEFT) enables efficient adaptation of large language models, but existing MoE-based PEFT methods typically improve capacity by storing multiple full LoRA experts, causing adapter storage to grow linearly with the number of experts and restricting adaptation to a fixed expert pool. We ask whether MoE-based PEFT can produce instance-specific adaptations without explicitly storing a separate LoRA module for each expert. To address this gap, we propose MoEGen, an adaptation framework that shifts MoE-based PEFT from expert selection to expert-conditioned parameter generation. Instead of storing each expert as a full LoRA adapter, MoEGen represents each expert as a small learnable vector, termed an expert code. It routes each input over these vectors and uses their weighted combination to condition a lightweight hypernetwork that generates input-specific low-rank updates. This design decouples expert capacity from adapter storage while enabling instance-conditioned adaptation. Experiments on eight commonsense reasoning benchmarks show consistent improvements over strong static and MoE-based PEFT baselines across three backbones. MoEGen also performs strongly in joint medical and legal-domain adaptation.

We build VectraYX-Vision-1B, a sub-2B Spanish/LATAM cybersecurity vision-language model for offline use, and measure what limits its visual grounding. A frozen 1.04B-parameter decoder is coupled to a frozen vision encoder through a trainable projector: first SigLIP, then the Qwen2-VL-2B tower with a projector shaped for llama.cpp's mmproj export. With SigLIP, after five fine-tuning defects were repaired, a nine-field extraction gate with a shuffled-image control passes the same 2/9 fields under every configuration that keeps the encoder and adds no text hint. A frozen-feature probe explains why. Transplanting the Qwen2-VL tower with its own merger reads an 8-nibble address SigLIP never read (0.00 to 0.81 exact match). On B8 (2,040 items, 34 fields, 16 templates, shuffled-image and best-constant controls), 9 fields pass on a Qwen-tower checkpoint, all within trained template-field pairs. Screenshot tool identification (B6) sat at exactly 0.000. That was not a perception ceiling. Stock Qwen2-VL-2B transcribes the same images at word recall 0.93 at our pixel budget (0.52 with our letterboxing). Training only the projector on B6's task format and render family lifts B6 tool identification to 0.96 and recall to 0.48-0.51. Perception of rendered text is therefore present and trainable in this frozen-backbone design. The evidence is narrow. B6 is nearly in-distribution for that projector, which also forgets part of B8 (mean field accuracy 0.53 to 0.29 on a retention check); runs are single-seed, and no checkpoint yet has both. We document four harness defects, retract an earlier B6 score, and release code, benchmarks and checkpoints.

We present the first systematic study of massive activations (MAs) in layer-interleaved Hybrid linear attention large language models (HLA LLMs), examining their architectural organization, training-time emergence, underlying mechanisms, and functional significance. Across five linear attention architectures, six hybridization configurations, and five input domains, we identify two architecture-aligned morphologies: pre-attention spikes (PAS) immediately before full attention and inter-spike plateaus (ISP) persisting through intervening linear attention layers. Denser full attention increasingly connects PAS through ISP, approaching the persistent MAs of conventional Transformers. This organization also recurs across 12 public checkpoints spanning 1.2B-397B parameters, covering linear attention and state-space hybrids. Controlled pretraining of Gated DeltaNet (GDN) hybrids up to 1.3B reveals early emergence and consolidation of both morphologies, alongside asymmetric gating effects. Specifically, full attention output gates strongly attenuate MA magnitudes without eliminating their organization, whereas removing GDN output gates yields modest amplification. Mechanistically, we develop a shared systematic-outlier account: PAS follows a localized write-sink-cancel process, while ISP is consistent with delayed cancellation. Functionally, our interventions show that deleting only the four largest-magnitude PAS coordinates at each full attention input reduces mean downstream accuracy by 21.9%-63.6% relative to normal inference. Moreover, reference-conditioned spike-to-plateau connection consistently improves mean real-world retrieval accuracy, yielding relative gains of 1.1%-12.6% without retraining. Our code is available at https://github.com/StartLuxLabs/Massive-Activations-HLA.

Understanding language requires tracking entities across discourse - i.e., knowing where things are and how they change, even when not explicitly stated. Whether language models perform such tracking in a human-like fashion remains unclear, in part because existing evaluations rely on artificial tasks, far removed from natural language comprehension, and lack comparisons to humans. Here, we evaluate entity tracking in both language models and humans (N = 48) using naturalistic narratives at multiple levels of complexity. In humans, we find that entity tracking degrades specifically with narrative complexity, not narrative length. In language models, we find that human-level entity tracking is already present at 410 million parameters - well below the multi-billion parameter, code-specialised models identified by prior work - and improves with scale, with contemporary models far exceeding human performance. Together, these results demonstrate that entity tracking, a core component of language understanding, emerges at model scales far smaller than previously thought.

Current cultural evaluations for large language models (LLMs) often reduce culture to single-turn factual recall via MCQs, failing to capture a common use case: users seeking practical help over multiple turns in culturally grounded scenarios. We introduce CultureConverse, a scalable, multilingual simulation and evaluation harness for culturally grounded assistant dialogue that covers 10 East and Southeast Asian regions, 58 subgroup identities, and 7 domains. Each simulated and evaluated episode produces a scored interaction where the assistant assists the user and infers cultural constraints from partial information. The resulting CultureConverse-DS dataset contains 14,610 benchmark (evaluation) episodes and 274,295 oracle-guided (gold-mode) dialogues. In our benchmark evaluation of 18 models, GPT-5 mini achieves the highest assistance quality. Human annotation experiments suggest that our evaluation framework is a sufficient proxy for human judgment. Performance gains from fine-tuning on 27,860 high-quality CultureConverse-DS samples improve in-domain assistance and transfer out-of-domain to cultural MCQ and safety classification benchmarks. We release the harness, both splits, and judge prompts to support interactive evaluation of cultural competency.

Specialized machine learning architectures encode structural assumptions (equivariance, permutation invariance, relational structure) that language-based foundation models lack by design. This review asks whether such assumptions can instead be acquired through language, drawing on a corpus of 186 papers published between 2016 and 2026 across nine modalities: tabular data, graphs, time series, vision, chemistry, code, knowledge graphs, point clouds, and protein structure. Methods are organized into eight representational regimes, ranging from language-only prompting to fully specialized architectures, and are assessed not only on predictive accuracy but on whether structural information is preserved by the representation and computed by the model. Language-mediated models prove competitive in extreme few-shot prediction, discretized symbolic tasks, and textually annotated knowledge graphs. No evidence of general replacement survives once structure itself is evaluated: where direct tests exist, apparent parity is traceable to information asymmetry, benchmark contamination, or computation performed outside the language model. Across independent research communities, missing structural inductive biases are reintroduced through graph modules, structure-aware tokens, or specialized attention, indicating that specialization is relocated rather than eliminated. The review closes by specifying a falsifiable controlled experiment that would measure the remaining gap directly. The derived data supporting the findings of this review are openly available in the repository at https://github.com/kiyan-rezaee/language-vs-structure.

Criminal judgment prediction requires models to infer statutory articles, charges, and sentencing outcomes from case facts. Unlike standard classification tasks, it involves a structured reasoning process in which statutes should be matched with facts, charges should be justified by statutes, and sentencing outcomes should remain consistent with charges. Existing approaches optimize final labels, and while some have attempted to evaluate reasoning quality, their evaluations are indirect, often relying on LLM-generated rubrics that reflect model-internal preferences rather than the inherent logical structure of legal adjudication. We propose Juris Policy Optimization (JPO), a post-training framework for structured legal reasoning in Chinese criminal judgment prediction. JPO first uses teacher-generated rationales to supervise a standardized four-step reasoning process, and then applies reinforcement learning with a composite reward over legal prediction quality, reasoning structure completeness, and cross-step consistency. JPO further introduces token-level advantage reweighting and adaptive clipping for legally salient reasoning segments. Experiments on multiple open-source language models and three Chinese legal benchmarks show that JPO consistently improves both judgment prediction and reasoning quality over supervised fine-tuning and reinforcement learning baselines.

Reranking methods, such as Minimum Bayes Risk (MBR) decoding and Quality Estimation (QE) reranking, have been widely used in modern neural machine translation (NMT) to select an output from a set of candidate hypotheses. However, the performance gains come at the cost of high inference latency. Existing acceleration methods target MBR decoding and reduce only the reranking computation, leaving QE reranking unaddressed and candidate generation---which can be the larger computational bottleneck---largely untouched. In this work, we propose Quit (Quantifying Uncertainty for Incremental Termination), a novel early-stopping strategy for the entire generation--reranking pipeline. Quit treats candidate generation as a sequential decision-making process under uncertainty. It incrementally generates and reranks candidates, stopping when the best reranking score stabilizes. Comprehensive experiments with three NMT models across 19 language pairs show that Quit achieves end-to-end speedups of $1.47$--$2.66\times$ for MBR decoding and $3.43$--$4.12\times$ for QE reranking while preserving automatic metric scores.

Speculative decoding accelerates LLM inference by drafting candidate tokens and verifying them in parallel. Tree-attention drafters such as EAGLE-3 are widely adopted, yet typically hold two decisions fixed: (1) a strict token-match verification rule and (2) a static draft-tree shape. Prior work relaxes each in isolation under limiting assumptions: long draft chains for training-free lossy verification, and adaptive tree shaping under a fixed token budget. We introduce AdaptiveSpec, a training-free per-step speculative decoding method that adapts both decisions from internal signals already produced during decoding. A per-step margin rule promotes a mismatched draft-proposed token when the ratio of the target's probability on the drafted token to its top-1 probability exceeds a threshold with no dependence on draft length or underlying drafter architecture. A per-step tree policy adjusts the draft tree's depth, width, and node count directly from a fused signal of draft top-1 confidence and a rolling acceptance history capturing recent draft-target agreement, allowing the total draft count to vary rather than only be redistributed. The two adaptations operate on orthogonal axes and compound in effect. Implemented on the SGLang production-grade serving engine, AdaptiveSpec improves throughput over the state-of-the-art autoregressive speculative decoding method EAGLE-3 by up to 56%, recovering 93% to fully lossless task accuracy across GSM8K, MATH-500, and HumanEval on three target models (DeepSeek-R1-Distill-Llama-8B, Llama-3.1-8B-Instruct, Qwen3-8B).

Counterfactual fairness audits of clinical language-model agents report a flip rate: how often an action changes when only the patient's demographic descriptor changes. Part of that rate is not demographic. A stochastic agent also changes its own action when nothing changes, and a flip rate cannot be interpreted without knowing how often. We measured it. Re-running one condition ten times over sixteen synthetic vignettes at default sampling changed a clinical agent's action in 8.7 percent of replicate pairs, from 2.2 percent for intensive-care escalation to 17.9 percent for controlled-substance caution, an output given no operational criteria. Across six models from five vendors, pooled floors ranged from 2.5 to 23.7 percent; in this panel neither disclosed size, vendor, nor hosting ordered them. The floor depends on the decoding configuration: majority voting over five draws removed 39 percent of it (95 percent confidence interval (CI) 18 to 64); at temperature 0 three of four locally served models showed no disagreement, but a hosted model still did. We also show that, for a binary action, the flip rate expected under no demographic effect equals the floor and a real effect adds only its square, so a flip rate inside the floor is not evidence of fairness, and direction must be tested with a signed paired test. We give a four-step reporting procedure and release FairMedAgent, the harness, with its protocol, vignettes and analysis scripts, so any team can measure the floor for its own agent.

For scientific progress, we need benchmarks that test the limits of state-of-the-art models, and evaluation methods that inform us about failure cases. As models get stronger, standard benchmarks for machine translation are approaching saturation. Further, automatic translation metrics are unreliable, opaque, and vulnerable to reward-hacking. Even gold human evaluation is not problem-free, because it often lacks reproducibility, objectivity, and scalability. Overall, this prevents us from tracking progress in the field and identifying pathways for improvement. We introduce the Last Translation Benchmark, a collection of human-authored and peer-reviewed examples (texts, images, audio, videos) that break leading machine translation models. We also present a new evaluation approach: each example comes with handcrafted verification rules describing concrete failure cases on that example, therefore allowing reliable and actionable future evaluation. The Last Translation Benchmark is a live dataset that accepts ongoing contributions. The latest version is LTBv1, containing accepted contributions prior to September 1st 2026, with future releases planned as new data is continuously collected.

How fast does a language model degrade when trained on its own outputs? Theory traces it to gradually accumulating errors, while experiments report repeated phrases within ten generations. Under a fixed sampling seed in vLLM, the fast loss of lexical diversity comes from the sampler. When vLLM serves a batch from one seeded sampling configuration, every request receives the same random draws, and a fixed seed replays them every generation. Fine-tuning raises the tokens that won, and the replayed draws let them win by more. Sharing across requests and replay across generations matter only together. Remove either one, by changing the shared seed every generation or by giving each request its own seed that repeats every generation, and the unique-4-gram fraction of two StableLM checkpoints stays near its starting value of about 0.98 through generation 3. Keep both, and the replayed shared seed takes seven checkpoints from five families to between 0.045 and 0.38 by then. Three generations of replay write the favoured phrases into the weights: decoded with one seed per request, the generation-3 weights of the replayed StableLM-2-1.6B chain recover most of their diversity, yet the phrase that filled every sample under the shared seed still opens 46% of them. Without replay, five checkpoints drift slowly, consistent with the gradual accumulation that theory describes, and three turn incoherent though their diversity scores stay high. One peer-reviewed model-collapse pipeline that fine-tunes Gemma-2-27B samples identical prompts under one seeded configuration, and three quarters of the rows it released for one iteration repeat nearly as often as one such batch copies them. A seed per request restores the fresh sample that stability analyses assume.

We demonstrate that a general-purpose agent can directly drive a physical robot throughout task execution without any task-specific or environment-specific training. We introduce Agent as Policy (AGP), which places task planning and execution under the agent's control. Given a task and a robot interface, the agent interprets visual evidence, writes executable programs, issues motion commands, and revises its actions in response to physical outcomes. This brings the agent's reasoning and programming capabilities into continuous interaction with the physical world. We study AGP across multiple real-world manipulation tasks spanning precision manipulation, dynamic motions, and deformable objects. These include assembly from human videos, block construction from goal images, dice flipping, targeted throwing, and bimanual towel folding. AGP achieves success rates of at least 80% in seven of eight task configurations and significantly outperforms previous agentic robot systems. We further study efficiency through task experience accumulation and find that reusing saved procedures and programs shortens execution time across repeated trials. These findings support a path for general-purpose agents to act as robot policies, extending their autonomy to physical manipulation through runtime reasoning, programming, and interaction.

Large language models can hallucinate even when the knowledge required for a correct answer is already available. We study this failure through a latent-key view of inference, where answer selection depends on competition among associations acquired during pretraining. We show that model predictions can be highly sensitive to individual query keywords, that these influential keywords exhibit entity-specific binding, and that their effects are systematically shaped by pretraining frequency. Multiple bindings can also compete and exhibit higher-order interactions within the same query. Based on this mechanism, we introduce a two-stage keyword-perturbation method for hallucination detection. By removing influential keywords and measuring how the model reorganizes its prediction, the method distinguishes errors caused by misleading key associations from correct decisions supported by diagnostic evidence. Across multiple models and benchmarks, perturbation provides a strong and transferable detection signal, reaching $0.910$ AUROC on probe-known ScientistQA. Finally, we extend the same probabilistic framework to four hallucination regimes: knowledge deficit, wrong knowledge, context distraction, and unstable inference. Their operational distributions across benchmarks provide diagnostic context for why different detector families succeed in different settings.

Large language models (LLMs) have demonstrated strong capabilities in document key-value pair (KVP) extraction, yet controlled evaluations of their robustness to optical character recognition (OCR) output remain limited. This leaves an important gap in understanding their reliability in real-world OCR-to-LLM pipelines. Unlike end-to-end Vision-Language Models (VLMs), which jointly perform visual perception and semantic extraction, modular pipelines allow these stages and their errors to be isolated and audited. We introduce a controlled benchmark that distinguishes downstream LLM extraction behavior from upstream OCR degradation. It evaluates 136 experimental configurations and 17,688 document-level inferences generated with deterministic decoding across five instruction-tuned open-weight LLMs (2B-8B parameters), three datasets, and four text-quality conditions. The evaluation combines a full zero-shot comparison, targeted one- to three-shot experiments, and a sensitivity analysis of 40 configurations across 20 frozen demonstration sets. By separating Key Recall (annotated-field recovery) from Exact Match and Value F1 (exact and partial value recovery, respectively), we test whether OCR degradation affects field identification and value reproduction differently across models. Our findings challenge three practical assumptions: (1) clean-text performance reliably predicts real-world robustness, (2) model rankings remain consistent across annotation-derived Gold and OCR-derived text, and (3) additional few-shot demonstrations monotonically improve extraction accuracy. The observed model-ranking reversals and unstable few-shot gains expose important reliability risks under noisy document conditions. We release the benchmarking framework, dataset splits, and evaluation scripts to support reproducible research.

We built a node that grounds a replaceable language model in a maintained ontology corpus, then asked what its successful-looking evaluation could support. Across ten models, grounding raised target-name recall from 0.265 unaided to about 0.92. A copy baseline, the recall a verbatim copy of the shown context already achieves, scores 0.964, and every model sits 0.022 to 0.067 below it. Copying therefore scores higher on this limited recall measure, which does not assess whether answers are better. The comparison tests what a recall score establishes; it does not test whether reasoning occurred, because a reasoned answer and a copy score alike when the answer name is already in context. We report exposure accounting (four counts classifying each gold item by whether the context exposed it and the answer recovered it) and a model-judged audit of 423 sampled item observations. A separate paired production study found a model-judged quality gain of +0.27 [+0.11, +0.45] on a 0-5 scale. Operational studies found failures that recall alone would not show: rephrasing questions out of the graph's vocabulary cut exposure from 0.964 to 0.328, yet the absence-keyed fallback would have fired on only 2 of 506; and inserting extracted facts degraded judged pages in every arm, so that step was disabled. Five-arm controls show that any well-formed on-corpus block beats no context but do not establish that the specific content matters, and no matched comparison against flat-text retrieval was run. The corpus is public and largely LLM-generated, which establishes neither training exposure nor novelty. Each study has its own outcome measure. Where gold derives from the injected corpus, we recommend reporting the accounting beside quality judgements, not in place of them.

We present the first experiments on isolated sign language recognition (ISLR) for Icelandic Sign Language (\'ITM). We use \'ITM SignWiki, a dataset derived from a bilingual Icelandic--\'ITM online dictionary. It is genuinely low-resource: 1,845 videos cover 849 classes, 86% of which have only two examples, making the full task effectively one-shot recognition across signers. We compare two open-source ISLR frameworks, OpenHands and SPOTER, on three tasks of increasing vocabulary size (22, 117 and 849 classes), and evaluate three pose estimators and two forms of cross-lingual transfer. With \'ITM data alone, SPOTER outperforms OpenHands on all three tasks, and MediaPipe poses give better results than AlphaPose or SDPose. Cross-lingual transfer brings the largest gains: pretraining SPOTER on American Sign Language data before finetuning on \'ITM raises accuracy by 14--24 percentage points, to 72.7%, 47.9% and 22.6% on the three tasks, and multilingual training with data from six other sign languages lifts OpenHands from 1.41% to 28.86% on the full task. Although far from practical use, the results suggest that transfer from better-resourced sign languages is promising for very low-resource ones. We release our adapted versions of both frameworks.

LLM judges score outputs against rubrics well enough to have become the norm, both in benchmarks and as rewards for training. Jev, a classifier-like alternative its creators call a "decision model", returns probabilities over permitted answers with a calibrated confidence score, which LLM judges do not natively provide. We compare Jev with three flash-tier LLM judges on nine panels from seven benchmarks with human judgments, giving every judge identical criterion texts. The LLM judges run in two setups: holistically, reading a whole rubric at once as Jev does, and one criterion at a time. Jev can often stand in for them. They cost 16 to 325 times as much and take 28 to 350 times as long, yet in each setup Jev's accuracy differs significantly from theirs in at most 8 of 27 paired comparisons, ahead mostly on binary checklist criteria and behind only on ordinal ones. Despite their different designs, the two kinds of judge err alike. On ordinal criteria, all LLM judges and Jev depart from the human raters together, agreeing more with one another than with the labels and mostly assigning lower levels. On Jev's most confident errors, about 96% of LLM verdicts repeat its wrong answer, where independent errors would give about half. Intuitively, calibrated confidence should make Jev an ideal first stage of a cascade that defers uncertain verdicts to an LLM judge. Yet such cascades only lower cost while adding little accuracy: even with oracle thresholds, none beats the best single judge by more than 2.7 points. Calibration can tell a cascade when to defer, but the cascade also needs a fallback that errs elsewhere; these judges are wrong in the same places. These findings, which hold in both setups and at high reasoning effort, suggest that a cascade of judges succeeds only when its judges make complementary errors, and that future decision models should be designed afresh with that aim.

Multilingual text-vision embedding models are essential for cross-lingual image-text retrieval, but Southeast Asian languages remain poorly supported due to the region's linguistic diversity and limited data and computing resources. In this paper, we introduce SEA-CLIP-Tiny, a compact multilingual text-vision embedding model for Southeast Asia with fewer than 50M parameters. Our model adapts a CLIP-KD-style framework to Southeast Asian multilingual settings through regional data curation and multilingual teacher guidance. Experiments across seven Southeast Asian languages show that SEA-CLIP-Tiny achieves the strongest average retrieval performance among the evaluated student models, reaching 12.9%, 31.5%, and 42.2% at R@1, R@5, and R@10, respectively. Compared with MobileCLIP2, it improves average R@10 by 12.1 points while using 38.4% fewer parameters and lower measured CPU latency. These results highlight the importance of region-aware training for efficient multilingual text-vision models in Southeast Asia.

Document rerankers determine what evidence reaches the downstream model in RAG and deep research, yet mainstream rerankers select by relevance matching, and individually relevant documents rarely constitute the complete, complementary, non-redundant set a complex information need demands. Prior work rewards a set by its aggregate rubric score, shifting the objective from ranking documents to composing sets. Yet that score is one scalar shared by every document in the set, so the supervision is sparse: a redundant document is rewarded with the rest whenever the set scores well, and a decisive one penalized with the rest whenever it does not; credit assignment leaves contributors indistinguishable from free riders. On-policy distillation could densify this supervision, but existing methods give every rollout the same fixed guidance, too prescriptive for strong rollouts and too abstract for weak ones. We therefore propose AdaTutoRank, a setwise reranker trained with Adaptive Tutoring Optimization (ATO) under a three-level hierarchy of nine rubric dimensions, which supplies silver labels for the cold start, rewards for reinforcement learning, and hints for distillation. ATO draws three hint forms of increasing specificity from the policy's own frozen snapshot: the rubrics alone, a self-selector's sibling-set chosen under rubrics, and a self-reflector's reflection contrasting the rollout with that sibling-set; each rollout receives the form matched to its quality. Re-scoring that rollout under the hint-conditioned frozen teacher and the hint-free snapshot distills the hint's effect into a token-level advantage that complements the group-relative outcome advantage. Across ten benchmarks spanning RAG, deep research, and setwise evaluation, AdaTutoRank attains the best overall performance while issuing fewer retrieval calls.

Multidisciplinary tumor boards integrate multimodal clinical observations and longitudinal patient histories through specialist discussions, yet benchmarks rarely capture these real-world trajectories. We introduce OpenTumorBoard, a benchmark with 611 patient cases and 19,157 discussion turns across ten specialist roles, transcribed from 12,534 minutes of publicly available tumor board recordings on YouTube. The benchmark evaluates two settings: SPECIALIST TURN, in which an LLM responds to a clinically significant question posed during a real discussion, and BOARD SIMULATION, in which it generates an entire back-and-forth discussion and reaches a consensus on therapy recommendations, surgical plans, next actions and clinical trial matching. Evaluation of 14 general-purpose frontier and medical LLMs reveals substantial limitations: the best models score 3.43 out of 5 in clinical equivalence to specialist answers and 2.78 out of 5 in alignment with recorded board conclusions. Supervised finetuning and reinforcement learning improve performance on a held-out test set, suggesting that real-world discussion trajectories can support model adaptation. Three M.D. experts review a subset of the benchmark, finding high information coverage and factuality of patient cases and strong fidelity of extracted consensus conclusions. We will release OpenTumorBoard and its automated curation pipeline to support the development and evaluation of LLMs for multidisciplinary, personalized cancer decision-making.

Generating a correct answer does not mean that a language model will select it. We separate factual recall into three steps: generating a correct candidate, ranking the available candidates, and selecting the final answer. Pre-generation readouts predict factual recall and which questions sampling will cover across three model families, but say little about whether an available correct answer will ultimately be selected. Explicit verification with $P(\mathrm{True})$ improves within-question ranking over mean log-likelihood in Gemma, Qwen3, and Llama, with AUROC gains of $0.08$--$0.12$. In a prospectively defined Gemma cohort, verification raises plurality accuracy by about $5$ points, and still gains about $2$ points over chat-template likelihood, a stronger generative baseline. The advantage is strongest for relations with common-answer priors and depends on access to the entity; masking the entity removes the ranking advantage in larger Qwen models. Finally, the measured benefit depends on how correctness is defined: recall-oriented reference matching can credit option lists favored by likelihood and substantially understate the improvement seen under human semantic judgments. Prior work shows that models can carry latent factual knowledge and judge candidate answers; we show that these capabilities do not collapse into a single notion of ``knowing,'' and trace where information is gained, lost, or mismeasured between availability, ranking, and final choice.

Large language models are trained to always produce an answer, regardless of whether they possess the relevant knowledge, which leads them to fabricate facts. Prior work has shown that LLMs' confidence estimates correspond poorly to their actual performance, and that fine-tuning can substantially improve them. However, what models actually learn during such training remains poorly understood. We investigate how LLMs acquire metacognitive monitoring, the ability to know what one knows, by training 10 open-weight LLMs to predict their own accuracy on factual multiple-choice questions before answering them. We find that trained confidence reflects two distinct signals. While on questions close to the training data, it tracks the model's true accuracy, in other domains, it instead tracks output consistency: the concentration of the model's answer distribution. Output consistency tracking emerges early in training and generalizes across datasets, whereas accuracy tracking develops later and remains local to the training distribution. These results suggest that calibration training may not teach models to generally detect errors they commit confidently, and they raise broader questions about the nature of metacognition in artificial systems.

A single random Gaussian probe gives an unbiased estimate of the squared Frobenius norm of a layer's quantization error. The estimator is well-behaved because round-to-nearest error is spectrally flat. Across 1,683 tensors from a 35B MoE and a 9B dense model, effective dimensionality is 0.93 to 0.96 times the i.i.d. noise value of the same shape, and on the MoE the median is unchanged from 2-bit to 8-bit. The probe coefficient of variation is predictable from tensor shape. One probe measures per-tensor sensitivity to within 4 to 7%; twenty probes reach 1.3 to 1.4%.RAM applies the propagated form of this estimator to budget-targeted mixed-precision quantization with no calibration data. Gaussian probes carrying the network's own input statistics score every tensor at six bit-widths. A knapsack solver allocates bits under an exact byte budget, with guardrails against catastrophic 2-bit assignments. One probe pass serves any budget. Isolated and propagated scores rank tensors independently on Qwen3.5-35B-A3B (Spearman -0.01), yet the propagated probe rank-correlates 0.81 to 0.83 with the GPTQ layer objective from real activations, while the isolated estimator is uncorrelated with it. That objective is the wrong allocation target: at matched bytes on Qwen3.8-27B, a block-output probe beats a vendor IQ3_M mix and an oracle that allocates from the real-activation objective. On Qwen3-8B the propagated probe ties HAWQ-V2 at matched bytes. Across seven architectures from 8B to 122B, with probe timing up to a 400B model in nine minutes on one workstation, RAM reaches 3.5 to 13.6% lower median WikiText-2 perplexity than size-comparable uniform 4-bit builds on the tested MoE models. (Black Sheep Ai baa.ai)

The strong version of the stochastic parrot argument claims that, although large language models (LLMs) may exceed rote regurgitation, they cannot move beyond statistical pattern matching into abstraction or reasoning, remaining ontologically near the lower bound of pattern reuse despite producing alluringly fluent text. We test this hypothesis using conlang-like tasks. Several LLMs are given only natural-language descriptions of fictional languages that subvert prominent superficial patterns in training data by combining statistically uncommon and unattested features. Crucially, no example outputs are given. We argue that if the models exhibit rule-following behaviour, they cannot be relying solely on superficial statistical patterns; such patterns often work against the correct output. Instead, successful performance requires representations of the constraints specified in the prompt. Across three complementary task families, models systematically move in the meaning-predicted direction: they distinguish prompt exposure from instructed use, alter semantic relationships in response to novel constraints, and sometimes produce exact matches to complex translation answer keys. Although performance varies across the spectrum of models used, these results provide evidence for meaning-mediated abstraction in LLMs and refute the strong stochastic parrot hypothesis. Our work shows that, under appropriate architectural and contextual constraints, statistical learning can produce meaning-mediated abstractions, although generation remains strongly constrained by superficial plausibility. We discuss implications for model development and for understanding how increasingly abstract representations may emerge from plausible-text-generation objectives.

Existing open-ended generation metrics measure likelihood, lexical diversity, or distributional similarity in generic representation space, yet can miss fundamental dimensions of quality. A prominent blind spot is global coherence: a generated passage may be locally fluent while remaining globally contradictory, causally inconsistent, or topically disconnected. We identify representation as a central bottleneck in detecting these failures and introduce CHORD (Coherence-aware Hidden-state Open-generation Reference Distance), a coherence-sensitive distributional metric. CHORD encodes generated and human-written corpora in the hidden-state space of a frozen LLM using a coherence-eliciting prompt, and compares the resulting distributions using RBF-MMD. To test coherence sensitivity and selectivity, we construct a counterfactual evaluation suite pairing graded coherence-degrading perturbations with meaning-preserving controls. CHORD selectively detects relation, discourse, structural, and mixture failures that perplexity, entropy, MAUVE, FBD, and MMD-based baselines either miss or cannot separate from benign rewriting. Factorial ablations show that representation is the primary source of coherence sensitivity, while RBF-MMD improves sample efficiency. Larger backbones capture finer-grained distinctions, but coherence prompting improves selectivity only when the backbone can follow the prompt. On unconditional generation and prefix continuation, CHORD yields model rankings that strongly align with human judgments of whether outputs make sense and appear human-written. Together, these results establish representation design as central to reliable distributional evaluation. Code: https://github.com/MAPS-research/CHORD. Experiments: https://github.com/MAPS-research/CHORD-Experiment.

Of the approximately 4,500 Oracle Bone Inscription (OBI) characters discovered from the Shang dynasty, only about 1,600 have been deciphered. Many computational approaches compare OBI with glyphs from one historical period at a time. However, during the evolution of Chinese characters, significant structural or semantic changes often occur in uncertain dynasties. A single-period reference may be insufficient when relevant forms change substantially between observed eras. Therefore, we propose the \textbf{Manifold-based Script Evolution Framework (MSEF)}, a framework that models the evolution series (OBI, Bronze, Seal, Clerical, Regular) of Chinese characters as the continual evolution of a manifold space. MSEF represents each character as an era-specific manifold point and learns continuous inter-era transition rules via Neural Ordinary Differential Equations. Both manifold space and transition dynamics can be trained end-to-end through character evolution pairs across any two eras.

In recent years, Sentiment Analysis (SA) and Emotion Recognition (ER) have been increasingly popular in the Bangla language, which is the seventh most spoken language throughout the entire world. However, the language is structurally complicated, which makes this field arduous to extract emotions in an accurate manner. Several distinct approaches such as the extraction of positive and negative sentiments as well as multiclass emotions, have been implemented in this field of study. Nevertheless, the extraction of multiple sentiments is an almost untouched area in this language. Which involves identifying several feelings based on a single piece of text. Therefore, this study demonstrates a thorough method for constructing an annotated corpus based on scrapped data from Facebook to bridge the gaps in this subject area to overcome the challenges. To make this annotation more fruitful, the context-based approach has been used. Bidirectional Encoder Representations from Transformers (BERT), a well-known methodology of transformers, have been shown the best results of all methods implemented. Finally, a web application has been developed to demonstrate the performance of the pre-trained top-performer model (BERT) for multi-label ER in Bangla.

Retrieval-Augmented Generation (RAG) enhances Large Language Models (LLMs) by retrieving relevant information from external knowledge bases to provide more accurate, contextually informed, and up-to-date responses. However, this reliance on external knowledge introduces significant security vulnerabilities, as many RAG systems (e.g., Google Search) rely on large and unsanitized data repositories (e.g., Reddit). In this paper, we unveil a novel threat in which attackers steer the RAG system's response by injecting malicious passages into its knowledge base. When a user's query contains attacker-specified trigger words, the RAG retrieves and refers to these malicious passages, enabling the attacker to steer the response without altering the user input or modifying the RAG weights. BadRAG operates in two phases: (i) malicious passages are optimized to be retrieved exclusively when trigger words appear in user queries; (ii) these passages are meticulously crafted to achieve adversarial generation objectives, including denial of service, sentiment manipulation, context leakage, and tool misuse. Our experiments show that injecting just 10 malicious passages (0.04\% of the external corpora) achieves a 98.2\% retrieval success rate and increases negative response rates from 0.22\% to 72\% for queries containing triggers.

Modern large language models increasingly require long contexts for reasoning and multi-document tasks, but attention's quadratic complexity creates a severe computational bottleneck. We present Block Sparse Flash Attention (BSFA), a drop-in replacement that accelerates long-context inference while preserving model quality. Unlike methods that predict importance before computing scores, BSFA computes exact query-key similarities to select the top-k most important value blocks for each query. By comparing per-block maximum scores against calibrated thresholds, we skip approximately 50% of the computation and memory transfers for pruned blocks. Our training-free approach requires only a one-time threshold calibration on a small dataset to learn the per-layer and per-head attention score distributions. We provide a CUDA kernel implementation that can be used as a drop-in replacement for FlashAttention. On Llama-3.1-8B, BSFA achieves up to 1.13x end-to-end speedup on LongBench with only a 1.1% accuracy drop, and up to 1.24x on Needle-in-a-Haystack retrieval at a 1% accuracy drop. The attention kernel itself accelerates by up to 1.38x. We compare BSFA against five recent sparse attention baselines (SpargeAttention, MInference, FlexPrefill, XAttention, and BLASST), and verify the method on Qwen2.5-7B and on A6000 and H100 GPUs. The implementation is available at https://github.com/Danielohayon/Block-Sparse-Flash-Attention.

Aligning large language models (LLMs) to serve users with heterogeneous and potentially conflicting preferences is a central challenge for personalized and trustworthy AI. We formalize an ideal notion of universal alignment through test-time scaling: for each prompt, the model produces $k\ge 1$ candidate responses and a user selects their preferred one. We introduce $(k,f(k))$-robust alignment, which requires the $k$-output model to have win rate $f(k)$ against any other single-output model, and asymptotic universal alignment (U-alignment), which requires $f(k)\to 1$ as $k\to\infty$. Our main result characterizes the optimal convergence rate: there exists a family of single-output policies whose $k$-sample product policies achieve U-alignment at rate $f(k)=\frac{k}{k+1}$, and no method can achieve a faster rate in general. We show that popular post-training methods, including Nash learning from human feedback (NLHF), can fundamentally underutilize the benefits of test-time scaling. Even though NLHF is optimal for $k=1$, sampling from the resulting (often deterministic) policy cannot guarantee win rates above $\tfrac{1}{2}$ except for an arbitrarily small slack. This stems from a lack of output diversity: existing alignment methods can collapse to a single majority-preferred response, making additional samples redundant. In contrast, our approach preserves output diversity and achieves the optimal test-time scaling rate. In particular, we propose a family of symmetric multi-player alignment games and prove that any symmetric Nash equilibrium policy of the $(k+1)$-player alignment game achieves the optimal $(k,\frac{k}{k+1})$-robust alignment. Finally, we provide theoretical convergence guarantees for self-play learning dynamics in these games and extend the framework to opponents that also generate multiple responses.

LLM agents are increasingly expected to operate as general-purpose systems that resolve real-world user requests, yet their dynamic scaling behavior in realistic environments remains poorly understood. In this paper, we systematically investigate two principal test-time scaling axes of LLM agents: sequential scaling through extended interaction and parallel scaling through trajectory sampling. We first introduce a realistic benchmark that provides one unified framework for evaluating LLM agents across search, coding, reasoning, and tool-use domains, more faithfully reflecting the heterogeneity of real-world deployments. Evaluating ten leading LLM agents reveals substantial performance degradation when transitioning from domain-specific evaluations to this realistic setting. Building on this foundation, we progressively scale test-time compute along fine-grained increments to characterize the performance upper bound. We find that neither scaling axis can consistently yield meaningful gains from additional test-time compute in realistic environments, a phenomenon we attribute to two fundamental limitations: the scaling plateau that bottlenecks sequential scaling and the verification gap that undermines parallel scaling. Code is publicly available at https://github.com/cxcscmu/General-AgentBench.

Audiovisual speech recognition (AVSR) combines acoustic and visual cues to improve transcription robustness under challenging conditions but remains out of reach for most under-resourced languages due to the lack of labeled video corpora for training. Synthetic visual data have been shown to be an effective augmentation strategy for addressing AV data scarcity. However, a more challenging scenario arises for languages such as Catalan, where no real audiovisual data are available for training. In this study, we investigate whether AVSR can be bootstrapped in such a zero-AV-resource setting, using synthetic visual data as the sole source of visual supervision. We synthesize over 700 hours of talking-head video and fine-tune a pre-trained AV-HuBERT model. On a manually annotated Catalan benchmark, our model achieves near state-of-the-art (SOTA) performance with much fewer parameters and training data than SOTA ASR systems such as Whisper-large-v3, outperforms an identically trained audio-only baseline, and preserves multimodal advantages under acoustic degradation. Scalable synthetic video thus offers a viable substitute for real recordings in zero-AV-resource AVSR.

Large language models (LLMs) frequently generate confident yet factually incorrect content when used for language generation (a phenomenon often known as hallucination). Retrieval augmented generation (RAG) tries to reduce factual errors by identifying information in a knowledge corpus and putting it in the context window of the model. While this approach is well-established for document-structured data, it is non-trivial to adapt it for Knowledge Graphs (KGs), especially for queries that require multi-node/multi-hop reasoning on graphs. We introduce UltRAG, a training-free KG-RAG recipe that combines LLM query generation, a fully inductive neural query executor, and LLM arbitration. This off-the-shelf composition achieves state-of-the-art results on Knowledge Graph Question Answering (KGQA) tasks without retraining the LLM or executor, while enabling language models to interface with Wikidata-scale graphs (116M entities, 1.6B relations) at comparable or lower costs. Our ablation studies indicate that these gains come from the full system design rather than from any single component.

We call query--key relevance absolute when its values lie on a fixed bounded scale, depend on neither competing keys nor sequence length, require no sequence-length-dependent calibration, and can all be zero. To realize this notion, we introduce screening, whose explicit threshold transforms bounded query--key similarities into relevance values, enabling exact rejection, empty selection, and direct inspection on a common scale. In a controlled comparison of 12 attention mechanisms on a matched Transformer backbone, only screening maintains both low long-context perplexity and robust retrieval beyond the training context; notably, it does so without inference-time scaling. Building on screening, we introduce Multiscreen, a language-model architecture composed of parallel gated screening tiles. Multiscreen retains these long-context gains while achieving greater parameter efficiency, stronger general zero-shot downstream performance, lower training cost at larger scales, and lower model-side time to first token than Transformer baselines. We further develop a normalization design that keeps Multiscreen training stable even at a learning rate of $1$ and show that an adapted version likewise stabilizes Transformer at the same learning rate.

Chain-of-thought (CoT) distillation transfers reasoning behaviors from a strong teacher to a smaller student, but prior work reports a capacity gap: distillation may fail when the teacher-student capability mismatch is large. We revisit the capacity gap from a practical perspective by re-examining commonly used experimental settings. Notably, we find that CoT distillation often degrades performance compared to the student's pre-distillation baseline, and that some settings used in prior work, while suitable for establishing the capacity gap as a phenomenon, do not reflect realistic deployment scenarios. Complementing prior work that establishes the capacity gap, we evaluate its practical impact under more realistic settings and find that it does not consistently dominate; stronger teachers tend to be preferable when candidate teachers differ substantially in performance. Our results offer practical guidance for selecting teacher-student pairs in CoT distillation.

Global structured pruning requires channels from different layers to compete under a shared sparsity budget, raising two coupled challenges: identifying which channels should be retained and making their scores comparable across layers. Raw channel scores can contain block-common scale that leaves within-block ordering unchanged but distorts model-wide competition. Our experiment indicates that similar layer-wise allocations can retain substantially different FFN channels, so layer allocation alone does not determine channel identity. Motivated by this separation, we introduce Global Relative Kinetic Utility (Global RKU), a label-free criterion that separates channel importance estimation from cross-layer comparison. Global RKU measures channel participation using a final-hidden-state activation-gradient signal, then applies block-relative normalization to mitigate block-common scale while preserving within-block ordering, requires only unlabeled calibration inputs, and produces a static pruning topology in a single calibration stage. Under questions-only calibration on Qwen-2.5-7B, RKU-GISP Mean3 margins are -0.98, +3.79, and +8.61 points at 30%, 40%, and 50% sparsity, respectively (average +3.81). Additional Qwen evaluations cover non-mathematical reasoning, recovery, held-out transfer, and physical deployment. Separately, replacing Wiki16K with questions-only Q16K improves RKU's Mean3 at every tested sparsity on Qwen, Llama, and Gemma. Our ablation study shows relative-normalization gains of 14.42 and 5.53 Mean3 points at 40% and 50% sparsity, respectively; the common-seed audit is positive in all 27 seed-task comparisons.

Adaptive optimizers such as Adam are standard for training Transformers, but storing gradient first and second moments incurs substantial memory overhead. We introduce PowerStep, a memory-efficient optimizer that achieves coordinate-wise adaptivity without storing second-moment statistics. Motivated by $\ell_p$-norm steepest descent, PowerStep applies a signed-power transform directly to one momentum buffer. We establish a finite-horizon stationarity bound for exact, unregularized updates, with an $O(1/\sqrt{T})$ term and a noise-dependent residual. Experiments on Transformers from 124M to 235B parameters show competitive validation quality while halving $\texttt{fp32}$ optimizer-state memory relative to AdamW. Combined with uniform $\texttt{int8}$ quantization, PowerStep remains numerically stable and reduces optimizer-state memory by $\sim8\times$ compared to $\texttt{fp32}$ AdamW. PowerStep thus provides a simple, memory-efficient alternative for large-scale training.

Automated code documentation is essential for modern software development, providing the contextual grounding that both human developers and coding agents rely on to navigate large codebases. Existing repository-level approaches process components independently, causing redundant retrieval and conflicting descriptions across documents while producing outputs that lack hierarchical structure. Therefore, we propose MemDocAgent, a long-horizon agentic framework that generates documentation within a single, integrated context spanning the entire repository. It combines two components: (i) Dependency-Aware Traversal Guiding that predetermines a traversal order respecting dependency and granularity hierarchies; (ii) Memory-Guided Agentic Interaction, in which the agent interacts with RepoMemory, a shared memory accumulating prior work traces through read, write, and verify operations. Through an in-depth multi-criteria evaluation, MemDocAgent achieves the best performance over both open- and closed-source baselines and demonstrates practical applicability in real software development workflows.

Frontier language models sometimes recognize that they are under evaluation and adjust their behavior which can undermine validity of benchmark results. Yet the field studies it without a shared foundation, conflating flaws of the evaluation with capabilities of the model, and detection with behavioral response. We ground evaluation awareness in social psychology, decomposing it into an environment component and a model component that separates recognition from propensity. We operationalize the environment component through eight categorized trigger factors, such as placeholder entities and grading-style output formats, and study recognition and behavior through chain-of-thought monitoring. Across nine frontier models and four benchmarks, recognition rates depend on the specific pairing of model and benchmark. Recognition rarely associates with behavioral change, and when it does, the direction depends on the type of evaluation perceived. Models are also more sensitive to safety than capability evaluations, placing safety benchmark validity at greater risk. To study which factors each model is sensitive to and how they interact, we propose \textbf{EvalAwareBench}, a factor-controlled benchmark of 100 paired safety-capability tasks where each of the eight factors can be independently toggled, varying evaluative signals while holding the underlying request fixed. Through EvalAwareBench, we find that no single factor uniformly affects all models, but stacking factors progressively raises evaluation awareness across all of them. Our framework and EvalAwareBench provide the tools to measure, attribute, and mitigate evaluation awareness, building the foundation for future solutions.

Reinforcement learning with verifiable rewards (RLVR) trains reasoning models without labeled trajectories, using groups of verifier-scored rollouts to explore alternative reasoning paths. Limited rollout diversity is a central bottleneck, typically addressed through adjustments to temperature, prefixes, or rollout selection. We identify the first token of the response as a structurally distinct target for diversification, largely overlooked in prior work. We find that the first-token distribution is sharply concentrated and only weakly related to downstream correctness, as lower-probability candidates can yield similarly accurate responses. Diversifying the first token can therefore broaden the reasoning paths explored within each rollout group with little loss in response quality. Motivated by this observation, we introduce REFT (Rollout Exploration with First-Token Diversification), a lightweight modification to RLVR. REFT samples first tokens uniformly from the policy's top-$N$ candidates and allocates rollouts evenly across the sampled tokens, leaving the rest of the pipeline unchanged. We evaluate REFT on eight models spanning multiple architectures and sizes (0.5B-14B), with mathematical reasoning and code-generation tasks under GRPO and DAPO. Across these settings, REFT consistently improves Pass@1, Pass@8, and Pass@64. It also outperforms competing diversification methods at every evaluated budget, incurring the lowest rollout cost.

Frontier reasoning models are produced by post-training base language models with reinforcement learning. Recent work has challenged this by showing that sampling from a sharpened version of the base model's distribution, a so-called power distribution, elicits comparable reasoning without additional training, curated datasets, or verifiers. However, making this method practical requires efficiently sampling from the power distribution. A sampler needs to "mix" to the power distribution, which necessitates moving between modes of the target distribution; intuitively, e.g., trying different reasoning strategies. The samplers proposed in prior works repeatedly select a "cut" position in the current reasoning trace uniformly at random and resample the suffix from that position onward. However, reasoning traces typically contain a few consequential decisions (e.g., the choice of proof strategy or algorithm), and we observe that a uniformly chosen cut tends to rewrite local details rather than revisit decision points. We introduce an algorithm (Entropy-Cut Metropolis-Hastings) that uses the base model's next-token entropy as a proxy to identify key decision points and resample from those positions. We empirically verify that entropy jumps are a useful proxy for decision points and, in a stylized model of reasoning, prove that our method's mixing time scales with the number of decisions in a trace rather than with the number of tokens, which can be much larger. Across MATH500, HumanEval, GPQA Diamond, and AIME26, our method consistently improves over baselines and RL-trained models, including best-of-N at the same token budget.

Rubric-based reinforcement learning (RL) uses an LLM-as-a-Judge (LaaJ) to score model outputs according to rubrics as rewards. However, policy models may exploit latent biases in the judge, leading to reward hacking and ineffective or unsafe training outcomes. In real-world rubric-based RL, such hacking behaviors are often subtle and entangled with multiple judge biases, making them difficult to analyze, detect, and mitigate. In this paper, we introduce CHERRL, a Controllable Hacking Environment for Rubric-based RL. By injecting known biases into LaaJ, CHERRL enables stable reproduction of reward hacking, explicit observation of reward divergence, and identification of hacking onset. This provides a clean experimental testbed for studying the mechanisms and mitigations of reward hacking in rubric-based RL. To demonstrate its utility, we analyze different judge biases from the perspectives of discoverability and exploitability, and explore an agent for automatically detecting reward hacking onset from training logs. The code and environment are publicly available at https://github.com/THUAIS-Lab/CHERRL.

Maintainability is a core dimension of software engineering, shaping how code is written, reviewed, and developed over time. While coding agents have demonstrated strong performance on single-issue tasks, it remains unclear how maintainable their code is when future agents build on top of it, potentially leading to compounding downstream effects. We investigate how agent code compares to human code in these maintenance settings, presenting CodeThread, a framework to construct controlled experiments from repository-level coding benchmarks. Applying CodeThread to four frontier coding agents and four benchmarks, we find that agents are less effective at resolving tasks when building on agent code compared to human code, with task resolve rate drops of up to 13.1%. Regression analysis reveals that many traditional software engineering maintainability metrics do not explain this difference. Instead, the clearest signals are subtler behavioral differences in agent code, such as changes to input validation and error handling, along with differences in downstream code size and task difficulty. These findings highlight the need to evaluate these systems not only by immediate task resolution but also by code maintainability, and point to potential sources of downstream errors introduced by agent code.