AI 每日进展速报 / Daily AI Digest - 2026-03-13

图像生成/编辑 / Image Generation/Editing

arXiv

• [2026-03-08] GRD-Net: Generative-Reconstructive-Discriminative Anomaly Detection with Region of Interest Attention Module 📖5 🆕NEW

  • 赛道归属: 工业视觉异常检测（缺陷检测/定位）
  • 核心创新点: 提出GRD-Net，将生成-重建-判别三类信号在同一框架内协同建模，并引入ROI注意力模块把模型容量聚焦到潜在缺陷区域，减少依赖“生成图-原图差分+后处理”的传统流水线。通过端到端的区域级关注与判别约束，提升缺陷定位的稳定性与可解释性。
  • 一句话总结: 用ROI注意力把生成式重建与判别式学习有效结合，增强工业表面缺陷检测的定位精度与鲁棒性。
    Track: Industrial visual anomaly detection (defect detection/localization)
    Core innovation: GRD-Net unifies generative, reconstructive, and discriminative cues in one model and adds an ROI attention module to concentrate learning on likely defect regions, reducing reliance on post-hoc “reconstruction difference + blob/image processing” pipelines. End-to-end region-focused supervision improves localization stability and interpretability.
    One-sentence summary: It boosts industrial defect localization by tightly coupling generative reconstruction with discriminative learning under ROI-focused attention.

• [2026-03-12] ZeroSense:How Vision matters in Long Context Compression 📖2 🆕NEW

  • 赛道归属: 多模态长上下文压缩与评测（视觉-文本压缩/VTC）
  • 核心创新点: 提出ZeroSense评测框架，将“下游任务表现”与“文本保真度”解耦，针对MLLM强语言先验导致的虚假高分问题，设计更能度量压缩后文本是否被真实保留的评价协议。通过强调视觉渲染在压缩中的作用与失真来源，提供更可靠的长上下文压缩诊断工具。
  • 一句话总结: 为VTC类长上下文压缩建立更可信的“保真度”评测，避免被MLLM语言先验掩盖的文本丢失。
    Track: Multimodal long-context compression & evaluation (visual-text compression, VTC)
    Core innovation: ZeroSense introduces an evaluation framework that decouples downstream-task success from text-preservation fidelity, addressing inflated scores caused by MLLMs’ strong linguistic priors. It provides protocols that more directly test whether the compressed representation truly retains the original text.
    One-sentence summary: It makes VTC evaluation more trustworthy by explicitly measuring preservation fidelity rather than relying on downstream performance.

• [2026-03-10] Towards Unified Multimodal Interleaved Generation via Group Relative Policy Optimization 📖1 🆕NEW

  • 赛道归属: 多模态交错生成（Interleaved Generation）/ 强化学习后训练
  • 核心创新点: 提出Group Relative Policy Optimization（GRPO）用于统一多模态模型的后训练，在不依赖大规模交错图文数据的前提下，通过组内相对偏好优化解锁“图文交错输出”能力。采用warm-up与基于相对奖励的策略更新，降低对绝对标注/奖励标定的依赖并提升训练稳定性。
  • 一句话总结: 用相对策略优化的RL后训练，让现有统一模型在缺少交错数据时也能学会高质量图文交错生成。
    Track: Multimodal interleaved generation / RL post-training
    Core innovation: The work proposes GRPO to post-train unified vision-language models for interleaved multimodal outputs without large-scale interleaved datasets, optimizing relative preferences within groups instead of relying on absolute reward calibration. A warm-up plus relative-reward policy updates improves stability and unlocks interleaved generation.
    One-sentence summary: It enables strong interleaved image-text generation via RL post-training even when interleaved supervision data is scarce.

• [2026-03-12] GRADE: Benchmarking Discipline-Informed Reasoning in Image Editing 🆕NEW

  • 赛道归属: 图像编辑评测基准（推理驱动/学科知识约束）
  • 核心创新点: 提出GRADE基准，将图像编辑从自然图像+浅层常识扩展到“学科知识+结构化约束”的推理评测，覆盖10个学科领域、520个精心样本，用于检验模型在遵循领域规则下的编辑与推理一致性。通过学科化约束设计，暴露统一多模态模型在可控编辑与知识推理耦合上的短板。
  • 一句话总结: GRADE用学科约束把图像编辑评测提升到“可验证推理”的层级，推动更可靠的知识驱动编辑模型。
    Track: Image editing benchmark (reasoning- and discipline-constrained)
    Core innovation: GRADE is a benchmark that evaluates image editing under structured, domain-specific constraints across 10 academic disciplines (520 curated samples), testing whether models can reason and edit consistently with discipline rules rather than only commonsense on natural images. Its constraint design surfaces failures in coupling controllable edits with knowledge-based reasoning.
    One-sentence summary: It upgrades image-editing evaluation to discipline-informed, verifiable reasoning, guiding more reliable knowledge-grounded editors.

• [2026-03-12] The Latent Color Subspace: Emergent Order in High-Dimensional Chaos 🆕NEW

  • 赛道归属: 文生图可解释性与可控生成（潜空间解析/控制）
  • 核心创新点: 解析FLUX.1的VAE潜空间颜色表征，提出Latent Color Subspace（LCS），揭示高维潜变量中涌现出的近似HSL（色相/饱和度/明度）结构，并用可预测、可操控的实验验证该解释。该工作把“颜色语义如何编码”从黑箱经验调参推进到可操作的潜空间几何理解。
  • 一句话总结: 通过发现并验证潜空间的颜色子空间结构，为文生图提供更可解释、更精细的颜色控制手段。
    Track: Text-to-image controllability & interpretability (latent-space analysis/control)
    Core innovation: The paper identifies a Latent Color Subspace in FLUX.1’s VAE latent space, showing an emergent structure aligned with HSL (hue/saturation/lightness) and validating it via predictive and controllable manipulations. This turns color control from heuristic prompting into actionable latent-space geometry.
    One-sentence summary: It enables more interpretable and fine-grained color control in T2I by uncovering a structured color subspace in the VAE latent.

• [2026-03-12] Trust Your Critic: Robust Reward Modeling and Reinforcement Learning for Faithful Image Editing and Generation 🆕NEW

  • 赛道归属: 图像生成/编辑的强化学习与奖励建模（对齐与保真度）
  • 核心创新点: 提出FIRM（Faithful Image Reward Modeling），针对现有reward model易幻觉、打分噪声大而误导RL优化的问题，构建更鲁棒的“评论家”训练与校准框架，使奖励更贴合编辑指令与视觉事实。通过提升奖励可靠性，改善RL驱动的图像编辑与T2I生成的忠实度与稳定性。
  • 一句话总结: 先把“critic”做可信，再做RL对齐，从根源提升图像编辑/生成的指令遵循与事实一致性。
    Track: RL & reward modeling for image generation/editing (alignment & faithfulness)
    Core innovation: FIRM builds robust reward models to mitigate hallucinated/noisy scoring that misguides RL for image editing and T2I generation, introducing training/calibration strategies so the critic better reflects instruction adherence and visual faithfulness. More reliable rewards translate into more stable and faithful RL optimization.
    One-sentence summary: By making the critic trustworthy, it improves RL-aligned image editing/generation in instruction following and factual consistency.

• [2026-03-12] GlyphBanana: Advancing Precise Text Rendering Through Agentic Workflows 🆕NEW

  • 赛道归属: 文生图文字渲染（复杂字符/公式）与Agentic工作流
  • 核心创新点: 提出GlyphBanana及其基准，面向复杂文字与数学公式渲染，采用agentic工作流把“生成—检查—纠错/重试”流程化，以缓解模型在分布外字符与公式指令上的跟随失败。通过基准+工作流的组合，系统性提升精确排版与符号级正确率的可评测、可迭代能力。
  • 一句话总结: 用代理式闭环工作流攻克复杂文本/公式渲染的长尾难题，并提供可量化对比的专用基准。
    Track: Text rendering in image generation (complex glyphs/formulas) & agentic workflows
    Core innovation: GlyphBanana introduces a benchmark for complex character and math formula rendering and an agentic workflow that operationalizes generate–verify–correct/retry loops to handle out-of-distribution prompts where instruction following breaks down. The benchmark+workflow pairing enables measurable, iterative improvements in symbol-level accuracy.
    One-sentence summary: It tackles long-tail precise text/formula rendering via a closed-loop agentic pipeline with a dedicated benchmark for evaluation.

• [2026-03-12] Hoi3DGen: Generating High-Quality Human-Object-Interactions in 3D 🆕NEW

  • 赛道归属: 文生3D（人-物交互）/ 3D生成
  • 核心创新点: 提出Hoi3DGen，从文本生成高质量带纹理的人-物交互网格，针对基于T2I蒸馏常见的Janus问题与交互数据稀缺导致的不忠实，构建更适配交互几何与接触关系的生成框架。通过显式建模交互与提升数据/约束质量，实现更符合描述的3D交互姿态与外观。
  • 一句话总结: 面向AR/XR与游戏的关键需求，Hoi3DGen把“可用且可信”的人-物交互3D生成质量推上一个台阶。
    Track: Text-to-3D generation (human-object interaction)
    Core innovation: Hoi3DGen generates high-quality textured meshes of human-object interactions from text, addressing Janus artifacts common in T2I distillation and poor prompt faithfulness caused by scarce interaction data, via an interaction-aware generation framework better aligned with contact/geometry constraints. This yields more text-consistent HOI poses and appearances.
    One-sentence summary: It advances practical HOI 3D generation for AR/XR and games with higher fidelity and better prompt adherence.

• [2026-03-12] EvoTok: A Unified Image Tokenizer via Residual Latent Evolution for Visual Understanding and Generation 🆕NEW

  • 赛道归属: 统一视觉Tokenizer（理解-生成一体化）/ 多模态基础设施
  • 核心创新点: 提出EvoTok，通过Residual Latent Evolution在同一tokenizer中同时满足“理解所需的语义粒度”和“生成所需的像素细节”，缓解两类监督在同一表征上互相干扰或在双空间中不一致的问题。以演化式残差潜变量机制实现从粗到细的可兼容编码，提升MLLM在理解与生成任务间的统一性。
  • 一句话总结: EvoTok用“残差潜变量演化”打通理解与生成的粒度鸿沟，为统一多模态模型提供更一致的视觉离散表征。
    Track: Unified visual tokenizer for multimodal understanding & generation
    Core innovation: EvoTok proposes a single tokenizer using Residual Latent Evolution to bridge the granularity gap: semantic abstractions for understanding and fine-grained details for generation, avoiding interference from forcing both supervisions onto one representation or inconsistencies from splitting into separate spaces. The coarse-to-fine residual evolution yields compatible codes for both regimes.
    One-sentence summary: It provides a more consistent visual discretization that better unifies multimodal understanding and image generation in one model stack.

• [2026-03-12] Single Pixel Image Classification using an Ultrafast Digital Light Projector 🆕NEW

  • 赛道归属: 计算成像+超高速视觉分类（单像素成像/SPI）
  • 核心创新点: 将单像素成像（SPI）与低复杂度机器学习模型结合，并利用microLED-on-CMOS数字光投影实现多kHz级模式投射与采集，从实验上展示超高速帧率下的图像分类。该路线用硬件-算法协同在“极少传感器/高速度”约束下实现实时识别，为边缘与高速场景提供新范式。
  • 一句话总结: 通过超高速DLP驱动的SPI，把图像分类推进到多kHz实时水平，展示了计算成像在高速机器视觉中的潜力。
    Track: Computational imaging + ultrafast visual classification (single-pixel imaging)
    Core innovation: The project combines single-pixel imaging with a low-complexity ML classifier and a microLED-on-CMOS digital light projector to achieve multi-kHz pattern projection/acquisition, experimentally demonstrating ultrafast image classification. This hardware–algorithm co-design enables recognition under extreme sensing and speed constraints.
    One-sentence summary: It shows multi-kHz real-time classification using SPI enabled by an ultrafast projector, highlighting a promising path for high-speed machine vision.

视频生成/编辑 / Video Generation/Editing

arXiv

• [2026-03-09] Foley-Flow: Coordinated Video-to-Audio Generation with Masked Audio-Visual Alignment and Dynamic Conditional Flows 📖1 🆕NEW

  • 赛道归属: 视频到音频生成（Video-to-Audio Generation）
  • 核心创新点: 提出带“遮蔽式音视对齐”的对齐机制，在生成过程中以局部/片段级方式强化语义与节奏对齐，减少仅靠全局视频条件导致的错配；引入动态条件流（conditional flows）以更细粒度地调度不同时间段的条件信息，实现更协调的音频段落生成。
  • 一句话总结: 在不依赖传统“两阶段对比对齐+全局引导”的前提下，显著提升视频驱动拟音的时序对齐与一致性。
  • Track: Video-to-Audio Generation
  • Core innovation: Introduces masked audio-visual alignment to enforce local/segment-level semantic and rhythmic correspondence during generation, mitigating mismatches from purely global video conditioning; uses dynamic conditional flows to adaptively route conditioning over time for more coordinated audio segment synthesis.
  • One-sentence takeaway: Improves temporal alignment and coherence for video-driven Foley generation beyond the standard two-stage contrastive-alignment pipeline.

• [2026-03-12] EVATok: Adaptive Length Video Tokenization for Efficient Visual Autoregressive Generation 🆕NEW

  • 赛道归属: 视频生成（视觉自回归）/ 视频Tokenizer与推理效率优化
  • 核心创新点: 提出自适应长度的视频离散化方案，根据时空块的动态复杂度分配不同数量的token，避免静态/重复片段的token浪费；在保证重建质量的同时显著缩短AR生成序列长度，从而降低下游自回归生成的计算与延迟。
  • 一句话总结: 用“按内容分配token”的可变长度tokenization，在质量-成本之间取得更优折中，直接加速自回归视频生成。
  • Track: Autoregressive Video Generation / Video Tokenization & Inference Efficiency
  • Core innovation: Proposes adaptive-length video tokenization that allocates tokens based on spatiotemporal complexity, reducing waste on static/repetitive segments; shortens AR token sequences while preserving reconstruction fidelity, cutting downstream generation compute and latency.
  • One-sentence takeaway: Variable-length, content-adaptive tokenization delivers a better quality–cost trade-off and speeds up autoregressive video generation.

• [2026-03-12] DreamVideo-Omni: Omni-Motion Controlled Multi-Subject Video Customization with Latent Identity Reinforcement Learning 🆕NEW

  • 赛道归属: 视频生成（可控生成/个性化定制，多主体）
  • 核心创新点: 统一框架同时解决多主体身份保持与多粒度运动控制，通过“潜空间身份强化”的强化学习策略抑制身份漂移；支持全局到局部的全方位运动（omni-motion）控制，降低控制歧义并提升多主体协同一致性。
  • 一句话总结: 在多主体视频定制中同时把“像谁”和“怎么动”做得更可控、更稳定。
  • Track: Controllable Video Generation / Personalized Multi-subject Customization
  • Core innovation: A unified framework that jointly tackles multi-subject identity preservation and multi-granularity motion control, using latent identity reinforcement learning to reduce identity degradation; enables omni-motion control from global to fine-grained levels with less control ambiguity and better multi-subject consistency.
  • One-sentence takeaway: Makes multi-subject video customization more reliable by stabilizing identity while enabling precise, multi-scale motion control.

• [2026-03-12] DVD: Deterministic Video Depth Estimation with Generative Priors 🆕NEW

  • 赛道归属: 视频深度估计（生成式先验 + 判别式回归）
  • 核心创新点: 首次将预训练视频扩散模型确定性地改造成单次前向的深度回归器，避免扩散采样带来的随机几何幻觉与尺度漂移；利用扩散时间步作为结构锚点并结合生成先验，实现更稳定的时序几何一致性，同时减少对大规模标注数据的依赖。
  • 一句话总结: 把“生成式视频先验”变成可确定输出的深度估计器，兼顾稳定性与数据效率。
  • Track: Video Depth Estimation (Generative Priors + Deterministic Regression)
  • Core innovation: Deterministically adapts a pretrained video diffusion model into a single-pass depth regressor, avoiding stochastic geometric hallucinations and scale drift from sampling; leverages diffusion timestep as a structural anchor to enforce temporally consistent geometry while reducing reliance on massive labeled data.
  • One-sentence takeaway: Turns generative video priors into stable, deterministic depth estimation with improved consistency and label efficiency.

• [2026-03-12] A Two-Stage Dual-Modality Model for Facial Emotional Expression Recognition 🆕NEW

  • 赛道归属: 多模态理解（视频表情识别/情感计算）
  • 核心创新点: 两阶段流水线先提升人脸定位与帧级稳定性，再进行音视频双模态融合分类，以缓解野外视频中的姿态尺度变化、运动模糊与相邻帧抖动；通过分阶段建模将“检测/对齐误差”与“表情判别”解耦，提高鲁棒性。
  • 一句话总结: 用两阶段+音视频融合的工程化设计，提升复杂真实场景下的帧级表情识别稳定性。
  • Track: Multimodal Understanding (Video Facial Expression Recognition)
  • Core innovation: A two-stage pipeline that first improves face localization and frame-level stability, then performs audio-visual fusion for expression classification to handle pose/scale variation, motion blur, and temporal jitter; decouples alignment noise from expression discrimination for robustness in-the-wild.
  • One-sentence takeaway: A pragmatic two-stage AV approach that boosts frame-level expression recognition robustness in unconstrained videos.

• [2026-03-12] FlashMotion: Few-Step Controllable Video Generation with Trajectory Guidance 🆕NEW

  • 赛道归属: 视频生成（少步扩散蒸馏/轨迹可控生成）
  • 核心创新点: 面向轨迹引导的可控视频生成，提出适配少步生成的训练/蒸馏策略，避免将通用few-step蒸馏直接套用导致的轨迹控制失真；在极少去噪步数下仍保持对预定义运动轨迹的精确跟随，显著降低推理冗余与算力开销。
  • 一句话总结: 把“轨迹可控”从多步扩散带到few-step推理，实现更快且不丢控制精度的视频生成。
  • Track: Controllable Video Generation (Few-step Diffusion/Distillation with Trajectory Guidance)
  • Core innovation: Designs a few-step generation/distillation scheme tailored to trajectory-guided control, addressing control degradation when naively applying generic few-step distillation; preserves accurate adherence to predefined motion trajectories under very few denoising steps, reducing inference redundancy and compute.
  • One-sentence takeaway: Enables fast trajectory-controllable video generation with minimal steps while retaining control fidelity.

• [2026-03-12] Cornserve: A Distributed Serving System for Any-to-Any Multimodal Models 🆕NEW

  • 赛道归属: 推理优化/系统（多模态Any-to-Any模型分布式Serving）
  • 核心创新点: 针对Any-to-Any多模态请求在计算图中“路径可变”的特点，提出通用分布式服务系统，对不同模态组合进行组件级弹性伸缩与调度；通过面向组件的资源管理与路由，提升吞吐、降低尾延迟，并适配异构扩展特性。
  • 一句话总结: 为任意输入输出组合的多模态大模型提供可落地的分布式服务框架，解决路径多样带来的调度难题。
  • Track: Inference Optimization / Serving Systems for Any-to-Any Multimodal Models
  • Core innovation: A distributed serving system for Any-to-Any models where requests traverse different computation-graph paths depending on modality I/O; enables component-wise elastic scaling and scheduling/routing to improve throughput and tail latency under heterogeneous scaling behaviors.
  • One-sentence takeaway: Makes deployment of Any-to-Any multimodal models practical by systematizing routing and elastic scaling across variable execution paths.

• [2026-03-12] Coarse-Guided Visual Generation via Weighted h-Transform Sampling 🆕NEW

  • 赛道归属: 生成采样算法（扩散模型训练免/引导采样）/ 图像与视频生成
  • 核心创新点: 提出加权h-transform采样，将“粗参考（退化/低保真）”作为采样过程中的概率变换式引导信号，在无需额外训练与成对数据的情况下实现从粗到细的生成；相比常见训练免引导，改进了引导强度与稳定性的权衡，降低伪影与偏移。
  • 一句话总结: 用更 principled 的采样分布变换，把粗条件更稳定地注入扩散采样，实现训练免的高质量细化生成。
  • Track: Diffusion Sampling Algorithms (Training-free Guided Generation) / Image & Video Generation
  • Core innovation: Proposes weighted h-transform sampling to inject degraded/low-fidelity coarse references as a principled distribution-transform guidance during diffusion sampling, avoiding extra training and paired data; improves the guidance–stability trade-off, reducing artifacts and drift versus prior training-free guidance.
  • One-sentence takeaway: A training-free, more stable coarse-to-fine diffusion sampling method that better leverages coarse references for high-quality refinement.

• [2026-03-12] HomeSafe-Bench: Evaluating Vision-Language Models on Unsafe Action Detection for Embodied Agents in Household Scenarios 🆕NEW

  • 赛道归属: 多模态理解与评测（具身智能安全/视频VLM基准）
  • 核心创新点: 构建面向家庭场景具身体的“不安全动作检测”评测基准，强调动态过程与时序因果（而非静态危害识别）；系统性覆盖家庭任务中的风险类型，用于暴露VLM在延迟感知与常识安全推理上的短板。
  • 一句话总结: 用更贴近真实机器人部署的动态视频基准，推动VLM从“看见危险物”走向“识别危险行为”。
  • Track: Multimodal Understanding & Evaluation (Embodied Safety / Video VLM Benchmark)
  • Core innovation: Introduces a benchmark for unsafe action detection in household embodied-agent scenarios, emphasizing dynamic, temporal action risk rather than static hazard recognition; provides systematic coverage of household risk types to reveal VLM weaknesses in perception latency and commonsense safety reasoning.
  • One-sentence takeaway: A realistic video-centric safety benchmark that shifts evaluation toward detecting unsafe behaviors relevant to household robots.

• [2026-03-12] Controllable Egocentric Video Generation via Occlusion-Aware Sparse 3D Hand Joints 🆕NEW

  • 赛道归属: 视频生成（第一视角/手部动作可控生成，3D条件）
  • 核心创新点: 用稀疏3D手部关节作为可控条件，并显式建模遮挡（occlusion-aware），在严重第一视角遮挡下仍保持3D一致的精细手部关节运动；相比2D轨迹或隐式姿态条件，减少空间歧义与手部幻觉伪影，并提升跨主体/跨形体的可泛化控制。
  • 一句话总结: 以“可遮挡感知的稀疏3D手部骨架”作为控制信号，让第一视角手部视频生成更真实、更一致、更可控。
  • Track: Controllable Egocentric Video Generation (3D Hand Pose Conditioning)
  • Core innovation: Uses occlusion-aware sparse 3D hand joints as control signals to maintain 3D-consistent fine-grained articulation under severe egocentric occlusions; reduces spatial ambiguity and hallucinated artifacts compared to 2D trajectories or implicit pose cues, improving cross-embodiment generalization.
  • One-sentence takeaway: Sparse, occlusion-aware 3D hand-joint conditioning enables more realistic and controllable egocentric hand video generation.

HuggingFace Models

• Lightricks/LTX-2.3 🆕NEW

• unsloth/LTX-2.3-GGUF 🆕NEW

语言大模型 / Large Language Models

arXiv

• [2026-03-09] Gradually Excavating External Knowledge for Implicit Complex Question Answering 📖7 🆕NEW

  • 赛道归属: 开放域复杂问答 / 检索增强生成（RAG）与知识挖掘
  • 核心创新点: 提出“渐进式外部知识挖掘”框架，将隐式复杂问题拆解为可逐步验证与扩展的子目标，通过多轮检索/生成迭代补全LLM缺失或过时的领域知识，并缓解一次性生成导致的覆盖不全。

  • 一句话总结: 通过渐进式引入与校验外部知识，让LLM在开放域隐式复杂问答中更全面、更可靠。

  • Track: Open-domain complex QA / Retrieval-Augmented Generation (RAG) & knowledge excavation

  • Core innovation: Proposes a gradual external-knowledge excavation framework that iteratively decomposes implicit complex questions into verifiable sub-goals, alternating retrieval and generation to fill missing/outdated knowledge and overcome one-shot incompleteness.
  • One-sentence summary: Improves coverage and reliability of LLMs on implicit open-domain complex QA by progressively retrieving and validating external knowledge.

• [2026-03-11] AI Psychometrics: Evaluating the Psychological Reasoning of Large Language Models with Psychometric Validities 📖4 🆕NEW

  • 赛道归属: 大模型评测 / 心理推理与可解释性评估（Psychometrics）
  • 核心创新点: 将心理测量学的效度框架（如结构效度、聚合/区分效度等）系统引入LLM心理推理评估，把“像不像人”从主观印象转为可检验的测量学指标与实验范式。

  • 一句话总结: 用心理测量学的“效度”体系为LLM心理推理能力提供更可信、可解释的评估路径。

  • Track: LLM evaluation / Psychological reasoning & interpretability (psychometrics)

  • Core innovation: Introduces psychometric validity frameworks (e.g., construct, convergent/divergent validity) to LLM psychological-reasoning evaluation, turning subjective human-likeness into testable measurement criteria and protocols.
  • One-sentence summary: Provides a validity-driven, more interpretable methodology to evaluate LLMs’ psychological reasoning.

• [2026-03-06] Abductive Reasoning with Syllogistic Forms in Large Language Models 📖2 🆕NEW

  • 赛道归属: 推理能力评测 / 溯因推理（Abductive Reasoning）与逻辑形式测试
  • 核心创新点: 以三段论形式系统构造溯因推理任务，区分演绎有效性与“最佳解释”式推断，分析LLM在信念偏差等现象下的表现，从而更贴近人类真实推理机制而非仅用演绎标准苛责。

  • 一句话总结: 用结构化三段论溯因任务更公平地刻画LLM与人类在“解释性推理”上的能力与偏差。

  • Track: Reasoning evaluation / Abductive reasoning with formal logic tasks

  • Core innovation: Builds syllogism-form abductive reasoning benchmarks that separate deductive validity from “best-explanation” inference, enabling analysis of belief-bias-like behaviors in a setting closer to human reasoning than pure deduction.
  • One-sentence summary: Offers a structured way to assess LLMs’ explanatory (abductive) reasoning and its human-like biases beyond deductive-only tests.

• [2026-03-11] The Unlearning Mirage: A Dynamic Framework for Evaluating LLM Unlearning 📖1 🆕NEW

  • 赛道归属: 安全与合规 / 大模型遗忘（Unlearning）评测与鲁棒性
  • 核心创新点: 提出动态评测框架，针对实体别名、多跳推理、提示改写等“轻微扰动可恢复”攻击面系统生成查询与对抗测试，揭示静态基准导致的“遗忘有效性幻觉”。

  • 一句话总结: 通过动态、对抗式评测把LLM遗忘从“看起来忘了”推进到“在真实攻击下也忘得住”。

  • Track: Safety & compliance / LLM unlearning evaluation and robustness

  • Core innovation: Proposes a dynamic evaluation framework that systematically generates adversarial query variants (aliasing, multi-hop, paraphrases) to expose recoverability vulnerabilities missed by static benchmarks, debunking the “unlearning mirage.”
  • One-sentence summary: Moves unlearning evaluation toward real-world robustness by testing whether “forgotten” information remains unrecoverable under adaptive attacks.

• [2026-03-06] Supporting Artifact Evaluation with LLMs: A Study with Published Security Research Papers 📖1 🆕NEW

  • 赛道归属: 科研工具与自动化 / 可复现性与工件评审（Artifact Evaluation）辅助
  • 核心创新点: 以已发表安全论文为对象，研究LLM在工件评审中的可用性：从论文与工件描述中抽取可复现步骤、依赖与风险点，辅助生成检查清单与复现实验计划，以降低人工AE成本并提升规模化能力。

  • 一句话总结: 证明LLM可作为“AE助理”提升安全研究复现检查的效率与一致性。

  • Track: Research tooling & automation / Reproducibility and artifact evaluation assistance

  • Core innovation: Studies LLMs as AE assistants on published security papers by extracting reproducibility steps, dependencies, and pitfalls from paper/artifact descriptions, and generating structured checklists and execution plans to reduce manual AE burden.
  • One-sentence summary: Shows how LLMs can scale and standardize artifact evaluation for security research reproducibility.

• [2026-03-12] ZeroSense:How Vision matters in Long Context Compression 📖2 🆕NEW

  • 赛道归属: 多模态长上下文压缩与评测（视觉-文本压缩/VTC）
  • 核心创新点: 提出ZeroSense评测框架，将“下游任务表现”与“文本保真度”解耦，针对MLLM强语言先验导致的虚假高分问题，设计更能度量压缩后文本是否被真实保留的评价协议。通过强调视觉渲染在压缩中的作用与失真来源，提供更可靠的长上下文压缩诊断工具。
  • 一句话总结: 为VTC类长上下文压缩建立更可信的“保真度”评测，避免被MLLM语言先验掩盖的文本丢失。
    Track: Multimodal long-context compression & evaluation (visual-text compression, VTC)
    Core innovation: ZeroSense introduces an evaluation framework that decouples downstream-task success from text-preservation fidelity, addressing inflated scores caused by MLLMs’ strong linguistic priors. It provides protocols that more directly test whether the compressed representation truly retains the original text.
    One-sentence summary: It makes VTC evaluation more trustworthy by explicitly measuring preservation fidelity rather than relying on downstream performance.

• [2026-03-11] Counterweights and Complementarities: The Convergence of AI and Blockchain Powering a Decentralized Future 📖2 🆕NEW

  • 赛道归属: AI×区块链 / 去中心化AI治理与基础设施（观点与框架）
  • 核心创新点: 从“AI趋向中心化、区块链趋向去中心化”的张力出发，提出二者在数据/算力垄断、透明可审计、激励与安全等维度的互补关系与治理框架，用以讨论去中心化未来的技术与制度组合。

  • 一句话总结: 提供AI与区块链协同的宏观框架，阐明区块链如何在治理与激励层面对冲AI中心化趋势。

  • Track: AI × Blockchain / Decentralized AI governance & infrastructure (editorial/framework)

  • Core innovation: Articulates a complementary governance framing where blockchain’s decentralization, transparency, and incentive/security mechanisms counterbalance AI/LLM centralization driven by data and compute concentration.
  • One-sentence summary: Clarifies how blockchain primitives could mitigate centralization pressures in AI through governance, auditability, and incentives.

• [2026-03-07] Enhancing Consistency of Werewolf AI through Dialogue Summarization and Persona Information 📖2 🆕NEW

  • 赛道归属: 多智能体对话 / 社交推理游戏代理（Werewolf）与一致性控制
  • 核心创新点: 通过对话摘要与人格信息注入，构建可持续更新的“记忆-人格”状态表示，约束LLM在多轮博弈对话中的立场与叙事一致性，减少自相矛盾与角色漂移。

  • 一句话总结: 用摘要记忆与人格约束提升LLM博弈对话代理的长期一致性与可控性。

  • Track: Multi-agent dialogue / Social deduction game agents & consistency control

  • Core innovation: Improves long-horizon dialogue consistency by maintaining an updatable state via dialogue summarization plus persona conditioning, reducing contradictions and role drift in LLM-based Werewolf agents.
  • One-sentence summary: Enhances controllability and coherence of LLM game agents through summary-based memory and persona grounding.

• [2026-03-12] DatedGPT: Preventing Lookahead Bias in Large Language Models with Time-Aware Pretraining 📖1 🆕NEW

  • 赛道归属: 金融时序建模 / 时间感知预训练与数据泄漏（Lookahead Bias）防控
  • 核心创新点: 提出按年份严格切分语料并从零训练的一组时间版本模型（annual cutoffs），用“时间感知预训练”从源头避免模型在回测/预测中因训练见过未来信息而产生前视偏差，并配套指令微调以适配下游任务。

  • 一句话总结: 通过时间切分预训练为金融场景提供更可信的LLM回测与预测基座，显著降低前视偏差风险。

  • Track: Financial time-series NLP / Time-aware pretraining & leakage (lookahead bias) prevention

  • Core innovation: Trains a suite of models from scratch on temporally partitioned corpora with strict annual cutoffs, preventing future-information leakage in backtesting, and adds instruction tuning for downstream usability.
  • One-sentence summary: Delivers more valid financial forecasting/backtesting LLMs by eliminating lookahead bias via time-aware pretraining.

• [2026-03-11] Dynamics-Predictive Sampling for Active RL Finetuning of Large Reasoning Models 📖1 🆕NEW

  • 赛道归属: 推理优化 / 强化学习微调（RLHF/RLAIF）中的主动数据选择
  • 核心创新点: 提出“动力学预测采样”（Dynamics-Predictive Sampling），用训练动态信号预测样本对策略更新的边际收益，主动挑选最能带来有效梯度更新的提示/轨迹，相比只选“中等难度”样本更高效地提升推理能力。

  • 一句话总结: 通过可预测的训练动态来做主动采样，加速并强化大推理模型的RL微调效果。

  • Track: Reasoning optimization / RL finetuning with active data selection

  • Core innovation: Introduces Dynamics-Predictive Sampling that leverages training-dynamics signals to predict each example’s expected policy-improvement gain, actively selecting prompts/trajectories that yield more effective updates than difficulty-only heuristics.
  • One-sentence summary: Speeds up and strengthens RL finetuning for large reasoning models by selecting training data based on predicted learning impact.

HuggingFace Datasets

• [2026-03-04] TuringEnterprises/Open-RL 🆕NEW

  Open-RL
  
  Dataset Summary
  

This dataset contains self-contained, verifiable, and unambiguous STEM reasoning problems across Phys...

多模态大模型 / Multimodal Models

arXiv

• [2026-03-10] Prune Redundancy, Preserve Essence: Vision Token Compression in VLMs via Synergistic Importance-Diversity 📖1 🆕NEW

  • 赛道归属: 多模态理解 / VLM推理优化（视觉token压缩与剪枝）
  • 核心创新点: 提出训练无关的PruneSID，以“重要性-多样性协同”为目标进行视觉token压缩：先用PSCA提取主语义成分并聚类以去冗余，再在簇内/簇间联合选择以同时保留关键信息与覆盖多样语义。相较仅按注意力或相似度剪枝的方法，更系统地平衡“保真”和“信息覆盖”。
  • 一句话总结: 在不改动模型与无需训练的前提下，更稳健地减少视觉token冗余，从而显著降低VLM推理成本并尽量不损失语义能力。
  • Track: Multimodal Understanding / VLM Inference Optimization (visual token compression & pruning)
  • Core innovation: Introduces PruneSID, a training-free visual token compression method that explicitly optimizes a synergy between importance and diversity: PSCA extracts principal semantic components for clustering to remove redundancy, then performs joint intra-/inter-cluster selection to preserve both salient content and semantic coverage. This goes beyond attention/similarity-only pruning by balancing fidelity and diversity more directly.
  • One-sentence takeaway: Cuts redundant visual tokens without retraining while better preserving model capability, enabling cheaper VLM inference with minimal semantic loss.

• [2026-03-12] ZeroSense:How Vision matters in Long Context Compression 📖2 🆕NEW

  • 赛道归属: 多模态长上下文压缩与评测（视觉-文本压缩/VTC）
  • 核心创新点: 提出ZeroSense评测框架，将“下游任务表现”与“文本保真度”解耦，针对MLLM强语言先验导致的虚假高分问题，设计更能度量压缩后文本是否被真实保留的评价协议。通过强调视觉渲染在压缩中的作用与失真来源，提供更可靠的长上下文压缩诊断工具。
  • 一句话总结: 为VTC类长上下文压缩建立更可信的“保真度”评测，避免被MLLM语言先验掩盖的文本丢失。
    Track: Multimodal long-context compression & evaluation (visual-text compression, VTC)
    Core innovation: ZeroSense introduces an evaluation framework that decouples downstream-task success from text-preservation fidelity, addressing inflated scores caused by MLLMs’ strong linguistic priors. It provides protocols that more directly test whether the compressed representation truly retains the original text.
    One-sentence summary: It makes VTC evaluation more trustworthy by explicitly measuring preservation fidelity rather than relying on downstream performance.

• [2026-03-10] Towards Unified Multimodal Interleaved Generation via Group Relative Policy Optimization 📖1 🆕NEW

  • 赛道归属: 多模态交错生成（Interleaved Generation）/ 强化学习后训练
  • 核心创新点: 提出Group Relative Policy Optimization（GRPO）用于统一多模态模型的后训练，在不依赖大规模交错图文数据的前提下，通过组内相对偏好优化解锁“图文交错输出”能力。采用warm-up与基于相对奖励的策略更新，降低对绝对标注/奖励标定的依赖并提升训练稳定性。
  • 一句话总结: 用相对策略优化的RL后训练，让现有统一模型在缺少交错数据时也能学会高质量图文交错生成。
    Track: Multimodal interleaved generation / RL post-training
    Core innovation: The work proposes GRPO to post-train unified vision-language models for interleaved multimodal outputs without large-scale interleaved datasets, optimizing relative preferences within groups instead of relying on absolute reward calibration. A warm-up plus relative-reward policy updates improves stability and unlocks interleaved generation.
    One-sentence summary: It enables strong interleaved image-text generation via RL post-training even when interleaved supervision data is scarce.

• [2026-03-08] Holi-Spatial: Evolving Video Streams into Holistic 3D Spatial Intelligence 📖1 🆕NEW

  • 赛道归属: 多模态理解 / 3D空间智能（从视频到3D表征与数据构建）
  • 核心创新点: 面向“空间智能”提出从海量网络视频流中系统性演化/标注出细粒度3D信息的框架，突破以少量人工3D数据集生成QA的可扩展性瓶颈。通过直接从原始视频构建更大规模、更贴近真实分布的3D场景与评测资源，缓解数据规模与域差问题。
  • 一句话总结: 将空间智能的数据来源从小规模人工3D集扩展到可规模化的网络视频，为3D理解与推理提供更真实、更大规模的训练与评测基础。
  • Track: Multimodal Understanding / 3D Spatial Intelligence (video-to-3D representation & data construction)
  • Core innovation: Proposes a scalable pipeline to evolve raw web video streams into fine-grained 3D spatial supervision/benchmarks, moving beyond QA generation from a handful of manually annotated 3D datasets. By constructing large-scale, more in-the-wild 3D resources directly from videos, it addresses both scalability limits and domain gaps.
  • One-sentence takeaway: Scales 3D spatial intelligence by turning web videos into realistic large-scale 3D data/benchmarks, strengthening training and evaluation for 3D understanding and reasoning.

• [2026-03-07] MAviS: A Multimodal Conversational Assistant For Avian Species 📖1 🆕NEW

  • 赛道归属: 多模态理解 / 垂直领域多模态助手（细粒度物种识别与问答）
  • 核心创新点: 构建MAviS-Dataset，将鸟类物种相关的图像与文本知识/对话式问答进行大规模整合，以支持细粒度、物种特定的多模态对话能力。针对通用MLLM在专业鸟类知识与细粒度辨识上的短板，通过数据与任务设计提升专业领域可用性与可靠性。
  • 一句话总结: 通过面向鸟类物种的专用多模态数据与助手形态，推动MLLM在生物多样性监测等真实垂直场景中的可落地应用。
  • Track: Multimodal Understanding / Vertical-domain multimodal assistant (fine-grained species ID & QA)
  • Core innovation: Introduces MAviS-Dataset, a large-scale integration of bird-species images with knowledge and conversational QA to enable fine-grained, species-specific multimodal dialogue. It targets the gap where general-purpose MLLMs underperform on specialized avian expertise and subtle visual distinctions via domain-tailored data/task design.
  • One-sentence takeaway: Makes multimodal assistants more practical for biodiversity and ecological monitoring by grounding them in a dedicated, fine-grained avian dataset.

• [2026-03-07] The Model Knows Which Tokens Matter: Automatic Token Selection via Noise Gating 📖1 🆕NEW

  • 赛道归属: 多模态理解 / VLM推理优化（自动视觉token选择）
  • 核心创新点: 将视觉token剪枝重述为“容量受限通信”问题：在固定预算K下最大化保留视觉信息；提出AutoSelect，在冻结VLM上外挂轻量Scorer与Denoiser，并通过噪声门控(Noise Gating)训练，让模型自动学习“哪些token重要”。相较基于注意力幅值/相似度的启发式规则，提供可学习、任务自适应的token分配机制。
  • 一句话总结: 用可学习的噪声门控替代启发式剪枝，实现更自适应的视觉token预算分配，在降算力的同时更好维持性能。
  • Track: Multimodal Understanding / VLM Inference Optimization (automatic visual token selection)
  • Core innovation: Reframes visual token pruning as capacity-constrained communication—maximize preserved visual information under a fixed budget K. Proposes AutoSelect by attaching a lightweight Scorer and Denoiser to a frozen VLM and training them with noise gating so the model learns which tokens matter, replacing heuristic attention/similarity-based pruning with a learnable, task-adaptive allocation.
  • One-sentence takeaway: Achieves more reliable compute reduction by learning token importance via noise gating, preserving performance better than heuristic pruning.

• [2026-03-12] MM-CondChain: A Programmatically Verified Benchmark for Visually Grounded Deep Compositional Reasoning 🆕NEW

  • 赛道归属: 多模态推理评测 / 视觉条件链式组合推理（基准与验证）
  • 核心创新点: 提出MM-CondChain，用“程序化可验证”(programmatically verified)的方式构建视觉落地的深度组合条件链推理基准，覆盖带分支/提前终止的工作流式决策（如GUI操作中的多条件if-then链）。相较浅层组合或独立约束评测，更强调长链条件依赖与可自动判定的正确性。
  • 一句话总结: 为MLLM在真实视觉工作流中的“可验证深组合条件推理”提供了更贴近应用且可自动评测的基准。
  • Track: Multimodal reasoning evaluation / Visually grounded chained compositional reasoning (benchmark & verification)
  • Core innovation: Introduces MM-CondChain, a programmatically verified benchmark for deeply chained, visually grounded compositional conditions with branching/early termination, reflecting workflow decisions (e.g., GUI navigation with multi-condition if-then chains). It goes beyond shallow compositions or independent constraints by emphasizing long-horizon conditional dependencies with automatically checkable correctness.
  • One-sentence takeaway: Provides an application-faithful, automatically verifiable benchmark to measure MLLMs’ deep conditional reasoning in visual workflows.

• [2026-03-12] EndoCoT: Scaling Endogenous Chain-of-Thought Reasoning in Diffusion Models 🆕NEW

  • 赛道归属: 文生图 / 扩散模型推理增强（链式思维式指导）
  • 核心创新点: 提出EndoCoT，将“内生链式思维”(endogenous CoT)引入扩散生成：不再把MLLM仅当一次性文本编码器，而是通过多步推理产生更深的指导信号，并使指导在扩散采样过程中动态演化而非保持不变。该设计针对复杂空间/组合任务提升可控性与推理深度。
  • 一句话总结: 让扩散模型获得随采样过程演进的多步推理式指导，从而更好生成满足复杂约束的图像结果。
  • Track: Text-to-image / Diffusion model reasoning enhancement (CoT-style guidance)
  • Core innovation: Proposes EndoCoT to scale endogenous chain-of-thought reasoning inside diffusion generation: instead of using an MLLM as a one-shot text encoder, it elicits multi-step reasoning to produce deeper guidance and makes the guidance evolve during the diffusion sampling process rather than staying invariant. This targets improved controllability on complex spatial/compositional tasks.
  • One-sentence takeaway: Enables diffusion models to follow complex constraints better by injecting dynamic, multi-step reasoning guidance throughout sampling.

• [2026-03-12] SceneAssistant: A Visual Feedback Agent for Open-Vocabulary 3D Scene Generation 🆕NEW

  • 赛道归属: 3D生成 / 文本到3D场景生成（智能体与视觉反馈）
  • 核心创新点: 提出SceneAssistant，一个以“视觉反馈”为闭环的开放词表3D场景生成智能体：结合3D物体生成模型与空间推理模块，在生成过程中根据渲染/视觉结果进行迭代修正，从而摆脱预定义关系与受限域。通过“生成-观察-调整”的代理式流程提升开放指令下的场景一致性与可控性。
  • 一句话总结: 用视觉反馈驱动的代理式迭代生成，将文本到3D场景从受限模板推进到更开放、更可控的开放词表合成。
  • Track: 3D Generation / Text-to-3D scene generation (agentic loop with visual feedback)
  • Core innovation: Introduces SceneAssistant, a visual-feedback closed-loop agent for open-vocabulary text-to-3D scene generation: it combines modern 3D object generation with spatial reasoning and iteratively corrects the scene based on rendered visual feedback, reducing reliance on predefined relations or domain constraints. The generate-observe-adjust loop improves consistency and controllability under unconstrained prompts.
  • One-sentence takeaway: Advances open-vocabulary 3D scene synthesis by using an agentic visual-feedback loop to iteratively refine spatial layouts and objects.

• [2026-03-12] ForensicZip: More Tokens are Better but Not Necessary in Forensic Vision-Language Models 🆕NEW

  • 赛道归属: 多模态理解 / 多媒体取证VLM推理优化（面向取证的token剪枝）
  • 核心创新点: 提出ForensicZip，指出取证场景“更多token更好但非必要”，并针对现有语义驱动剪枝会丢失背景中高频伪造痕迹的问题，设计更贴合取证线索（高频异常、细微纹理/压缩痕迹等）的token保留策略。核心在于让剪枝目标从“语义显著”转向“取证证据显著”，在降算力下维持可解释取证能力。
  • 一句话总结: 将视觉token压缩从语义导向改为取证证据导向，使取证VLM在加速推理时不牺牲关键篡改线索。
  • Track: Multimodal Understanding / Multimedia forensics VLM optimization (forensics-aware token pruning)
  • Core innovation: Proposes ForensicZip, observing that while more tokens help forensic VLMs, they are not strictly necessary; it addresses a key failure of semantic-driven pruning that discards background regions containing manipulation traces (e.g., high-frequency anomalies, subtle texture/compression artifacts). The method shifts pruning objectives from semantic saliency to forensic-evidence saliency to retain interpretability and detection performance under reduced compute.
  • One-sentence takeaway: Makes token pruning safe for forensic VLMs by preserving manipulation evidence rather than just semantic objects, enabling faster inference without losing critical traces.

---

Generated automatically by Daily AI Digest Agent 生成时间: 2026-03-13 10:50:06