AI 每日进展速报 / Daily AI Digest - 2026-03-23

图像生成/编辑 / Image Generation/Editing

arXiv

• [2026-03-18] Interpretable Cross-Domain Few-Shot Learning with Rectified Target-Domain Local Alignment 📖1 🆕NEW

  • 赛道归属: 跨域小样本学习（CDFSL）/ 可解释视觉-语言模型适配
  • 核心创新点: 提出“目标域局部对齐”的可解释适配框架，通过在目标域上对细粒度局部区域进行校正式对齐，缓解微调CLIP只关注粗粒度显著区域、难以捕捉诊断级细节线索的问题。方法强调可解释的局部证据学习，使模型注意力更贴近目标域关键视觉线索。

  • 一句话总结: 让CLIP类模型在跨域小样本场景中学会“看对地方”，以更可解释的细粒度局部对齐提升目标域识别可靠性。

  • Track: Cross-domain few-shot learning (CDFSL) / Interpretable vision-language model adaptation

  • Core innovation: Introduces a rectified target-domain local alignment scheme that explicitly aligns fine-grained local regions in the target domain, addressing the tendency of fine-tuned CLIP to focus only on coarse salient areas. The method promotes evidence-based, interpretable recognition by steering attention toward target-domain critical cues.
  • One-sentence summary: It improves CDFSL by making CLIP-like models attend to the right fine-grained target-domain evidence, boosting interpretability and robustness.

• [2026-03-20] Beyond Single Tokens: Distilling Discrete Diffusion Models via Discrete MMD 🆕NEW

  • 赛道归属: 扩散模型蒸馏（离散扩散 / 文本或离散token生成）
  • 核心创新点: 提出离散矩匹配蒸馏（D-MMD），用离散MMD进行分布级矩匹配，将连续扩散蒸馏中有效的“矩匹配”思想迁移到离散扩散，避免以往离散蒸馏易崩塌（质量/多样性下降）的问题。该蒸馏在足够采样步数下保持高质量与多样性。

  • 一句话总结: 为离散扩散提供了更稳定的蒸馏范式，使其更接近连续扩散蒸馏的可用性与效果。

  • Track: Diffusion distillation (discrete diffusion / discrete-token generation)

  • Core innovation: Proposes Discrete Moment Matching Distillation (D-MMD), using discrete MMD to match distributional moments—bringing successful continuous diffusion distillation ideas into the discrete setting while avoiding collapse seen in prior discrete distillation methods. It preserves quality and diversity given sufficient sampling steps.
  • One-sentence summary: It delivers a more stable distillation recipe for discrete diffusion models, narrowing the gap with continuous diffusion distillation.

• [2026-03-20] Preference-Guided Debiasing for No-Reference Enhancement Image Quality Assessment 🆕NEW

  • 赛道归属: 图像质量评估（NR-IQA / 增强图像质量评估EIQA）
  • 核心创新点: 提出偏好引导的去偏框架：先学习连续的“增强偏好”嵌入空间（如用监督式对比学习）以刻画人类对不同增强风格/强度的偏好，再在此基础上抑制模型对特定增强算法伪迹模式的过拟合，从而提升跨算法泛化的无参考质量评估能力。

  • 一句话总结: 用“偏好表征+去偏学习”把EIQA从识别算法痕迹转向评估真实感知质量，显著增强泛化性。

  • Track: Image quality assessment (NR-IQA / enhancement IQA)

  • Core innovation: Introduces a preference-guided debiasing framework that first learns a continuous enhancement-preference embedding space (e.g., via supervised contrastive learning) and then uses it to reduce overfitting to enhancement-algorithm-specific artifacts, improving cross-method generalization for no-reference EIQA.
  • One-sentence summary: It shifts EIQA from “detecting enhancement signatures” to measuring true perceptual quality through preference-driven debiasing.

• [2026-03-20] Diffusion-Based Makeup Transfer with Facial Region-Aware Makeup Features 🆕NEW

  • 赛道归属: 图像编辑（人脸妆容迁移 / 扩散模型编辑）
  • 核心创新点: 提出面部区域感知的妆容特征建模与注入方式，替代直接用CLIP等通用基础模型提取的全局妆容条件；通过按面部区域（如唇、眼、肤）分解并注入妆容特征，实现更精细、可控且避免“全局涂抹式”迁移的扩散式妆容转移。

  • 一句话总结: 让扩散妆容迁移从“全脸同风格”升级为“分区精细上妆”，显著提升妆效还原与可控性。

  • Track: Image editing (face makeup transfer / diffusion-based editing)

  • Core innovation: Proposes facial region-aware makeup feature extraction and conditioning, replacing global makeup cues from generic foundation models (e.g., CLIP). By decomposing and injecting makeup features per facial region, it enables finer, more controllable transfer and mitigates over-globalized makeup application.
  • One-sentence summary: It upgrades diffusion makeup transfer to region-wise, high-fidelity, controllable editing rather than global style overlay.

• [2026-03-20] Timestep-Aware Block Masking for Efficient Diffusion Model Inference 🆕NEW

  • 赛道归属: 推理优化（扩散模型加速 / 动态计算图与跳层）
  • 核心创新点: 提出时间步感知的Block Masking：针对扩散去噪轨迹中不同时间步的特征动态，学习每个timestep应执行或旁路哪些网络块，并通过特征复用减少冗余计算，从而在不改训练的前提下对预训练DPM进行按步动态裁剪与加速。

  • 一句话总结: 用“按时间步动态跳块+特征复用”系统性降低扩散推理延迟，同时尽量保持生成质量。

  • Track: Inference optimization (diffusion acceleration / dynamic execution & block skipping)

  • Core innovation: Introduces timestep-aware block masking that learns per-timestep masks to decide which blocks to execute or bypass, leveraging feature reuse along the denoising trajectory. This optimizes the computation graph of pretrained diffusion models without retraining.
  • One-sentence summary: It reduces diffusion inference latency via per-timestep dynamic block execution while preserving output quality.

• [2026-03-20] Evaluating Image Editing with LLMs: A Comprehensive Benchmark and Intermediate-Layer Probing Approach 🆕NEW

  • 赛道归属: 图像编辑评测（文本引导图像编辑TIE Benchmark / 自动评估）
  • 核心创新点: 提出更系统的TIE评测基准TIEdit，并引入基于LLM的评估与“中间层探针”方法：通过探测模型/评估器的中间表征来更可靠地衡量指令对齐、感知质量与内容保持三者的权衡，提升与人类判断的相关性并诊断编辑失败模式。

  • 一句话总结: 用“更大更全的基准+中间层可诊断评估”补齐TIE评价短板，让编辑模型进步可量化、可解释。

  • Track: Image editing evaluation (text-guided image editing benchmark / automated evaluation)

  • Core innovation: Presents TIEdit, a comprehensive benchmark for text-guided image editing, and proposes LLM-based evaluation with intermediate-layer probing to better assess instruction alignment, perceptual quality, and content preservation, improving correlation with human judgments and enabling failure-mode diagnosis.
  • One-sentence summary: It makes TIE evaluation more reliable and diagnostic through a stronger benchmark plus intermediate-representation probing.

• [2026-03-20] WorldAgents: Can Foundation Image Models be Agents for 3D World Models? 🆕NEW

  • 赛道归属: 3D生成 / 世界模型评测（基于2D基础模型的Agent式3D世界合成）
  • 核心创新点: 系统评估2D图像生成模型与VLM是否具备隐式3D世界建模能力，并提出agentic框架将“生成—验证—迭代”的过程组织为代理行为，以驱动多视角一致性与3D世界合成；同时提供可对比的基准化评测来量化其3D能力边界。

  • 一句话总结: 通过“代理式闭环”把2D基础模型推向3D世界合成，并用系统评测回答其3D能力到底到哪一步。

  • Track: 3D generation / world-model benchmarking (agentic 3D world synthesis from 2D foundation models)

  • Core innovation: Systematically evaluates whether 2D image generators and VLMs exhibit implicit 3D world-modeling ability, and introduces an agentic loop (generate–check–iterate) to harness these models for multi-view-consistent 3D world synthesis, alongside standardized evaluation to map capability limits.
  • One-sentence summary: It turns 2D foundation models into agent-driven 3D world builders and rigorously benchmarks how far their implicit 3D understanding goes.

• [2026-03-20] ATHENA: Adaptive Test-Time Steering for Improving Count Fidelity in Diffusion Models 🆕NEW

  • 赛道归属: 文生图（扩散采样控制 / 计数一致性提升）
  • 核心创新点: 提出ATHENA测试时自适应引导框架：在采样过程中利用中间表征估计当前生成的目标计数，并施加计数感知的噪声校正/引导来纠正偏差；无需改结构、无需再训练，属于模型无关的test-time steering。

  • 一句话总结: 不训练也能显著提升扩散模型“数得对”的能力，为数值可控生成提供通用测试时方案。

  • Track: Text-to-image (diffusion sampling control / count fidelity)

  • Core innovation: Proposes ATHENA, a model-agnostic test-time adaptive steering method that estimates object counts from intermediate sampling representations and applies count-aware noise corrections to reduce counting errors—without architecture changes or retraining.
  • One-sentence summary: It improves numerical controllability (object counts) in diffusion models via a plug-and-play test-time steering mechanism.

• [2026-03-20] Toward High-Fidelity Visual Reconstruction: From EEG-Based Conditioned Generation to Joint-Modal Guided Rebuilding 🆕NEW

  • 赛道归属: 脑信号条件生成 / 多模态重建（EEG到图像重建）
  • 核心创新点: 从“强制EEG对齐到文本/图像语义空间”的耦合范式转向更高保真重建：先进行EEG条件生成，再引入联合模态（如文本/图像等）指导的重建/细化流程，以更好恢复空间关系与色彩细节，降低对单一对齐框架的依赖。

  • 一句话总结: 通过“EEG条件生成+联合模态引导重建”提升神经信号视觉重建的细节与稳定性。

  • Track: Brain-signal conditioned generation / multimodal visual reconstruction (EEG-to-image)

  • Core innovation: Moves beyond tightly coupled EEG-to-text/image semantic alignment by combining EEG-conditioned generation with joint-modality guided rebuilding/refinement, aiming for higher-fidelity recovery of spatial relations and chromatic details while reducing dependence on a single alignment framework.
  • One-sentence summary: It boosts EEG-based visual reconstruction fidelity by pairing EEG-conditioned synthesis with joint-modal guidance for refinement.

• [2026-03-20] MagicSeg: Open-World Segmentation Pretraining via Counterfactural Diffusion-Based Auto-Generation 🆕NEW

  • 赛道归属: 语义分割预训练（开放世界分割 / 扩散生成数据合成）
  • 核心创新点: 提出基于反事实（counterfactual）扩散自动生成的开放世界分割预训练管线：利用扩散模型生成可控场景与对应像素级标注/伪标注数据，构建覆盖更广类别与组合的训练集，减少对昂贵人工像素标注与大规模图文对的依赖。

  • 一句话总结: 用扩散模型“自造数据集”来做开放世界分割预训练，把像素标注瓶颈转化为可扩展的生成问题。

  • Track: Segmentation pretraining (open-world segmentation / diffusion-based data synthesis)

  • Core innovation: Introduces a counterfactual diffusion-driven auto-generation pipeline to create training data (with pixel-level labels or strong pseudo-labels) for open-world segmentation pretraining, expanding category/combination coverage while reducing reliance on costly human annotations and large image-text pairs.
  • One-sentence summary: It tackles the pixel-annotation bottleneck by using diffusion models to auto-generate scalable datasets for open-world segmentation pretraining.

视频生成/编辑 / Video Generation/Editing

arXiv

• [2026-03-16] Faster Inference of Flow-Based Generative Models via Improved Data-Noise Coupling 📖10 🆕NEW

  • 赛道归属: 流模型加速推理（Continuous Normalizing Flow / Conditional Flow Matching）
  • 核心创新点: 聚焦于“数据-噪声耦合”(data-noise coupling)对CFM采样轨迹长度与推理速度的决定性影响，在minibatch OT重配对的思路上改进其优化方式，使耦合更稳定/更高效，从而在不依赖仿真训练的前提下进一步缩短采样路径、加速生成推理。方法层面强调通过更优耦合来直接优化推理成本，而非改网络或改采样器的表层加速。

  • 一句话总结: 通过改进CFM中的数据-噪声配对策略，以更短的生成轨迹实现更快的流模型推理，为图像/视频生成提供扩散之外的高效路线。

  • Track: Flow-model inference acceleration (Continuous Normalizing Flow / Conditional Flow Matching)

  • Core innovation: Identifies data–noise coupling as a key lever controlling CFM trajectory length and inference speed, and improves minibatch OT-style reassignment with a more effective/stable optimization to obtain better couplings—thereby shortening sampling paths and accelerating generation without simulation-based training.
  • One-sentence summary: Improves CFM via better data–noise pairing to achieve faster flow-based generation, strengthening flows as an efficient alternative to diffusion for image/video synthesis.

• [2026-03-20] MME-CoF-Pro: Evaluating Reasoning Coherence in Video Generative Models with Text and Visual Hints 🆕NEW

  • 赛道归属: 视频生成评测（推理一致性/因果一致性 Benchmark）
  • 核心创新点: 提出“推理连贯性”(reasoning coherence)作为视频生成模型跨帧因果一致性的关键能力维度，并构建MME-CoF-Pro基准：用文本与视觉提示覆盖多类推理场景，专门检验生成事件在时间维度是否自洽、是否遵循因果链条，而不仅是画质或单帧语义对齐。

  • 一句话总结: 用系统化基准把“跨帧因果自洽”从主观观感中剥离出来，推动视频生成模型向可部署的可靠推理行为评估迈进。

  • Track: Video generation evaluation (reasoning/causal coherence benchmark)

  • Core innovation: Formalizes “reasoning coherence” as cross-frame causal consistency for video generators and introduces MME-CoF-Pro, a benchmark with text and visual hints spanning diverse reasoning categories to directly test whether generated events remain temporally and causally self-consistent beyond perceptual quality.
  • One-sentence summary: Provides a targeted benchmark to measure causal/temporal reasoning consistency in generated videos, enabling more deployment-relevant evaluation of video generative models.

• [2026-03-20] LumosX: Relate Any Identities with Their Attributes for Personalized Video Generation 🆕NEW

  • 赛道归属: 个性化视频生成（身份-属性绑定/人脸属性一致性）
  • 核心创新点: 针对多人物/多身份场景中“人脸身份与属性对齐不稳、组内一致性差”的痛点，引入显式的身份-属性关联建模，并配套构建面向人脸-属性对齐的数据资源与训练策略，使模型能在视频时序中稳定地把正确属性绑定到对应身份，同时兼顾前景人物与背景控制。

  • 一句话总结: 通过显式身份-属性绑定机制与数据支撑，显著提升个性化视频生成中多身份的属性一致性与可控性。

  • Track: Personalized video generation (identity–attribute binding / face-attribute consistency)

  • Core innovation: Addresses unstable face–attribute alignment in multi-identity videos by explicitly modeling identity–attribute relations and introducing face-attribute-aware data/training strategies, enabling temporally consistent attribute binding to the correct subject while retaining fine-grained foreground/background control.
  • One-sentence summary: Makes personalized video generation more reliable by ensuring attributes stay correctly attached to each identity across time.

• [2026-03-20] EgoForge: Goal-Directed Egocentric World Simulator 🆕NEW

  • 赛道归属: 生成式世界模型/具身智能仿真（第一人称目标导向视频生成）
  • 核心创新点: 面向第一人称(egocentric)世界模拟的难点（快速视角变化、手-物交互频繁、过程受隐式意图驱动），提出目标导向的生成式模拟框架，在不依赖密集监督（如精确相机轨迹、长视频标注）的情况下，建模动作动态与场景演化，使生成序列能随“目标/意图”推进而非仅做外观迁移或局部手部合成。

  • 一句话总结: 把第一人称视频从“看起来像”推进到“按目标演化”，为具身任务提供更可用的生成式世界模拟器。

  • Track: Generative world models / embodied simulation (goal-directed egocentric video generation)

  • Core innovation: Proposes a goal-directed egocentric world simulator that tackles rapid viewpoint changes and hand–object interactions by modeling action dynamics and scene evolution driven by latent intent, avoiding reliance on dense supervision such as exact camera trajectories or long annotated videos.
  • One-sentence summary: Enables egocentric world simulation that evolves with goals/intent, improving usefulness for embodied AI beyond appearance-only synthesis.

• [2026-03-20] X-World: Controllable Ego-Centric Multi-Camera World Models for Scalable End-to-End Driving 🆕NEW

  • 赛道归属: 自动驾驶世界模型（可控多摄像头自车视角生成/仿真评测）
  • 核心创新点: 面向端到端VLA自动驾驶的可扩展评测需求，提出可控的自车第一人称多摄像头世界模型：在给定候选动作/策略条件下生成未来多视角观测，并强调可复现、可覆盖长尾场景的仿真能力，从而减少对昂贵且偏置的真实道路测试依赖。

  • 一句话总结: 用可控多相机生成式世界模型把端到端驾驶评测从“上路试”转向“可复现的生成仿真”，提升规模化验证能力。

  • Track: Autonomous driving world models (controllable egocentric multi-camera generation / simulation for evaluation)

  • Core innovation: Introduces a controllable multi-camera egocentric world model that generates realistic future observations conditioned on proposed actions/policies, targeting reproducible and scalable evaluation for end-to-end VLA driving systems and broader scenario coverage than real-road testing.
  • One-sentence summary: Shifts end-to-end driving evaluation toward controllable, reproducible generative simulation with multi-view future prediction.

• [2026-03-20] FoleyDirector: Fine-Grained Temporal Steering for Video-to-Audio Generation via Structured Scripts 🆕NEW

  • 赛道归属: 视频到音频生成（V2A）/可控音频生成（细粒度时间编排）
  • 核心创新点: 提出结构化“脚本”(structured scripts)作为中间控制层，对多事件音效进行细粒度时间定位与引导，在DiT类V2A生成框架中实现精确的时序转向(temporal steering)，并在视觉线索不足（遮挡、小目标、画外音）时仍能通过脚本补足约束，同时尽量保持基座模型的音质与泛化能力。

  • 一句话总结: 用“脚本化时间控制”解决V2A多事件难以对齐的问题，让视频配音从随机命中走向可编排、可导演。

  • Track: Video-to-audio generation (V2A) / controllable audio synthesis (fine-grained temporal control)

  • Core innovation: Uses structured scripts as an explicit control interface to provide fine-grained temporal steering for multi-event audio in DiT-based V2A, enabling precise event timing even when visual cues are weak (occluded/off-screen/small regions) while preserving base model quality.
  • One-sentence summary: Turns V2A into a “directable” process by adding script-based, fine-grained temporal control for complex multi-event soundtracks.

• [2026-03-20] PCSTracker: Long-Term Scene Flow Estimation for Point Cloud Sequences 🆕NEW

  • 赛道归属: 3D点云序列理解（长时序Scene Flow估计/时序一致性）
  • 核心创新点: 针对现有scene flow多停留在两帧配对、长序列易漂移与不一致的问题，提出首个面向点云序列的端到端一致scene flow框架，通过迭代式几何匹配/跟踪机制在遮挡、形变与误差累积下维持跨时间的运动估计一致性，实现更长时程、更细粒度的3D运动分析。

  • 一句话总结: 把scene flow从“两帧估计”升级为“序列级一致跟踪”，为长期3D动态理解提供关键基础能力。

  • Track: 3D point cloud sequence modeling (long-term scene flow / temporal consistency)

  • Core innovation: Moves beyond pairwise scene flow by proposing an end-to-end sequence-level framework with iterative geometry matching/tracking to maintain temporally consistent motion estimates under occlusions, evolving geometry, and error accumulation.
  • One-sentence summary: Enables long-horizon, consistent 3D motion estimation in point cloud sequences, strengthening foundations for dynamic 3D understanding.

• [2026-03-20] Making Video Models Adhere to User Intent with Minor Adjustments 🆕NEW

  • 赛道归属: 可控视频生成/视频编辑（布局/框约束对齐优化）
  • 核心创新点: 发现控制失败往往来自“用户框与模型可生成分布不匹配”，提出在不改模型的前提下，对用户提供的bounding boxes做小幅可微优化/调整，使其更符合生成先验，从而同时提升画质与对控制输入的遵循度；本质上是把控制问题转化为“输入约束的轻量校准”。

  • 一句话总结: 通过“微调控制框而非微调模型”，以极低成本显著增强视频生成对用户意图的服从性。

  • Track: Controllable video generation / editing (layout & bounding-box adherence optimization)

  • Core innovation: Recasts control failures as a mismatch between user boxes and the model’s generative prior, and improves both quality and adherence by slightly optimizing (adjusting) the provided bounding boxes—achieving better controllability without modifying the underlying video diffusion model.
  • One-sentence summary: Boosts intent adherence cheaply by calibrating the control inputs (boxes) instead of retraining the model.

• [2026-03-20] OrbitNVS: Harnessing Video Diffusion Priors for Novel View Synthesis 🆕NEW

  • 赛道归属: 新视角合成（NVS）/视频扩散先验迁移（单视图到多视图）
  • 核心创新点: 将NVS重构为“轨道环绕视频生成”任务，利用预训练视频扩散模型的时序一致性先验来补全未观测区域；通过定制化模型结构与训练策略，把视频生成的动态一致性转化为多视角几何/外观一致性，尤其强化单视图输入下对背面/遮挡区域的合理合成。

  • 一句话总结: 把视频扩散的强先验用于NVS，通过“环绕视频化”显著提升少视角尤其单视图的新视角补全能力。

  • Track: Novel view synthesis (NVS) / leveraging video diffusion priors (single-view to multi-view)

  • Core innovation: Reformulates NVS as an orbiting video generation problem and adapts a pretrained video diffusion model with tailored architecture/training so temporal coherence priors translate into geometry/appearance consistency, improving plausibility for unobserved regions—especially from single-view input.
  • One-sentence summary: Uses video diffusion as a powerful prior for NVS by turning view synthesis into orbit-video generation, improving consistency and unseen-region completion.

• [2026-03-20] Physion-Eval: Evaluating Physical Realism in Generated Video via Human Reasoning 🆕NEW

  • 赛道归属: 视频生成评测（物理真实感/人类推理式评估）
  • 核心创新点: 提出Physion-Eval，从“人类推理”角度细粒度评估生成视频的物理真实性：不止做偏好/打分，而是通过更具诊断性的评测设计定位模型在动力学、接触、守恒等物理规律上的违背情形与原因，从而把视频模型作为世界模拟器的可信度评估落到可解释层面。

  • 一句话总结: 用可诊断的人类推理评测补齐自动指标盲区，推动视频生成向“物理可信的世界模拟”演进。

  • Track: Video generation evaluation (physical realism / human-reasoning-based assessment)

  • Core innovation: Introduces Physion-Eval to assess physical realism via human reasoning with more diagnostic protocols than preference/rubric scoring, aiming to reveal when and why generated dynamics violate physical laws (e.g., contacts, dynamics, conservation) for world-simulator use cases.
  • One-sentence summary: Provides an interpretable, diagnostic evaluation of physical plausibility in generated videos, complementing perceptual metrics for world-model reliability.

HuggingFace Models

• Lightricks/LTX-2.3

语言大模型 / Large Language Models

arXiv

• [2026-03-17] A Scoping Review of AI-Driven Digital Interventions in Mental Health Care: Mapping Applications Across Screening, Support, Monitoring, Prevention, and Clinical Education 📖42 🆕NEW

  • 赛道归属: 医疗健康AI应用综述（数字精神健康干预/GenAI+HCAI）
  • 核心创新点: 以PRISMA-ScR框架对AI驱动的精神健康数字干预进行范围综述，提出覆盖“筛查/分诊—治疗支持—远程监测—临床教育—人群预防”的五阶段全景映射，用统一维度梳理应用形态与落地环节。强调将生成式AI与以人为中心AI纳入同一分析框架，便于对不同干预类型的证据与空白进行对照。

  • 一句话总结: 该工作用“五阶段地图”系统化刻画AI精神健康干预的应用版图，为后续评估、监管与产品化落地提供结构化参照。

  • Track: Healthcare AI survey (digital mental health interventions / GenAI + HCAI)

  • Core innovation: Conducts a PRISMA-ScR scoping review that maps AI-driven mental health technologies onto a unified five-phase care pathway (screening/triage, therapeutic support, remote monitoring, clinical education, population prevention). It explicitly integrates GenAI and human-centered AI within the same taxonomy to surface evidence patterns and gaps across intervention types.
  • One-sentence summary: Provides a structured “five-phase” landscape of AI-enabled mental health interventions to guide evaluation, governance, and real-world deployment.

• [2026-03-18] IndicSafe: A Benchmark for Evaluating Multilingual LLM Safety in South Asia 📖1 🆕NEW

  • 赛道归属: LLM安全评测（多语言/低资源语言安全基准）
  • 核心创新点: 构建面向南亚12种Indic语言的安全评测基准，使用6000条“文化语境扎根”的提示词覆盖种姓、宗教、性别、健康、政治等敏感维度，并对10个主流LLM进行跨语言对照评测。通过翻译变体与统一协议，系统揭示多语言部署中安全行为的语言差异与文化特异风险。

  • 一句话总结: 该基准把LLM安全从英语中心扩展到南亚低资源语言与文化语境，为多语言安全对齐与红队测试提供了可复现的测量工具。

  • Track: LLM safety evaluation (multilingual / low-resource language benchmark)

  • Core innovation: Introduces a safety benchmark spanning 12 Indic languages with 6,000 culturally grounded prompts across caste, religion, gender, health, and politics, and evaluates 10 leading LLMs under a unified cross-lingual protocol. The design enables controlled comparisons via translated variants to expose language- and culture-specific safety failures.
  • One-sentence summary: Extends LLM safety measurement beyond English into South Asian languages and contexts, enabling reproducible multilingual safety alignment and red-teaming.

• [2026-03-16] LLM-Driven Discovery of High-Entropy Catalysts via Retrieval-Augmented Generation 📖1 🆕NEW

  • 赛道归属: 科学发现与材料化学（RAG增强的LLM科研助手）
  • 核心创新点: 提出面向催化剂发现的检索增强生成（RAG）框架，使GPT-4通过检索接入5万+材料/化学相关数据库条目，在“化学空间探索—候选生成—结果解释”链路中实现可追溯的知识落地。将LLM的生成能力与结构化外部证据绑定，以降低幻觉并提升对专业化学信息的可用性。

  • 一句话总结: 该工作展示了RAG如何把LLM变成“可证据支撑”的材料发现助手，从而加速CO₂还原催化剂的探索流程。

  • Track: Scientific discovery & materials chemistry (RAG-augmented LLM assistant)

  • Core innovation: Proposes a retrieval-augmented generation framework for catalyst discovery where GPT-4 is grounded by retrieval over a 50k+ entry database to navigate chemical space, propose candidates, and interpret outcomes with traceable evidence. It couples generative reasoning with external structured knowledge to reduce hallucinations and improve domain utility.
  • One-sentence summary: Demonstrates how RAG can turn LLMs into evidence-grounded assistants that accelerate CO₂ reduction catalyst exploration.

• [2026-03-16] Why the Valuable Capabilities of LLMs Are Precisely the Unexplainable Ones 📖1 🆕NEW

  • 赛道归属: LLM可解释性与理论分析（能力边界/不可解释性论证）
  • 核心创新点: 通过“专家系统等价”的反证思路提出观点：若LLM的全部能力可被人类可读的离散规则完整刻画，则其等价于专家系统，而专家系统历史上无法达到LLM展现的广泛能力，由此论证LLM最有价值的能力恰来自不可完全规则化的部分。将“不可解释性”从缺陷转化为能力来源的理论命题，重塑对可解释性目标的讨论边界。

  • 一句话总结: 该论文以理论论证挑战“完全可解释”的直觉，主张LLM的关键价值可能正来自难以规则化的能力成分。

  • Track: LLM interpretability & theory (capability limits / unexplainability argument)

  • Core innovation: Advances a proof-by-contradiction argument via expert-system equivalence: if an LLM’s full capabilities were capturable by a complete set of human-readable discrete rules, it would be functionally equivalent to an expert system—yet expert systems historically fail to match LLM breadth—implying the most valuable capabilities lie in what resists full rule-based explanation. This reframes “unexplainability” as a potential source of capability rather than merely a defect.
  • One-sentence summary: Argues that pursuing fully rule-level explanations may miss what makes LLMs powerful, because their most valuable capabilities may be inherently non-rule-capturable.

• [2026-03-20] LumosX: Relate Any Identities with Their Attributes for Personalized Video Generation 🆕NEW

  • 赛道归属: 个性化视频生成（身份-属性绑定/人脸属性一致性）
  • 核心创新点: 针对多人物/多身份场景中“人脸身份与属性对齐不稳、组内一致性差”的痛点，引入显式的身份-属性关联建模，并配套构建面向人脸-属性对齐的数据资源与训练策略，使模型能在视频时序中稳定地把正确属性绑定到对应身份，同时兼顾前景人物与背景控制。

  • 一句话总结: 通过显式身份-属性绑定机制与数据支撑，显著提升个性化视频生成中多身份的属性一致性与可控性。

  • Track: Personalized video generation (identity–attribute binding / face-attribute consistency)

  • Core innovation: Addresses unstable face–attribute alignment in multi-identity videos by explicitly modeling identity–attribute relations and introducing face-attribute-aware data/training strategies, enabling temporally consistent attribute binding to the correct subject while retaining fine-grained foreground/background control.
  • One-sentence summary: Makes personalized video generation more reliable by ensuring attributes stay correctly attached to each identity across time.

• [2026-03-20] AI Agents Can Already Autonomously Perform Experimental High Energy Physics 🆕NEW

  • 赛道归属: AI智能体（科学工作流自动化/代码执行型代理）
  • 核心创新点: 在高能物理（HEP）真实分析管线中评估LLM代理的端到端自治能力：在给定数据集、执行框架与既有文献语料的条件下，代理可自动完成事件筛选、背景估计、不确定性量化、统计推断到论文撰写等关键环节。将“文献检索+代码执行+多阶段任务编排”组合为可运行的实验分析闭环，验证代理在复杂科研流程中的可行性边界。

  • 一句话总结: 该工作表明LLM代理已能在真实HEP分析中实现高比例自动化，为“自治科研”范式提供了强实证案例。

  • Track: AI agents (scientific workflow automation / code-executing agents)

  • Core innovation: Evaluates an LLM-based agent in an end-to-end high-energy physics analysis pipeline, showing it can autonomously perform event selection, background estimation, uncertainty quantification, statistical inference, and paper drafting given a dataset, an execution environment, and prior literature. It operationalizes a closed-loop workflow combining retrieval, code execution, and multi-stage orchestration to test autonomy in a complex scientific setting.
  • One-sentence summary: Provides concrete evidence that LLM agents can automate substantial parts of real experimental HEP analyses, advancing the case for autonomous science workflows.

• [2026-03-20] Learning Dynamic Belief Graphs for Theory-of-mind Reasoning 🆕NEW

  • 赛道归属: 多智能体心智推理（Theory-of-Mind/动态信念建模）
  • 核心创新点: 提出“动态信念图”（Dynamic Belief Graph）学习框架，将个体的隐式信念表示为随时间演化、相互依赖的图结构状态，用于支撑LLM在不确定环境下的ToM推理与行为预测。相较将信念视为静态独立变量或纯提示推理的方法，该图式表示旨在提升跨轮次一致性与因果连贯性，减少随对话/事件推进而产生的心理模型漂移。

  • 一句话总结: 该工作用可更新的图结构信念状态增强LLM的心智推理，使其更适用于动态、高风险的人机协作场景。

  • Track: Theory-of-Mind reasoning (dynamic belief modeling / belief graphs)

  • Core innovation: Introduces a Dynamic Belief Graph learning framework that represents agents’ implicit beliefs as a time-evolving, interdependent graph state to support ToM reasoning under uncertainty. Compared to direct prompting or static/independent latent belief models, the graph formulation targets improved cross-turn coherence and reduced mental-model drift as situations evolve.
  • One-sentence summary: Strengthens LLM ToM reasoning by maintaining an updatable graph-structured belief state, better matching dynamic high-stakes human-in-the-loop settings.

• [2026-03-20] Semantic Token Clustering for Efficient Uncertainty Quantification in Large Language Models 🆕NEW

  • 赛道归属: 不确定性量化与可靠性（高效UQ/语义级token聚类）
  • 核心创新点: 提出Semantic Token Clustering，通过对生成分布中的token进行语义聚类来估计不确定性，在不依赖多次采样或额外模型的前提下降低UQ计算开销。其关键在于将“表面token多样性”映射为“语义层面的等价类”，从而更稳健地度量模型在含同义表达时的真实不确定性。

  • 一句话总结: 该方法以语义聚类替代昂贵的重复采样，为LLM输出可靠性提供更轻量的、不确定性评估路径。

  • Track: Uncertainty quantification & reliability (efficient UQ / semantic token clustering)

  • Core innovation: Proposes Semantic Token Clustering to estimate uncertainty by clustering tokens in the generation distribution at the semantic level, reducing the need for repeated sampling or auxiliary models. By mapping surface-form variability into semantic equivalence classes, it aims to measure uncertainty more robustly under paraphrastic variation.
  • One-sentence summary: Delivers a lower-overhead route to LLM uncertainty estimation by leveraging semantic clustering instead of costly multi-sample UQ.

• [2026-03-20] Reasoning Gets Harder for LLMs Inside A Dialogue 🆕NEW

  • 赛道归属: 对话场景推理评测（任务型对话中的推理鲁棒性）
  • 核心创新点: 聚焦任务型对话（TOD）中“边生成边推理”的真实约束，系统分析对话框架、角色/格式/风格指令等因素如何使LLM推理变难，从而解释“单轮基准高分”与“对话中推理失真”之间的落差。通过将推理嵌入多轮指令遵循与文本生成过程，提出更贴近产品形态的推理鲁棒性评测视角。

  • 一句话总结: 该研究指出对话交互会系统性削弱LLM推理表现，推动从静态基准转向更真实的对话推理评估与改进。

  • Track: Dialogue reasoning evaluation (reasoning robustness in task-oriented dialogue)

  • Core innovation: Investigates why reasoning degrades in task-oriented dialogue where models must reason while simultaneously generating text under role/format/style constraints, highlighting a mismatch between isolated benchmark setups and real usage. By analyzing how dialogue framing and instruction-following pressures interact with reasoning, it motivates evaluation protocols that better reflect deployed TOD conditions.
  • One-sentence summary: Shows that dialogue constraints can systematically make LLM reasoning harder, arguing for more realistic dialogue-based reasoning benchmarks and mitigation strategies.

• [2026-03-20] Revisiting Gene Ontology Knowledge Discovery with Hierarchical Feature Selection and Virtual Study Group of AI Agents 🆕NEW

  • 赛道归属: 生物信息学知识发现（LLM智能体协作/层次特征选择）
  • 核心创新点: 将层次化特征选择与“虚拟学习小组”式多智能体协作结合，用于从Gene Ontology中围绕衰老相关主题进行知识发现与归纳。通过代理分工讨论/互评来提升信息提取与假设凝练质量，并利用GO层次结构进行特征筛选以降低噪声、增强生物学可解释性与检索聚焦度。

  • 一句话总结: 该工作把多智能体协作引入GO知识挖掘流程，并结合层次特征选择，提高衰老相关生物知识发现的系统性与可解释性。

  • Track: Bioinformatics knowledge discovery (multi-agent LLMs / hierarchical feature selection)

  • Core innovation: Combines hierarchical feature selection with a “virtual study group” of collaborating AI agents to extract and synthesize aging-related knowledge from the Gene Ontology. The approach uses agent role specialization and critique to improve hypothesis distillation, while leveraging GO’s hierarchy to filter features, reduce noise, and enhance biological interpretability and focus.
  • One-sentence summary: Introduces an agentic, hierarchy-aware pipeline for GO-based aging knowledge discovery, improving both systematic exploration and interpretability.

多模态大模型 / Multimodal Models

arXiv

• [2026-03-20] LumosX: Relate Any Identities with Their Attributes for Personalized Video Generation 🆕NEW

  • 赛道归属: 个性化视频生成（身份-属性绑定/人脸属性一致性）
  • 核心创新点: 针对多人物/多身份场景中“人脸身份与属性对齐不稳、组内一致性差”的痛点，引入显式的身份-属性关联建模，并配套构建面向人脸-属性对齐的数据资源与训练策略，使模型能在视频时序中稳定地把正确属性绑定到对应身份，同时兼顾前景人物与背景控制。

  • 一句话总结: 通过显式身份-属性绑定机制与数据支撑，显著提升个性化视频生成中多身份的属性一致性与可控性。

  • Track: Personalized video generation (identity–attribute binding / face-attribute consistency)

  • Core innovation: Addresses unstable face–attribute alignment in multi-identity videos by explicitly modeling identity–attribute relations and introducing face-attribute-aware data/training strategies, enabling temporally consistent attribute binding to the correct subject while retaining fine-grained foreground/background control.
  • One-sentence summary: Makes personalized video generation more reliable by ensuring attributes stay correctly attached to each identity across time.

• [2026-03-20] Adaptive Greedy Frame Selection for Long Video Understanding 🆕NEW

  • 赛道归属: 长视频多模态理解 / 视频问答推理优化（帧选择）
  • 核心创新点: 提出问题自适应的贪心帧选择策略，在选择过程中同时优化“与问题的相关性”和“语义多样性/覆盖度”，避免仅按相关性导致的近重复帧坍塌与证据缺失。通过在有限帧预算下更好覆盖时间上分散的关键证据，降低视觉token与推理成本并提升长视频QA稳健性。

  • 一句话总结: 用“相关性+覆盖度”的联合贪心选择，在不显著增加计算的前提下更可靠地抓住长视频中的决定性时刻。

  • Track: Long-video multimodal understanding / VideoQA inference optimization (frame selection)

  • Core innovation: Proposes a question-adaptive greedy frame selection method that jointly optimizes query relevance and semantic diversity/coverage, preventing relevance-only selection from collapsing into near-duplicate frames and missing temporally distant evidence. This improves robustness under tight frame/token budgets while reducing inference cost.
  • One-sentence summary: A relevance-plus-coverage greedy selector that captures decisive moments in long videos more reliably without paying extra compute.

• [2026-03-20] The Robot's Inner Critic: Self-Refinement of Social Behaviors through VLM-based Replanning 🆕NEW

  • 赛道归属: 具身智能 / 社交机器人多模态规划与自我改进（VLM辅助重规划）
  • 核心创新点: 提出CRISP闭环框架，让机器人使用VLM充当“类人社交批评者”对自身社交行为进行评估并触发重规划，而非依赖预定义动作库或人工反馈。系统性结合从机器人描述文件中抽取可动关节与约束、生成候选行为、基于批评进行迭代修正，实现更自主的社交行为自我精炼。

  • 一句话总结: 把VLM变成机器人的“内在批评机制”，用闭环批评-重规划实现社交行为的自主迭代优化。

  • Track: Embodied AI / Social-robot behavior planning and self-improvement (VLM-guided replanning)

  • Core innovation: Introduces CRISP, a closed-loop framework where a VLM acts as a human-like social critic to evaluate the robot’s own behaviors and trigger replanning, reducing reliance on scripted motions or human feedback. It integrates joint/constraint extraction from robot description files, candidate action generation, and iterative critique-driven refinement.
  • One-sentence summary: Turns a VLM into an “inner critic” that enables autonomous critique-and-replan loops for improving robot social behaviors.

• [2026-03-20] Detached Skip-Links and $R$-Probe: Decoupling Feature Aggregation from Gradient Propagation for MLLM OCR 🆕NEW

  • 赛道归属: 多模态模型训练优化 / MLLM OCR（特征融合与梯度干扰抑制）
  • 核心创新点: 揭示多层特征融合中skip-link带来的“梯度直通”会让高层语义目标反向覆盖早期视觉层的低层细节，从而损害OCR所需的精细视觉表征并导致训练不稳定。提出Detached Skip-Links将特征聚合与梯度传播解耦，并配合R-Probe诊断/度量该问题，实现更稳健的细粒度文本识别能力。

  • 一句话总结: 通过“断开式跳连”抑制语义梯度对低层视觉细节的污染，显著提升MLLM在OCR上的可训练性与细节保真。

  • Track: Multimodal training optimization / MLLM OCR (feature fusion & gradient interference mitigation)

  • Core innovation: Identifies that skip-links in multi-layer fusion create direct backprop paths that let high-level semantic objectives overwrite early low-level visual signals, destabilizing training and hurting OCR. Proposes Detached Skip-Links to decouple feature aggregation from gradient propagation, along with an R-Probe to diagnose/measure the issue, improving fine-grained text recognition.
  • One-sentence summary: Improves MLLM OCR by preventing semantic gradients from corrupting low-level visual detail via detached skip connections.

• [2026-03-20] MedSPOT: A Workflow-Aware Sequential Grounding Benchmark for Clinical GUI 🆕NEW

  • 赛道归属: 多模态评测基准 / 医疗GUI视觉定位与序列式工作流理解
  • 核心创新点: 提出MedSPOT基准，将临床软件中的GUI grounding从单步静态查询扩展为“工作流驱动的多步序列定位”，显式建模任务跨步骤演化与界面状态动态变化。该设定更贴近真实临床操作链路，用于系统性检验MLLM在高风险环境中的持续定位、状态跟踪与步骤级推理能力。

  • 一句话总结: 用“序列工作流”把医疗GUI定位评测拉回真实临床场景，暴露并量化MLLM在多步交互中的可靠性短板。

  • Track: Multimodal benchmarking / Clinical GUI visual grounding with sequential workflow reasoning

  • Core innovation: Introduces MedSPOT, a workflow-aware sequential grounding benchmark that moves beyond isolated single-step GUI queries to multi-step tasks with evolving goals and dynamic interface states. This better reflects real clinical workflows and stress-tests MLLMs’ grounding, state tracking, and step-wise reasoning in high-stakes software.
  • One-sentence summary: A realistic sequential clinical-GUI benchmark that measures whether MLLMs can ground reliably across evolving workflow steps.

• [2026-03-20] HiPath: Hierarchical Vision-Language Alignment for Structured Pathology Report Prediction 🆕NEW

  • 赛道归属: 医学多模态理解 / 病理图像到结构化报告生成（层级对齐）
  • 核心创新点: 提出HiPath，将病理报告预测从“扁平标签/自由文本”提升为以结构化、多粒度输出为核心训练目标，并通过层级式视觉-语言对齐来匹配报告中不同粒度字段（诊断、分级、部位等）。在冻结UNI2与Qwen3骨干的前提下，仅引入约15M可训练模块实现轻量适配，兼顾结构化可控输出与训练效率。

  • 一句话总结: 用轻量层级对齐把病理VLM从“会描述”推进到“会按结构写报告”，更贴近临床文书需求。

  • Track: Medical multimodal understanding / Pathology image-to-structured report generation (hierarchical alignment)

  • Core innovation: Proposes HiPath, treating structured, multi-granular pathology report prediction as the primary objective and aligning vision-language representations hierarchically to different report fields (diagnosis, grades, sites, etc.). Built on frozen UNI2 and Qwen3 backbones with only ~15M trainable parameters, enabling efficient adaptation with controllable structured outputs.
  • One-sentence summary: A lightweight hierarchical alignment approach that enables VLMs to generate clinically structured pathology reports rather than flat labels or free text.

• [2026-03-20] MedQ-Engine: A Closed-Loop Data Engine for Evolving MLLMs in Medical Image Quality Assessment 🆕NEW

  • 赛道归属: 医学多模态数据引擎 / 医学图像质量评估（闭环数据生成与模型演化）
  • 核心创新点: 提出MedQ-Engine闭环数据引擎，针对Med-IQA中“高成本描述性标注+模型弱点随训练变化”的痛点，通过迭代发现模型失效模式、定向采样/生成与补充高价值样本、再训练评估的循环机制，让数据分布随模型能力共同演化。重点支持带临床推理的描述性质量评估，而非仅回归分数，从而更贴近专家判读。

  • 一句话总结: 用闭环“找弱点—补数据—再进化”的数据引擎，把医学IQA从一次性标注推进到持续自我改进。

  • Track: Medical multimodal data engine / Medical image quality assessment (closed-loop data evolution)

  • Core innovation: Introduces MedQ-Engine, a closed-loop data engine that continuously identifies model failure modes and acquires/constructs targeted high-value samples to address evolving weaknesses, mitigating the cost of descriptive annotations. It emphasizes descriptive, clinically reasoned quality assessment beyond scalar scoring, enabling co-evolution of data and MLLMs.
  • One-sentence summary: A closed-loop pipeline that keeps improving medical IQA MLLMs by iteratively targeting and fixing their current weaknesses with new data.

• [2026-03-20] IsoCLIP: Decomposing CLIP Projectors for Efficient Intra-modal Alignment 🆕NEW

  • 赛道归属: 表征学习 / CLIP对齐优化（单模态对齐与高效投影器分解）
  • 核心创新点: 聚焦CLIP在图像-图像等单模态任务中的“单模态不对齐”问题，指出共享嵌入空间前的projector在其中扮演关键角色。通过分解/重构CLIP投影器以实现更高效的单模态对齐，在不改变主干编码器的情况下提升图像侧（或文本侧）内部检索与匹配性能。

  • 一句话总结: 从projector入手修复CLIP的单模态对齐缺陷，让CLIP更适合图像到图像等非跨模态检索场景。

  • Track: Representation learning / CLIP alignment optimization (intra-modal alignment via projector decomposition)

  • Core innovation: Studies CLIP’s intra-modal misalignment in image-to-image (and other unimodal) tasks and pinpoints the projector as a key contributor. Proposes decomposing/reworking CLIP projectors to achieve more efficient intra-modal alignment, improving unimodal retrieval/matching without modifying the backbone encoders.
  • One-sentence summary: Makes CLIP work better for unimodal retrieval by fixing intra-modal alignment through projector decomposition.

• [2026-03-20] From Plausibility to Verifiability: Risk-Controlled Generative OCR for Vision-Language Models 🆕NEW

  • 赛道归属: 多模态OCR / 可靠性与风险控制推理（可验证生成）
  • 核心创新点: 指出生成式OCR的部署错配：自回归解码偏向“语义合理”，而OCR需要“视觉可证伪/几何可验证”的输出，导致过生成与无依据替换等高风险错误。提出风险可控的生成式OCR框架，将输出与视觉证据进行可验证约束（如几何/定位一致性），把“看起来对”转为“可被证据证明对”，以降低长尾灾难性失误。

  • 一句话总结: 把生成式OCR从追求“合理”升级为追求“可验证”，用风险控制机制显著降低不可接受的识别错误。

  • Track: Multimodal OCR / Reliability & risk-controlled inference (verifiable generation)

  • Core innovation: Identifies a deployment misalignment in generative OCR: autoregressive decoding optimizes semantic plausibility, while OCR demands visually grounded, geometrically verifiable outputs—otherwise causing over-generation and unsupported substitutions. Proposes a risk-controlled generative OCR approach that enforces verifiability against visual evidence (e.g., geometric/grounding consistency) to reduce rare but severe failures.
  • One-sentence summary: Shifts generative OCR from “plausible” to “verifiable,” reducing high-risk errors through evidence-based constraints.

• [2026-03-20] One Model, Two Minds: Task-Conditioned Reasoning for Unified Image Quality and Aesthetic Assessment 🆕NEW

  • 赛道归属: 多模态图像评估 / 统一IQA与美学评估的任务条件推理
  • 核心创新点: 提出“One Model, Two Minds”的任务条件化推理范式，指出IQA与IAA在证据类型与推理策略上本质不同：IQA依赖低层客观失真线索、适合简洁聚焦的推理；IAA依赖语义与主观判断、需要更审慎的推理与不同的学习信号。通过为不同任务引入差异化的推理/奖励或训练目标，在单一模型内实现更匹配任务本质的统一评估。

  • 一句话总结: 在一个模型里为“质量”和“美学”配两套思维方式，避免一刀切推理策略带来的系统性失配。

  • Track: Multimodal image assessment / Unified IQA & aesthetic assessment with task-conditioned reasoning

  • Core innovation: Proposes “One Model, Two Minds,” arguing IQA and IAA require fundamentally different evidence and reasoning: IQA benefits from concise distortion-focused cues, while IAA needs deliberative semantic judgment and different learning signals. Introduces task-conditioned reasoning (and corresponding rewards/objectives) to unify both tasks in one model without a one-size-fits-all strategy.
  • One-sentence summary: A unified assessor that adapts its reasoning style to the task, aligning one model with two distinct evaluation mindsets (quality vs aesthetics).

---

Generated automatically by Daily AI Digest Agent 生成时间: 2026-03-23 09:00:25