HuggingFace 每日AI论文速递

duan

每天10分钟，带您快速了解当日HuggingFace热门AI论文内容。每个工作日更新，欢迎订阅。 📢播客节目在小宇宙、Apple Podcast平台搜索【HuggingFace 每日AI论文速递】 🖼另外还有图文版，可在小红书搜索并关注【AI速递】

  1. 4h ago

    2026.06.05 | ArcANE框架量化角色弧线；TIDE模型实现主动洞察

    【目录】本期的 15 篇论文如下： [00:31] 🎭 ArcANE: Do Role-Playing Language Agents Stay in Character at the Right Time?（ArcANE：角色扮演语言代理在正确时刻保持角色一致性吗？）[01:26] 🔍 TIDE: Proactive Multi-Problem Discovery via Template-Guided Iteration（TIDE：通过模板引导的迭代实现主动多问题发现）[02:27] 🤖 AdaPlanBench: Evaluating Adaptive Planning in Large Language Model Agents under World and User Constraints（AdaPlanBench：在世界与用户约束下评估大语言模型智能体的自适应规划能力）[03:14] 🎥 VideoKR: Towards Knowledge- and Reasoning-Intensive Video Understanding（VideoKR：迈向知识和推理密集型视频理解）[04:09] 🤖 RobotValues: Evaluating Household Robots When Human Values Conflict（机器人价值观：当人类价值观冲突时评估家用机器人）[05:01] 🌐 Reinforcement Learning Elicits Contextual Learning of Unseen Language Translation（强化学习引发对未见语言的上下文翻译学习）[05:58] 🎬 LoomVideo: Unifying Multimodal Inputs into Video Generation and Editing（LoomVideo：统一多模态输入的视频生成与编辑）[06:49] 📸 Personal AI Agent for Camera Roll VQA（个人相机胶卷视觉问答的AI助手）[07:36] 🧠 Rethinking Continual Experience Internalization for Self-Evolving LLM Agents（重新思考持续经验内化以实现自演化的大语言模型智能体）[08:27] ⚖ Complexity-Balanced Diffusion Splitting（复杂度平衡扩散分割）[09:28] 🤖 Dream.exe: Can Video Generation Models Dream Executable Robot Manipulation?（Dream.exe：视频生成模型能否构想出可执行的机器人操作？）[10:33] 🔬 Unsupervised Skill Discovery for Agentic Data Analysis（面向智能体数据分析的无监督技能发现）[11:25] 🔍 LLMs Can Leak Training Data But Do They Want To? A Propensity-Aware Evaluation of Memorization in LLMs（大型语言模型可能泄露训练数据，但它们愿意吗？一种基于倾向性的记忆评估方法）[12:17] 🎯 Towards One-to-Many Temporal Grounding（迈向一对多时序定位）[13:16] 💰 The Shadow Price of Reasoning: Economic Perspective on Optimal Budget Allocation for LLMs（推理的影子价格：大型语言模型最优预算分配的经济学视角） 【关注我们】您还可以在以下平台找到我们，获得播客内容以外更多信息小红书: AI速递 【赞助商】OpenClaw快报每天五分钟，听听 OpenClaw 快报，带你了解最新动态和业内讨论传送门 https://www.xiaoyuzhoufm.com/podcast/6a1732a2dffa135d0ab5ef43 在小宇宙查看该单集文稿

    14 min

  2. 1d ago

    2026.06.04 | 全模态统一框架；音频实时主动交互

    【目录】本期的 15 篇论文如下： [00:31] 🌌 Cosmos 3: Omnimodal World Models for Physical AI（宇宙3：面向物理AI的全模态世界模型）[01:36] 🎧 Audio Interaction Model（音频交互模型）[02:31] 🔍 Where Do Deep-Research Agents Go Wrong? Span-Level Error Localization in Agent Trajectories（深度研究型智能体错在哪里？智能体轨迹中的跨度级错误定位）[03:30] 🔍 Reproducing, Analyzing, and Detecting Reward Hacking in Rubric-Based Reinforcement Learning（在基于评分标准的强化学习中复现、分析与检测奖励作弊行为）[04:25] 🧭 OVO-S-Bench: A Hierarchical Benchmark for Streaming Spatial Intelligence in Multimodal LLMs（OVO-S-Bench：面向多模态大语言模型流式空间智能的分层基准）[05:27] ⚡ Qwen-Image-Flash: Beyond Objective Design（Qwen-Image-Flash：超越客观设计）[06:18] 🧠 M$^3$Eval: Multi-Modal Memory Evaluation through Cognitively-Grounded Video Tasks（M$^3$Eval：基于认知视频任务的多模态记忆评估）[07:13] 🎥 Echo-Infinity: Learning Evolving Memory for Real-Time Infinite Video Generation（回声无限：面向实时无限视频生成的可学习演化记忆）[08:14] 🧠 ThoughtFold: Folding Reasoning Chains via Introspective Preference Learning（思维折叠：通过内省偏好学习折叠推理链）[09:08] 🧪 Benchmarks are Not Enough: RAMP for Runtime Assessing of Agentic Models in Production Systems（基准测试并不足够：用于生产系统中智能体模型运行时评估的RAMP框架）[10:15] ⚡ Streaming Communication in Multi-Agent Reasoning（多智能体推理中的流式通信）[11:08] 🎯 Self-Distilled Policy Gradient（自蒸馏策略梯度）[12:13] 🧠 MemTrain: Self-Supervised Context Memory Training（MemTrain：自监督上下文记忆训练）[13:05] 🧩 Eliciting Complex Spatial Reasoning in MLLMs through Wide-Baseline Matching（通过宽基线匹配激发多模态大语言模型中的复杂空间推理能力）[14:11] 🤖 MMG2Skill: Can Agents Distill In-the-Wild Guides into Self-Evolving Skills?（MMG2Skill：智能体能否从野外指南中蒸馏出自我进化的技能？） 【关注我们】您还可以在以下平台找到我们，获得播客内容以外更多信息小红书: AI速递 【赞助商】OpenClaw快报每天五分钟，听听 OpenClaw 快报，带你了解最新动态和业内讨论传送门 https://www.xiaoyuzhoufm.com/podcast/6a1732a2dffa135d0ab5ef43 在小宇宙查看该单集文稿

    15 min

  3. 2d ago

    2026.06.03 | 信任区域教小模型；人形GPT追踪动作

    【目录】本期的 15 篇论文如下： [00:31] 🎯 Trust Region On-Policy Distillation（信任区域同策略蒸馏）[01:17] 🤖 Humanoid-GPT: Scaling Data and Structure for Zero-Shot Motion Tracking（人形GPT：扩展数据与结构实现零样本运动追踪）[02:07] 🧠 A Local Perturbation Theory for Cross-Domain Interference and Recovery in Multi-Domain RL（多领域强化学习中跨域干扰与恢复的局部微扰理论）[03:06] 🧠 World Models Meet Language Models: On the Complementarity of Concrete and Abstract Reasoning（世界模型与语言模型：具体与抽象推理的互补性）[03:57] 🏥 AutoMedBench: Towards Medical AutoResearch with Agentic AI Models（AutoMedBench：面向医疗自主研究的智能体AI模型基准）[05:09] 🖼 Decoupled Residual Denoising Diffusion Models for Unified and Data Efficient Image-to-Image Translation（解耦残差去噪扩散模型用于统一且数据高效的图像到图像翻译）[06:12] 😴 Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories（语言模型需要睡眠：学习自我修改与记忆巩固）[07:09] 🧩 TRON: Targeted Rule-Verifiable Online Environments for Visual Reasoning RL（TRON：面向视觉推理强化学习的目标驱动、规则可验证的在线环境）[08:07] 💬 $Ψ$-Bench: Evaluating Persona-Sensitive Influencing in Persuasive Dialogues（Ψ-Bench：评估说服性对话中个性感知影响能力）[09:08] 🧩 Decentralized Instruction Tuning: Conflict-Aware Splitting and Weight Merging（去中心化指令微调：冲突感知分割与权重合并）[10:05] 🎯 Small RL Controller, Large Language Model: RL-Guided Adaptive Sampling for Test-Time Scaling（小型强化学习控制器与大型语言模型：基于强化学习引导的自适应采样实现测试时扩展）[11:09] 📄 PaddleOCR-VL-1.6: Expanding the Frontier of Document Parsing with Under-Optimized Region Refinement and Progressive Post-Training（PaddleOCR-VL-1.6：通过欠优化区域精炼与渐进式后训练扩展文档解析前沿）[12:14] 🗺 PlatonicNav: Unveiling Semantic Correspondence in Navigation with Platonic Topological Maps（柏拉图导航：利用柏拉图拓扑图揭示导航中的语义对应关系）[13:16] 🔍 Diagnosing Harmful Continuation in Answer-Correct Long-CoT Training Traces（诊断正确答案长链思维训练轨迹中的有害延续）[14:05] 🎵 MERIT: Learning Disentangled Music Representations for Audio Similarity（MERIT：学习用于音频相似性的解耦音乐表示） 【关注我们】您还可以在以下平台找到我们，获得播客内容以外更多信息小红书: AI速递 【赞助商】OpenClaw快报每天五分钟，听听 OpenClaw 快报，带你了解最新动态和业内讨论传送门 https://www.xiaoyuzhoufm.com/podcast/6a1732a2dffa135d0ab5ef43 在小宇宙查看该单集文稿

    15 min

  4. 3d ago

    2026.06.02 | 多智能体框架生成可编辑图表；参数高效微调支撑百万个性化模型

    【目录】本期的 15 篇论文如下： [00:33] 🎨 Crafter: A Multi-Agent Harness for Editable Scientific Figure Generation from Diverse Inputs（Crafter：一种用于从多样化输入生成可编辑科学图形的多智能体框架）[01:39] 🧩 On the Scaling of PEFT: Towards Million Personal Models of Trillion Parameters（关于参数高效微调的规模化：迈向万亿参数级别的百万个性化模型）[02:35] 🧪 A Matter of TASTE: Improving Coverage and Difficulty of Agent Benchmarks（品味之道：提升智能体基准测试的覆盖度与难度）[03:25] 🌐 K-BrowseComp: A Web Browsing Agent Benchmark Grounded in Korean Contexts（K-BrowseComp：基于韩国语境的网页浏览代理基准测试）[04:21] ⚡ Draft-OPD: On-Policy Distillation for Speculative Draft Models（Draft-OPD：面向推测草稿模型的在策略蒸馏）[05:10] 🎓 VLMs are Good Teachers for Video Reasoning via Adaptive Test-Time Optimization（视觉语言模型作为视频推理的优质教师：通过自适应测试时优化）[06:18] 📡 X-Stream: Exploring MLLMs as Multiplexers for Multi-Stream Understanding（X-Stream：探索多模态大语言模型作为多流理解的多路复用器）[07:13] 🎬 VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion（VideoMLA：用于分钟级自回归视频扩散的低秩潜在KV缓存）[07:59] 🤖 SkillAdaptor: Self-Adapting Skills for LLM Agents from Trajectories（SkillAdaptor：面向LLM智能体的自适应技能从轨迹中学习）[08:54] 🧠 Which Pretraining Paradigm Better Serves Spatial Intelligence? An Empirical Comparison of Vision-Language and Video Generation Models（哪种预训练范式更好地服务于空间智能？视觉语言模型与视频生成模型的实证比较）[09:51] 🧠 NITP: Next Implicit Token Prediction for LLM Pre-training（NITP：面向大语言模型预训练的下一隐式词元预测）[10:50] 👀 Where to Look: Can Foundation Models Reach a Target Viewpoint Through Active Exploration?（该看向哪里：基础模型能否通过主动探索达到目标视角？）[11:46] 🎬 LVSA: Training-Free Sparse Attention for Long Video Diffusion（LVSA：面向长视频扩散的无训练稀疏注意力机制）[12:38] 🛑 ESPO: Early-Stopping Proximal Policy Optimization（早期停止的近端策略优化）[13:37] 🎤 StreamChar: Long-Horizon Streaming Character Audio-Video Generation with Decoupled Orchestration（StreamChar：基于解耦编排的长时流式角色音频-视频生成） 【关注我们】您还可以在以下平台找到我们，获得播客内容以外更多信息小红书: AI速递 【赞助商】OpenClaw快报每天五分钟，听听 OpenClaw 快报，带你了解最新动态和业内讨论传送门 https://www.xiaoyuzhoufm.com/podcast/6a1732a2dffa135d0ab5ef43 在小宇宙查看该单集文稿

    15 min

  5. 4d ago

    2026.06.01 | 知识蒸馏炼技能；表示强制破瓶颈

    【目录】本期的 15 篇论文如下： [00:30] 🧠 COLLEAGUE.SKILL: Automated AI Skill Generation via Expert Knowledge Distillation（COLLEAGUE.SKILL：通过专家知识蒸馏实现自动化AI技能生成）[01:17] 🧠 Representation Forcing for Bottleneck-Free Unified Multimodal Models（表示强制：无瓶颈统一多模态模型）[02:07] 🎙 SwanVoice: Expressive Long-Form Zero-Shot Speech Synthesis for Both Monologue and Dialogue（SwanVoice：面向独白与对话的表现力丰富长文本零样本语音合成）[02:58] 🔍 LongTraceRL: Learning Long-Context Reasoning from Search Agent Trajectories with Rubric Rewards（长迹强化学习：利用评分奖励从搜索代理轨迹中学习长上下文推理）[03:59] 🎧 Towards Streaming Synchronized Spatial Audio Generation via Autoregressive Diffusion Transformer（面向流式同步空间音频生成的自回归扩散Transformer）[04:48] 🖼 GGT-100K: Generative Ground Truth for Generalizable Real-World Image Restoration（GGT-100K：面向通用真实世界图像恢复的生成式真实标签）[05:39] 🎤 Comprehensive Benchmarking of Long-Form Speech Generation in Diverse Scenarios（多样化场景下长篇语音生成的综合基准测试）[06:46] 🛋 Function2Scene: 3D Indoor Scene Layout from Functional Specifications（从功能规格到场景：基于功能说明的3D室内布局生成）[07:36] 🎥 SANA-Streaming: Real-time Streaming Video Editing with Hybrid Diffusion Transformer（SANA-Streaming：基于混合扩散Transformer的实时流式视频编辑）[08:29] 🧠 Task-Focused Memorization for Multimodal Agents（面向多模态智能体的任务聚焦记忆机制）[09:30] 🤖 Exploring Autonomous Agentic Data Engineering for Model Specialization（探索面向模型专业化的自主代理数据工程）[10:15] 🎓 Not All Disagreement Is Learnable: Token Teachability in On-Policy Distillation（并非所有分歧都是可学习的：在线策略蒸馏中的令牌可教性）[11:10] 🧩 dMoE: dLLMs with Learnable Block Experts（dMoE：具有可学习块级专家机制的扩散大语言模型）[12:12] 🛠 Recovering Policy-Induced Errors: Benchmarking and Trajectory Synthesis for Robust GUI Agents（恢复策略诱导的错误：面向鲁棒GUI智能体的基准测试与轨迹合成）[13:07] 🛡 From Prompt Injection to Persistent Control: Defending Agentic Harness Against Trojan Backdoors（从提示注入到持久控制：防御智能体框架中的特洛伊后门） 【关注我们】您还可以在以下平台找到我们，获得播客内容以外更多信息小红书: AI速递 【赞助商】OpenClaw快报每天五分钟，听听 OpenClaw 快报，带你了解最新动态和业内讨论传送门 https://www.xiaoyuzhoufm.com/podcast/6a1732a2dffa135d0ab5ef43 在小宇宙查看该单集文稿

    15 min

  6. 4d ago

    【月末特辑】5月最火AI论文 | 多智能体世界建模；开源机器人VLA模型

    【目录】本期的 10 篇论文如下：[00:45] TOP1(🔥407) | 🌍 Gamma-World: Generative Multi-Agent World Modeling Beyond Two Players（Gamma-World：超越双玩家的生成式多智能体世界建模）[03:09] TOP2(🔥347) | 🤖 MolmoAct2: Action Reasoning Models for Real-world Deployment（MolmoAct2：面向实际部署的動作推理模型）[05:30] TOP3(🔥269) | 🔍 CiteVQA: Benchmarking Evidence Attribution for Trustworthy Document Intelligence（CiteVQA：为可信文档智能建立证据归因基准）[07:51] TOP4(🔥231) | 🧠 Mean Mode Screaming: Mean--Variance Split Residuals for 1000-Layer Diffusion Transformers（均值模式尖叫：面向千层扩散Transformer的均值-方差分裂残差）[10:04] TOP5(🔥219) | 🏗 MinT: Managed Infrastructure for Training and Serving Millions of LLMs（MinT：用于训练和服务数百万大语言模型的托管基础设施）[11:59] TOP6(🔥217) | 🧠 Heterogeneous Scientific Foundation Model Collaboration（异构科学基础模型协作）[14:17] TOP7(🔥210) | 🤖 Code as Agent Harness（代码作为智能体框架）[16:26] TOP8(🔥210) | 🧠 SkillOpt: Executive Strategy for Self-Evolving Agent Skills（SkillOpt：面向自进化智能体技能的执行策略）[18:39] TOP9(🔥204) | 🎯 DelTA: Discriminative Token Credit Assignment for Reinforcement Learning from Verifiable Rewards（DelTA：面向可验证奖励强化学习的判别性令牌信用分配）[20:25] TOP10(🔥195) | 🧠 Anti-Self-Distillation for Reasoning RL via Pointwise Mutual Information（基于点互信息的反自蒸馏用于推理强化学习） 【关注我们】您还可以在以下平台找到我们，获得播客内容以外更多信息小红书: AI速递 在小宇宙查看该单集文稿

    23 min

  7. 6d ago

    【周末特辑】5月第5周最火AI论文 | γ-World：多智能体世界建模新框架；SkillOpt：智能体技能优化新策略

    【目录】本期的 5 篇论文如下：[00:41] TOP1(🔥357) | 🌍 Gamma-World: Generative Multi-Agent World Modeling Beyond Two Players（Gamma-World：超越双玩家的生成式多智能体世界建模）[02:49] TOP2(🔥204) | 🧠 SkillOpt: Executive Strategy for Self-Evolving Agent Skills（SkillOpt：面向自进化智能体技能的执行策略）[04:37] TOP3(🔥131) | 🎯 DVAO: Dynamic Variance-adaptive Advantage Optimization for Multi-reward Reinforcement Learning（DVAO：面向多奖励强化学习的动态方差自适应优势优化）[06:17] TOP4(🔥124) | 🎯 LocateAnything: Fast and High-Quality Vision-Language Grounding with Parallel Box Decoding（LocateAnything：基于并行框解码的快速高质量视觉-语言定位）[08:56] TOP5(🔥110) | 🛡 AgentDoG 1.5: A Lightweight and Scalable Alignment Framework for AI Agent Safety and Security（AgentDoG 1.5：一种轻量级且可扩展的AI代理安全与安保对齐框架） 【关注我们】您还可以在以下平台找到我们，获得播客内容以外更多信息小红书: AI速递 在小宇宙查看该单集文稿

    12 min

  8. May 29

    2026.05.29 | AgentDoG 1.5实现毫秒级安全防护；Qwen-VLA统一跨任务动作建模。

    【目录】本期的 14 篇论文如下：[00:25] 🛡 AgentDoG 1.5: A Lightweight and Scalable Alignment Framework for AI Agent Safety and Security（AgentDoG 1.5：一种轻量级且可扩展的AI代理安全与安保对齐框架）[01:06] 🤖 Qwen-VLA: Unifying Vision-Language-Action Modeling across Tasks, Environments, and Robot Embodiments（Qwen-VLA：统一跨任务、环境和机器人本体的视觉-语言-动作建模）[02:02] 🌐 OmniRetrieval: Unified Retrieval across Heterogeneous Knowledge Sources（OmniRetrieval：跨异构知识源的统一检索）[02:52] 🎨 CollectionLoRA: Collecting 50 Effects in 1 LoRA via Multi-Teacher On-Policy Distillation（集合LoRA：通过多教师同策略蒸馏将50种效果收集到一个LoRA中）[03:47] 🎬 minWM: A Full-Stack Open-Source Framework for Real-Time Interactive Video World Models（minWM：一个用于实时交互式视频世界模型的全栈开源框架）[04:39] 🎥 YoCausal: How Far is Video Generation from World Model? A Causality Perspective（YoCausal：视频生成距离世界模型还有多远？一个因果视角）[05:42] 🎨 GenClaw: Code-Driven Agentic Image Generation（GenClaw：代码驱动的智能体图像生成）[06:40] ⚡ EarlyTom: Early Token Compression Completes Fast Video Understanding（EarlyTom：早期令牌压缩实现快速视频理解）[07:37] 🎯 UniSteer: Text-Guided Flow Matching in Activation Space for Versatile LLM Steering（UniSteer：文本引导的激活空间流匹配实现多功能大语言模型操控）[08:25] 🧠 How LoRA Remembers? A Parametric Memory Law for LLM Finetuning（LoRA如何记忆？大语言模型微调中的参数化记忆定律）[09:20] 🔗 LoMo: Local Modality Substitution for Deeper Vision-Language Fusion（本地模态替换：实现更深入的视觉-语言融合）[10:24] 🔍 LaRA: Layer-wise Representation Analysis for Detecting Data Contamination in RL Post-Training（LaRA：基于逐层表示分析的RL后训练数据污染检测方法）[11:16] 🧠 Skill0.5: Joint Skill Internalization and Utilization for Out-of-Distribution Generalization in Agentic Reinforcement Learning（Skill0.5：面向智能体强化学习中分布外泛化的技能内化与利用联合框架）[12:17] 🔍 Xetrieval: Mechanistically Explaining Dense Retrieval（Xetrieval：机制性解释稠密检索） 【关注我们】您还可以在以下平台找到我们，获得播客内容以外更多信息小红书: AI速递 在小宇宙查看该单集文稿

    14 min
See All (618)

Ratings & Reviews

5
out of 5
2 Ratings

About

每天10分钟，带您快速了解当日HuggingFace热门AI论文内容。每个工作日更新，欢迎订阅。 📢播客节目在小宇宙、Apple Podcast平台搜索【HuggingFace 每日AI论文速递】 🖼另外还有图文版，可在小红书搜索并关注【AI速递】

Information

You Might Also Like