HuggingFace 每日AI论文速递

duan
  • 5.0 (2)
  • Technology
  • Daily

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. 7h ago

    【周末特辑】10月第5周最火AI论文 | 把物理预测学进行动；多视图视频补上几何

    ⏱️ Highlights[00:29] 🤖 PDRVLA: A Vision-Language-Action Model with Physical Dynamics Representations（PDRVLA：具有物理动力学表征的视觉–语言–动作模型）[01:58] 🎥 OmniShowcase: Omnidirectional Subject-Consistent Video Generation from Multi-View Images（OmniShowcase：基于多视图图像的全方位主体一致视频生成）[03:27] 🎙️ Efficient Streaming Audio-Visual Target Speaker Extraction for Real-World Acoustic Scenes（面向真实世界声学场景的高效流式音视目标说话人提取）[05:02] 👁️ Look Only When Needed: Certificate-Guided Visual Routing for Efficient Diffusion VLMs（只在需要时看图：以凭据引导视觉路由的高效扩散视觉语言模型）[06:39] 🌍 Hy-Nomas: Towards Free-form Interactive World Generation（Hy-Nomas：迈向自由形式的交互式世界生成） 【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 在小宇宙查看该单集文稿

  2. 12h ago

    2026.11.02 | 视频先验统一生图；自博弈内化编排技能

    ⏱️ Highlights[00:25] 🎞️ FrameForge: Harnessing Priors of Video Foundation Models for Unified Image Generation（FrameForge：利用视频基础模型先验实现统一图像生成）[01:04] 🔁 R^2Orch: Recursive Self-Improvement for Agent Orchestration via Self-Play within a Skill-Internalization Loop（R^2Orch：在技能内化循环中通过自博弈实现智能体编排的递归自我改进）[01:42] 🧠 Does Scaling Toward AGI Depend on What System-1 Lets System-2 Execute?（迈向 AGI 的规模扩展，取决于 System-1 允许 System-2 执行什么吗？）[02:20] 🤖 REA: A Routed Embodied Agent Runtime for Asynchronous Brain-Cerebellum Control（REA：面向异步大脑—小脑控制的路由式具身智能体运行时）[02:58] 🎙️ Apriel Voice 27B Preview: An Open Model for Speech, Reasoning, and Tool Use（Apriel Voice 27B 预览版：面向语音、推理与工具使用的开放模型）[03:37] 🎮 Wizard: Manipulate Any World in Real Time via Agentic Control（Wizard：通过智能体控制实时操纵任意世界）[04:14] 💭 Latent Reasoning in Continuous Space for Unified Multimodal Models（统一多模态模型在连续空间中的潜在推理）[04:51] 🔭 HiFOR: Hierarchical Foresight with Action-Guided World Grounding for Vision-Language-Action Model（HiFOR：面向视觉—语言—动作模型的分层前瞻与动作引导世界信息锚定）[05:31] 🔎 WeFair: Fidelity-Aware Alignment from Human Feedback for Personalized Query Rewriting in Short-Video Search（WeFair：面向短视频搜索个性化查询改写的保真度感知人类反馈对齐）[06:11] ⚡ OSWAM: An Effective World Action Model with One-Step Action Generation（OSWAM：具备单步动作生成能力的高效世界动作模型）[06:49] 🧵 ReLift: Dense 3D Tracking as Lifted Trajectory Repair（ReLift：将稠密三维跟踪视为提升后轨迹的修复）[07:27] 🤝 CEPO: A General Framework for Improving Prompt Optimization in Multi-Agent Systems（CEPO：改进多智能体系统提示词优化的通用框架）[08:06] 🕵️ Venus-DefakerOne2: A Vision-Forensics Foundation Model with Broader Coverage（Venus-DefakerOne2：覆盖范围更广的视觉取证基础模型）[08:43] 🦾 CaPtain: A Pipeline for Post-Training Code-as-Policy Robot Agents（CaPtain：代码即策略机器人智能体的后训练流程）[09:22] 🧭 Probing Spatial State Representations in Visual Foundation Models（探测视觉基础模型中的空间状态表征） 【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

  3. 4d ago

    2026.10.29 | 扩散模型如何分功劳；嘈杂声场锁定目标声音

    ⏱️ Highlights[00:27] 🧠 DTCA: Denoising Token Credit Assignment for Reinforcement Learning in Diffusion Language Models（DTCA：用于扩散语言模型强化学习的去噪词元信用分配）[01:19] 🎧 Efficient Streaming Audio-Visual Target Speaker Extraction for Real-World Acoustic Scenes（面向真实声学场景的高效流式音视频目标说话人提取）[02:13] 🎬 OmniShowcase: Omnidirectional Subject-Consistent Video Generation from Multi-View Images（OmniShowcase：从多视图图像生成全方位主体一致的视频）[03:04] 🤖 Sensorimotor Alignment: A Platonic Proxy for Embodied Policy Success（感觉运动对齐：具身策略成功表现的柏拉图式代理指标）[03:56] ⚙️ TAO: Tool-Aware Joint Scheduling for LLM-Based Agentic Serving（TAO：面向基于大语言模型的智能体服务的工具感知联合调度）[04:53] 💻 RealCLI: Toward a Unified Evaluation of Human-Agent Collaboration for CLI Agents（RealCLI：迈向命令行智能体的人机协作统一评估）[05:43] 🏦 TauIndianBankBench: Procedure-Aware, Action-Oriented Evaluation of Tool-Using Conversational Agents（TauIndianBankBench：面向工具使用型对话智能体的流程感知、行动导向评估）[06:36] 📚 LongZero: Enhancing Long-Context Reasoning in LLMs from Zero Data（LongZero：从零数据增强大语言模型的长上下文推理）[07:27] 🔗 CausGround: Causal Interaction Grounding for Adaptive Embodied Planning（CausGround：面向自适应具身规划的因果交互落地）[08:18] 🦾 HuGe: Scaling Humanoid Generalist Vision-Language-Action Models with Egocentric Experience（HuGe：利用第一视角经验扩展人形通用视觉—语言—动作模型）[09:13] 🔭 OmniSpectra: A Unified Foundation Model for Universal Spectra Representation Learning（OmniSpectra：用于通用光谱表征学习的统一基础模型） 【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

  4. Oct 24

    【周末特辑】10月第4周最火AI论文 | 业务智能体长时程评估；具身智能跨门槛未实用

    【目录】本期的 5 篇论文如下： [00:43] TOP1(🔥494) | ⚠ BizBench: Unveiling Hidden Challenges in Deploying Long-Horizon Agents for Real-World Business（BizBench：揭示真实业务中部署长时程智能体的隐藏挑战）[02:15] TOP2(🔥472) | 🤖 Has GPT-6-Astra Achieved Embodied Intelligence?（GPT-6-Astra 是否已实现具身智能？）[04:17] TOP3(🔥438) | 🎬 StreamEval: Evaluating Interactive Streaming Video Generation Models（StreamEval：评估交互式流式视频生成模型）[06:21] TOP4(🔥332) | 🔍 Spectral-Spatial Interpretation（光谱-空间解释）[07:55] TOP5(🔥246) | 🧠 Scaling Parameter-Efficient Fine-Tuning for Brain Encoding（扩展用于大脑编码的参数高效微调） 【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

  5. Oct 23

    2026.10.23 | MLLM潜变量赋能电影视频生成；开放网络搜索基准暴露瓶颈

    【目录】本期的 15 篇论文如下： [00:24] 🎬 Making MLLM Latents Matter: Factor-Aware Predictive Learning for Cinematic Video Generation（让MLLM潜变量发挥作用：面向电影级视频生成的因子感知预测学习）[01:07] 🔍 Deep Search in Real-World Web Environments（真实世界网络环境中的深度搜索）[01:48] 👁 WeVisOne-VL: Unified Visual Perception from Objects to Pixels（WeVisOne-VL：从物体到像素的统一视觉感知）[02:34] 🧩 Editing Images into Reusable Layers via Compositional Edit-Prior Distillation（通过组合式编辑先验蒸馏将图像编辑为可复用图层）[03:15] 🤖 IQuest-Q1: Advancing Foundation Capabilities for Agentic CLI Systems（IQuest-Q1：提升面向智能体 CLI 系统的基础能力）[03:58] 🤖 Toward Recursive Self-Improvement in Physical AI（迈向物理人工智能中的递归自我改进）[04:38] 🧠 Deep Fusion Memory: A Chunk-Level Parametric Memory for Large Language Models（深度融合记忆：面向大语言模型的块级参数化记忆）[05:27] 🔍 What Generation Knows: Unlocking Fine-Grained Understanding in Unified Multimodal Models（生成知道什么：解锁统一多模态模型中的细粒度理解）[06:18] 🌍 LaLa-World: Scaling Latent World Models with Natural Language Actions（LaLa-World：以自然语言动作扩展潜在世界模型）[07:11] 🧩 Meta Expert: Cross-Layer Expert Sharing for Storage- and Retrieval-Efficient MoE（Meta Expert：面向存储与检索高效的跨层专家共享 MoE）[08:05] 🎧 Heartbeat Thinking: Speculative Reasoning and Tool Execution While Listening in Realtime Voice Agents（心跳思考：实时语音代理在聆听过程中的推测性推理与工具执行）[08:47] 🎛 Policy-Mask: Budget-Aware Adaptation of Diffusion Features to Discriminative Tasks（Policy-Mask：面向判别任务的扩散特征预算感知适配）[09:31] 🎬 Code2Video-Compass: A Comprehensive Benchmark for General-Purpose Code-Based Video Generation across T2V, I2V, and V2V（Code2Video-Compass：面向T2V、I2V和V2V的通用代码视频生成综合基准）[10:16] 🧠 Cola2: Co-Adaptive Hierarchical Language Modeling with Continuous Semantic Compression and Balanced Discrete Generation（Cola2：具有连续语义压缩与平衡离散生成的协同自适应分层语言建模）[11:01] 🎥 Beyond Watching and Searching: Benchmarking Agents for Video-centered Deep Research Reports（超越观看与搜索：面向以视频为中心的深度研究报告的智能体基准测试） 【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

  6. Oct 22

    2026.10.22 | 流式视频评测揭示时序瓶颈；VLM三维建图仍难精准

    【目录】本期的 15 篇论文如下： [00:25] 🎬 StreamEval: Evaluating Interactive Streaming Video Generation Models（StreamEval：评估交互式流式视频生成模型）[01:08] 🗺 SpatialScout: Can VLM Agents Actively Explore and Map 3D Scenes?（SpatialScout：VLM智能体能否主动探索并构建3D场景地图？）[01:55] 🎞 Looking Ahead: Modeling Future Evolution in Autoregressive Video Generation（展望未来：自回归视频生成中的未来演化建模）[02:42] 🗄 Cache Engineering for Agents（面向智能体的缓存工程）[03:24] 👐 DONIC: Generalist Dexterous Bimanual Tracking Control from Human Demonstrations（DONIC：基于人类演示的通用灵巧双手跟踪控制）[04:11] 🌍 Agent World Model in the Wild: Computer-Use Agent Environment Synthesis（真实环境中的智能体世界模型：计算机使用智能体环境合成）[04:52] 🔍 Spectral-Spatial Interpretation（光谱-空间解释）[05:36] 🧩 PrismaCoder: Pushing the Boundaries of Open Multimodal Code Intelligence through Decomposed Rubric Synthesis（PrismaCoder：通过分解式评分细则合成突破开放多模态代码智能的边界）[06:22] 🎓 Socratic Skill Learning: Interleaving Teaching and Practice for Transferable Agent Skills（苏格拉底式技能学习：交错教学与实践以培养可迁移的智能体技能）[07:07] 🎮 Scaling in Games: Continued Pre-Training For Embodied Agents In Diverse Virtual Worlds（游戏中的规模化：面向多样虚拟世界具身智能体的持续预训练）[07:43] 🤖 AgenticRoboRL: Bringing Coding-Agent Post-Training to Agentic Robotics（AgenticRoboRL：将编程智能体后训练引入智能体机器人）[08:31] 🧠 Memory-Centric Continual Learning for AI Agents（以记忆为中心的AI智能体持续学习）[09:17] 🤖 Multimodal Embodied Agents: A Unified Capability-centric Perspective from Computer-Use to Robot-Use（多模态具身智能体：从计算机使用到机器人使用的统一能力中心视角）[10:00] 🧠 TraceForge: Distilling with Your Own Thinking（TraceForge：用自身思考进行蒸馏）[10:40] 🛡 Right Screen, Wrong Transition: World Models as Verifiers for GUI Agents（正确的屏幕，错误的转移：世界模型作为GUI智能体的验证器） 【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

  7. Oct 21

    2026.10.21 | 全模态GUI动态聚焦；状态跟踪进度对齐

    【目录】本期的 15 篇论文如下： [00:23] 🎯 Focused Omni-Modal GUI Agents（聚焦式全模态GUI智能体）[01:06] 🔁 PAR: Progress-Aligned Recurrence for Length Generalization in State Tracking（PAR：面向状态跟踪长度泛化的进度对齐循环）[01:52] 🎮 Matrix-Game 4.0: Harnessing Video Models for Real-Time Playable Game Worlds（Matrix-Game 4.0：利用视频模型打造实时可玩游戏世界）[02:38] 🤖 Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human Experience（Ropedia W0：借助人类经验跨具身扩展移动操作世界模型）[03:23] 🧭 S-Agent 2.0: Spatial Reasoning with a General Coding Agent（S-Agent 2.0：基于通用编程智能体的空间推理）[04:08] 🧪 Sage-A1.5: From Long-Horizon Execution to Scientific Exploration with Capability Composition（Sage-A1.5：从长程执行到基于能力组合的科学探索）[04:55] 🔄 VLA-Rolling: Fewer Denoising Steps, Better Control via Adaptive Action Reuse（VLA-Rolling：通过自适应动作复用实现更少去噪步骤与更好控制）[05:41] 🧠 CoEvo-RSI: When Harnesses Evolve, Which Models Should Follow?（CoEvo-RSI：当执行框架演化时，哪些模型应当跟进？）[06:25] 🎥 Learning the 4D World through Point Trajectory Forecasting（通过点轨迹预测学习四维世界）[07:11] 🎬 SolarRenderer: Few-Step Generative Reference Rendering（SolarRenderer：少步生成式参考渲染）[07:56] 🚀 SWIFT: Self-guided Weak-to-strong dIFfusion models via Target extrapolation（SWIFT：基于目标外推的自引导弱到强扩散模型）[08:42] 🛠 Correction Training for Diffusion Models（扩散模型的校正训练）[09:31] 🏢 MMEnterprise-Bench: Evaluating Enterprise Agents over Evolving Multimodal Records（MMEnterprise-Bench：在持续演进的多模态记录上评估企业智能体）[10:18] 🧩 LLM2R: From Large Language Models to 3D Reassemblers（LLM2R：从大语言模型到3D重装配器）[10:58] ⚡ PhysSkin++: Real-Time and Generalizable Physics-Based Animation via Self-Supervised Neural Skinning（PhysSkin++：基于自监督神经蒙皮的实时可泛化物理动画） 【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

  8. Oct 20

    2026.10.20 | 机群经验赋能机器人群体学习；原生记忆驱动基础模型演化

    【目录】本期的 15 篇论文如下： [00:22] 🤖 Learning from Fleet Experience: A Collective Learning Paradigm for Robots（从机群经验中学习：一种面向机器人的集体学习范式）[01:12] 🧠 MemT-1.0: Towards Memory-Native Evolution（MemT-1.0：迈向记忆原生演化）[02:00] 🎥 EgoXperience: Learning World-Grounded Egocentric Representations from Human Experience（EgoXperience：从人类经验中学习世界锚定的自我中心表征）[02:47] 🧠 Scaling Parameter-Efficient Fine-Tuning for Brain Encoding（扩展用于大脑编码的参数高效微调）[03:39] 🌍 Holi-World: Building Holistic World Intelligence from Human-Ego Experience（Holi-World：从人类自我中心经验构建整体世界智能）[04:24] 🧰 OpenVLHarness: Real World Visual Agentic Tool-Use with Multimodal Memory（OpenVLHarness：面向真实世界的视觉智能体工具使用与多模态记忆）[05:08] 🧭 ScaleNav: Scalable Internet Video Pretraining for Language-Conditioned Local Navigation（ScaleNav：用于语言条件局部导航的可扩展互联网视频预训练）[05:49] 🪆 SpecMRL: Redesign Matryoshka Representation Learning within Parametric Spectrum（SpecMRL：在参数化频谱中重新设计套娃表示学习）[06:42] 🔁 Self-Evolving Coding Agents via Harness-Policy Optimization（通过执行框架与策略优化实现自进化编程智能体）[07:35] 🖼 MT Forcing: Towards Stable and Consistent Multi-Turn Image Editing with Autoregressive Diffusion（MT Forcing：面向稳定且一致的多轮图像编辑的自回归扩散方法）[08:18] 🧮 OR-License: Evaluating LLM Agents as Real-World Operations Research Practitioners（OR-License：评估LLM智能体作为真实世界运筹学实践者）[09:15] 🤝 CoworkEnv: Synthesizing Training Environments for Long-Horizon Cowork Agents（CoworkEnv：面向长时程协作智能体的训练环境合成）[10:00] 🧩 AnyGoal: On World Model Generalization Beyond Expert Rollouts（AnyGoal：论世界模型在专家轨迹之外的泛化）[10:42] 🖼 RoAR-VAE: Improving Latent Diffusability through Robust Prediction（RoAR-VAE：通过鲁棒预测提升潜变量可扩散性）[11:20] 🌍 InfiniDream: Open-Ended Photorealistic 3D Worlds Through Procedural Generation（InfiniDream：通过程序化生成开放式照片级真实感3D世界） 【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

See All (732)

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