HuggingFace 每日AI论文速递

本期的 15 篇论文如下：[00:20] 🔍 Watching, Reasoning, and Searching: A Video Deep Research Benchmark on Open Web for Agentic Video Reasoning（观察、推理与搜索：面向智能体视频推理的开放网络视频深度研究基准）[01:01] 👶 BabyVision: Visual Reasoning Beyond Language（BabyVision：超越语言的视觉推理）[01:45] 🚀 PaCoRe: Learning to Scale Test-Time Compute with Parallel Coordinated Reasoning（PaCoRe：通过并行协调推理学习扩展测试时计算）[02:24] 🧠 X-Coder: Advancing Competitive Programming with Fully Synthetic Tasks, Solutions, and Tests（X-Coder：基于全合成任务、解决方案与测试推进竞争性编程）[03:03] ⚡ MHLA: Restoring Expressivity of Linear Attention via Token-Level Multi-Head（MHLA：通过令牌级多头机制恢复线性注意力的表达能力）[03:41] ⚡ GlimpRouter: Efficient Collaborative Inference by Glimpsing One Token of Thoughts（GlimpRouter：通过瞥见思维令牌实现高效协同推理）[04:17] 🤖 OS-Symphony: A Holistic Framework for Robust and Generalist Computer-Using Agent（OS-Symphony：一个用于鲁棒且通用的计算机使用智能体的整体框架）[05:20] 📉 Lost in the Noise: How Reasoning Models Fail with Contextual Distractors（迷失于噪声之中：推理模型如何因上下文干扰物而失效）[06:00] 🚀 Beyond Hard Masks: Progressive Token Evolution for Diffusion Language Models（超越硬掩码：扩散语言模型的渐进式令牌演化）[06:30] 🧠 Controllable Memory Usage: Balancing Anchoring and Innovation in Long-Term Human-Agent Interaction（可控内存使用：在长期人机交互中平衡锚定与创新）[07:10] 🚗 DrivingGen: A Comprehensive Benchmark for Generative Video World Models in Autonomous Driving（DrivingGen：自动驾驶生成式视频世界模型的综合基准）[07:58] 🤖 MegaFlow: Large-Scale Distributed Orchestration System for the Agentic Era（MegaFlow：面向智能体时代的大规模分布式编排系统）[08:26] 🎨 Boosting Latent Diffusion Models via Disentangled Representation Alignment（通过解耦表征对齐提升潜在扩散模型）[09:08] 🤔 What Users Leave Unsaid: Under-Specified Queries Limit Vision-Language Models（用户未言明之处：欠明确的查询限制视觉语言模型）[09:45] 🔧 ET-Agent: Incentivizing Effective Tool-Integrated Reasoning Agent via Behavior Calibration（ET-Agent：通过行为校准激励有效的工具集成推理智能体）【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

本期的 15 篇论文如下：[00:20] 🗺 Thinking with Map: Reinforced Parallel Map-Augmented Agent for Geolocalization（借助地图思考：用于地理定位的强化并行地图增强智能体）[01:03] 🧠 MMFormalizer: Multimodal Autoformalization in the Wild（MMFormalizer：面向真实世界的多模态自动形式化方法）[01:38] 🧬 The Molecular Structure of Thought: Mapping the Topology of Long Chain-of-Thought Reasoning（思维分子结构：长链思维推理的拓扑映射）[02:21] 🎭 CaricatureGS: Exaggerating 3D Gaussian Splatting Faces With Gaussian Curvature（CaricatureGS：基于高斯曲率夸张3D高斯泼溅人脸）[03:04] 🔍 Chaining the Evidence: Robust Reinforcement Learning for Deep Search Agents with Citation-Aware Rubric Rewards（证据链构建：基于引文感知评分奖励的深度搜索智能体鲁棒强化学习）[03:47] ⚙ EnvScaler: Scaling Tool-Interactive Environments for LLM Agent via Programmatic Synthesis（EnvScaler：通过程序化合成扩展LLM智能体的工具交互环境）[04:22] 🔮 Can We Predict Before Executing Machine Learning Agents?（我们能在执行前预测机器学习智能体的行为吗？）[04:59] 🖼 AgentOCR: Reimagining Agent History via Optical Self-Compression（AgentOCR：通过光学自压缩重构智能体历史）[05:39] 🎬 VideoAR: Autoregressive Video Generation via Next-Frame & Scale Prediction（VideoAR：通过下一帧与尺度预测的自回归视频生成）[06:29] 🔍 Qwen3-VL-Embedding and Qwen3-VL-Reranker: A Unified Framework for State-of-the-Art Multimodal Retrieval and Ranking（Qwen3-VL-Embedding与Qwen3-VL-Reranker：用于最先进多模态检索与排序的统一框架）[07:23] 🔍 Illusions of Confidence? Diagnosing LLM Truthfulness via Neighborhood Consistency（自信的幻觉？通过邻域一致性诊断大语言模型的真实性）[08:07] 🔄 Orient Anything V2: Unifying Orientation and Rotation Understanding（Orient Anything V2：统一物体朝向与旋转理解的增强基础模型）[08:37] 🔍 SmartSearch: Process Reward-Guided Query Refinement for Search Agents（SmartSearch：面向搜索代理的流程奖励引导查询优化框架）[09:23] ⚙ Goal Force: Teaching Video Models To Accomplish Physics-Conditioned Goals（目标力：教导视频模型实现物理条件目标）[10:11] 📊 Same Claim, Different Judgment: Benchmarking Scenario-Induced Bias in Multilingual Financial Misinformation Detection（相同声明，不同判断：多语言金融虚假信息检测中场景诱导偏见的基准测试）【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

本期的 5 篇论文如下：[00:39] TOP1(🔥126) | 📈 GDPO: Group reward-Decoupled Normalization Policy Optimization for Multi-reward RL Optimization（GDPO：面向多奖励强化学习优化的组奖励解耦归一化策略优化）[02:31] TOP2(🔥108) | 🌍 NeoVerse: Enhancing 4D World Model with in-the-wild Monocular Videos（NeoVerse：利用野外单目视频增强4D世界模型）[04:40] TOP3(🔥107) | 🤖 Youtu-Agent: Scaling Agent Productivity with Automated Generation and Hybrid Policy Optimization（Youtu-Agent：通过自动化生成与混合策略优化扩展智能体生产力）[07:00] TOP4(🔥93) | 🔍 InfiniDepth: Arbitrary-Resolution and Fine-Grained Depth Estimation with Neural Implicit Fields（InfiniDepth：基于神经隐式场的任意分辨率与细粒度深度估计）[09:40] TOP5(🔥87) | 🎬 LTX-2: Efficient Joint Audio-Visual Foundation Model（LTX-2：高效的联合视听基础模型）【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

本期的 15 篇论文如下：[00:21] 📈 GDPO: Group reward-Decoupled Normalization Policy Optimization for Multi-reward RL Optimization（GDPO：面向多奖励强化学习优化的组奖励解耦归一化策略优化）[01:05] ⚖ Learnable Multipliers: Freeing the Scale of Language Model Matrix Layers（可学习的乘数：释放语言模型矩阵层的尺度）[01:33] 🌙 RL-AWB: Deep Reinforcement Learning for Auto White Balance Correction in Low-Light Night-time Scenes（RL-AWB：基于深度强化学习的低光照夜间场景自动白平衡校正）[02:07] 🤖 RoboVIP: Multi-View Video Generation with Visual Identity Prompting Augments Robot Manipulation（RoboVIP：基于视觉身份提示的多视角视频生成增强机器人操作）[02:56] 🤝 RelayLLM: Efficient Reasoning via Collaborative Decoding（RelayLLM：基于协作解码的高效推理框架）[03:31] 🌲 AT$^2$PO: Agentic Turn-based Policy Optimization via Tree Search（AT²PO：基于树搜索的智能体回合制策略优化）[04:24] 🤔 VideoAuto-R1: Video Auto Reasoning via Thinking Once, Answering Twice（VideoAuto-R1：通过思考一次、回答两次实现视频自动推理）[04:57] 🎬 VerseCrafter: Dynamic Realistic Video World Model with 4D Geometric Control（VerseCrafter：具有4D几何控制的动态逼真视频世界模型）[05:34] 🔍 The Illusion of Specialization: Unveiling the Domain-Invariant "Standing Committee" in Mixture-of-Experts Models（专业化的幻象：揭示混合专家模型中的领域不变“常务委员会”）[06:09] 🎯 Few Tokens Matter: Entropy Guided Attacks on Vision-Language Models（少数令牌至关重要：针对视觉语言模型的熵引导攻击）[06:40] 🎥 Plenoptic Video Generation（全光视频生成）[07:12] ⚖ Agent-as-a-Judge（智能体作为评审者）[07:43] 📄 DocDancer: Towards Agentic Document-Grounded Information Seeking（DocDancer：面向智能体化的文档驱动信息检索）[08:20] 🧠 Re-Align: Structured Reasoning-guided Alignment for In-Context Image Generation and Editing（Re-Align：基于结构化推理引导对齐的上下文图像生成与编辑）[09:05] 🧠 DiffCoT: Diffusion-styled Chain-of-Thought Reasoning in LLMs（DiffCoT：大语言模型中的扩散风格思维链推理）【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

本期的 15 篇论文如下：[00:21] ⚖ Entropy-Adaptive Fine-Tuning: Resolving Confident Conflicts to Mitigate Forgetting（熵自适应微调：解决置信冲突以缓解遗忘）[01:15] 🧠 Evolving Programmatic Skill Networks（演化式程序化技能网络）[01:51] 🧠 Atlas: Orchestrating Heterogeneous Models and Tools for Multi-Domain Complex Reasoning（Atlas：面向多领域复杂推理的异构模型与工具编排框架）[02:31] 📊 Benchmark^2: Systematic Evaluation of LLM Benchmarks（基准测试的基准测试：大语言模型评估基准的系统性评估）[03:12] 🎬 Klear: Unified Multi-Task Audio-Video Joint Generation（Klear：统一的多任务音视频联合生成）[03:53] 🎬 Choreographing a World of Dynamic Objects（动态物体的编排：一个通用生成式流水线）[04:36] ✅ Agentic Rubrics as Contextual Verifiers for SWE Agents（作为上下文验证器的智能评分标准在软件工程代理中的应用）[05:11] ⚗ MDAgent2: Large Language Model for Code Generation and Knowledge Q&A in Molecular Dynamics（MDAgent2：用于分子动力学代码生成与知识问答的大语言模型）[05:55] 🚀 E-GRPO: High Entropy Steps Drive Effective Reinforcement Learning for Flow Models（E-GRPO：高熵步驱动流模型的有效强化学习）[06:53] 🛡 RedBench: A Universal Dataset for Comprehensive Red Teaming of Large Language Models（RedBench：一个用于大型语言模型全面红队测试的通用数据集）[07:36] 📊 EpiQAL: Benchmarking Large Language Models in Epidemiological Question Answering for Enhanced Alignment and Reasoning（EpiQAL：面向增强对齐与推理的流行病学问答大语言模型基准评测）[08:15] 🧠 Enhancing Linguistic Competence of Language Models through Pre-training with Language Learning Tasks（通过语言学习任务预训练增强语言模型的语言能力）[08:48] 🔬 Why LLMs Aren't Scientists Yet: Lessons from Four Autonomous Research Attempts（为什么大语言模型还不是科学家：来自四次自主研究尝试的教训）[09:25] 🤖 ThinkRL-Edit: Thinking in Reinforcement Learning for Reasoning-Centric Image Editing（ThinkRL-Edit：基于强化学习的思维式推理中心图像编辑）[10:17] 🧠 MAGMA: A Multi-Graph based Agentic Memory Architecture for AI Agents（MAGMA：一种基于多图的AI智能体记忆架构）【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

本期的 15 篇论文如下：[00:25] 🔍 InfiniDepth: Arbitrary-Resolution and Fine-Grained Depth Estimation with Neural Implicit Fields（InfiniDepth：基于神经隐式场的任意分辨率与细粒度深度估计）[01:07] 🎙 MOSS Transcribe Diarize: Accurate Transcription with Speaker Diarization（MOSS转录与说话人分离：带说话人归属和时间戳的准确转录）[01:46] 🔬 SciEvalKit: An Open-source Evaluation Toolkit for Scientific General Intelligence（SciEvalKit：一个用于科学通用智能的开源评估工具包）[02:32] 🎬 LTX-2: Efficient Joint Audio-Visual Foundation Model（LTX-2：高效的联合视听基础模型）[03:26] 🦄 UniCorn: Towards Self-Improving Unified Multimodal Models through Self-Generated Supervision（UniCorn：通过自生成监督实现自改进统一多模态模型）[04:06] 🎨 DreamStyle: A Unified Framework for Video Stylization（DreamStyle：视频风格化的统一框架）[04:38] 🧠 CogFlow: Bridging Perception and Reasoning through Knowledge Internalization for Visual Mathematical Problem Solving（CogFlow：通过知识内化桥接感知与推理，用于视觉数学问题求解）[05:25] ⚡ MiMo-V2-Flash Technical Report（MiMo-V2-Flash 技术报告）[06:15] 🎮 NitroGen: An Open Foundation Model for Generalist Gaming Agents（NitroGen：通用游戏智能体的开放基础模型）[06:58] 🤖 SOP: A Scalable Online Post-Training System for Vision-Language-Action Models（SOP：一种可扩展的视觉-语言-动作模型在线后训练系统）[07:43] 🛡 OpenRT: An Open-Source Red Teaming Framework for Multimodal LLMs（OpenRT：一个用于多模态大语言模型的开源红队测试框架）[08:31] 📍 The Sonar Moment: Benchmarking Audio-Language Models in Audio Geo-Localization（声纳时刻：音频语言模型在音频地理定位中的基准测试）[09:14] 🔍 X-MuTeST: A Multilingual Benchmark for Explainable Hate Speech Detection and A Novel LLM-consulted Explanation Framework（X-MuTeST：一个用于可解释仇恨言论检测的多语言基准及一种新颖的LLM咨询解释框架）[09:57] 🧠 Parallel Latent Reasoning for Sequential Recommendation（并行潜在推理用于序列推荐）[10:27] 🤖 WebGym: Scaling Training Environments for Visual Web Agents with Realistic Tasks（WebGym：利用真实任务扩展视觉网络代理的训练环境）【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

本期的 15 篇论文如下：[00:21] 🧠 K-EXAONE Technical Report（K-EXAONE技术报告）[00:56] 🚀 NextFlow: Unified Sequential Modeling Activates Multimodal Understanding and Generation（NextFlow：统一序列建模激活多模态理解与生成）[01:36] 🎭 DreamID-V:Bridging the Image-to-Video Gap for High-Fidelity Face Swapping via Diffusion Transformer（DreamID-V：通过扩散Transformer弥合图像到视频的鸿沟以实现高保真人脸交换）[02:19] 🎨 VAR RL Done Right: Tackling Asynchronous Policy Conflicts in Visual Autoregressive Generation（VAR强化学习优化之道：解决视觉自回归生成中的异步策略冲突）[03:04] 🚀 GARDO: Reinforcing Diffusion Models without Reward Hacking（GARDO：无需奖励黑客攻击的扩散模型强化方法）[03:41] 🎨 VINO: A Unified Visual Generator with Interleaved OmniModal Context（VINO：一种具有交错式全模态上下文的统一视觉生成器）[04:17] ♾ InfiniteVGGT: Visual Geometry Grounded Transformer for Endless Streams（InfiniteVGGT：面向无尽流数据的视觉几何基础Transformer）[04:54] 🧠 Can LLMs Predict Their Own Failures? Self-Awareness via Internal Circuits（大型语言模型能否预测自身失败？通过内部电路实现自我感知）[05:23] 🚀 Falcon-H1R: Pushing the Reasoning Frontiers with a Hybrid Model for Efficient Test-Time Scaling（Falcon-H1R：通过混合模型实现高效测试时扩展，推动推理前沿）[05:57] 🔄 Talk2Move: Reinforcement Learning for Text-Instructed Object-Level Geometric Transformation in Scenes（Talk2Move：基于强化学习的文本指令场景物体几何变换框架）[06:43] 🔄 Recursive Language Models（递归语言模型）[07:12] 🧠 KV-Embedding: Training-free Text Embedding via Internal KV Re-routing in Decoder-only LLMs（KV-嵌入：通过仅解码器大语言模型内部KV重路由实现免训练文本嵌入）[07:51] ⚠ COMPASS: A Framework for Evaluating Organization-Specific Policy Alignment in LLMs（COMPASS：评估大语言模型中组织特定政策对齐性的框架）[08:52] 🛰 Toward Stable Semi-Supervised Remote Sensing Segmentation via Co-Guidance and Co-Fusion（通过协同引导与协同融合实现稳定的半监督遥感分割）[09:40] 🧱 SWE-Lego: Pushing the Limits of Supervised Fine-tuning for Software Issue Resolving（SWE-Lego：推动软件问题解决的监督微调极限）【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

本期的 12 篇论文如下：[00:22] 🤖 Youtu-Agent: Scaling Agent Productivity with Automated Generation and Hybrid Policy Optimization（Youtu-Agent：通过自动化生成与混合策略优化扩展智能体生产力）[00:52] 🌍 NeoVerse: Enhancing 4D World Model with in-the-wild Monocular Videos（NeoVerse：利用野外单目视频增强4D世界模型）[01:27] 🤖 Avatar Forcing: Real-Time Interactive Head Avatar Generation for Natural Conversation（Avatar Forcing：面向自然对话的实时交互式头部虚拟人生成）[01:59] 🚀 SenseNova-MARS: Empowering Multimodal Agentic Reasoning and Search via Reinforcement Learning（SenseNova-MARS：通过强化学习赋能多模态代理推理与搜索）[02:35] 🎭 Taming Hallucinations: Boosting MLLMs' Video Understanding via Counterfactual Video Generation（驯服幻觉：通过反事实视频生成提升多模态大语言模型的视频理解能力）[03:14] 🎬 AdaGaR: Adaptive Gabor Representation for Dynamic Scene Reconstruction（AdaGaR：用于动态场景重建的自适应Gabor表示）[03:47] 🧠 Deep Delta Learning（深度Delta学习）[04:11] 🧠 Nested Learning: The Illusion of Deep Learning Architectures（嵌套学习：深度学习架构的幻象）[04:47] 🧠 Diversity or Precision? A Deep Dive into Next Token Prediction（多样性还是精确性？深入探究下一个词元预测）[05:23] 🧠 Fast-weight Product Key Memory（快速权重乘积键值记忆）[05:58] 🧬 InfoSynth: Information-Guided Benchmark Synthesis for LLMs（InfoSynth：面向大语言模型的信息引导基准合成框架）[06:28] 🌀 MorphAny3D: Unleashing the Power of Structured Latent in 3D Morphing（MorphAny3D：释放结构化隐空间在3D形变中的力量）【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

本期的 5 篇论文如下：[00:33] TOP1(🔥132) | 🧠 mHC: Manifold-Constrained Hyper-Connections（mHC：流形约束的超连接）[02:32] TOP2(🔥100) | 🧠 Mindscape-Aware Retrieval Augmented Generation for Improved Long Context Understanding（面向提升长文本理解的思维景观感知检索增强生成）[04:45] TOP3(🔥94) | 🎬 InsertAnywhere: Bridging 4D Scene Geometry and Diffusion Models for Realistic Video Object Insertion（InsertAnywhere：连接4D场景几何与扩散模型以实现逼真的视频对象插入）[07:17] TOP4(🔥86) | 🔗 Coupling Experts and Routers in Mixture-of-Experts via an Auxiliary Loss（通过辅助损失耦合专家混合模型中的专家与路由器）[09:23] TOP5(🔥62) | 🎬 LiveTalk: Real-Time Multimodal Interactive Video Diffusion via Improved On-Policy Distillation（LiveTalk：通过改进的策略内蒸馏实现实时多模态交互式视频扩散）【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

本期的 3 篇论文如下：[00:19] 🧠 Dynamic Large Concept Models: Latent Reasoning in an Adaptive Semantic Space（动态大型概念模型：自适应语义空间中的潜在推理）[00:56] 🧠 DiffThinker: Towards Generative Multimodal Reasoning with Diffusion Models（DiffThinker：基于扩散模型的生成式多模态推理）[01:27] 🔄 On the Role of Discreteness in Diffusion LLMs（论离散性在扩散语言模型中的作用）【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

本期的 15 篇论文如下：[00:22] 🚀 Youtu-LLM: Unlocking the Native Agentic Potential for Lightweight Large Language Models（Youtu-LLM：解锁轻量级大语言模型的原生智能体潜力）[01:00] 🤖 Let It Flow: Agentic Crafting on Rock and Roll, Building the ROME Model within an Open Agentic Learning Ecosystem（任其流动：摇滚乐上的智能体构建，在开放智能体学习生态系统中建立ROME模型）[01:52] 🧠 mHC: Manifold-Constrained Hyper-Connections（mHC：流形约束的超连接）[02:25] 🔍 GaMO: Geometry-aware Multi-view Diffusion Outpainting for Sparse-View 3D Reconstruction（GaMO：面向稀疏视图三维重建的几何感知多视角扩散外绘）[03:11] 🔮 Scaling Open-Ended Reasoning to Predict the Future（扩展开放端推理以预测未来）[04:00] 🧠 AI Meets Brain: Memory Systems from Cognitive Neuroscience to Autonomous Agents（AI遇见大脑：从认知神经科学到自主智能体的记忆系统）[04:32] 🎬 PhyGDPO: Physics-Aware Groupwise Direct Preference Optimization for Physically Consistent Text-to-Video Generation（PhyGDPO：面向物理一致性文本到视频生成的物理感知分组直接偏好优化）[05:16] 🦾 GR-Dexter Technical Report（GR-Dexter技术报告）[05:59] 🎬 SpaceTimePilot: Generative Rendering of Dynamic Scenes Across Space and Time（SpaceTimePilot：跨时空动态场景的生成式渲染）[06:56] 🔍 Fantastic Reasoning Behaviors and Where to Find Them: Unsupervised Discovery of the Reasoning Process（神奇推理行为及其发现：推理过程的无监督探索）[07:28] 🧠 BEDA: Belief Estimation as Probabilistic Constraints for Performing Strategic Dialogue Acts（BEDA：将信念估计作为执行策略性对话行为的概率约束）[08:06] 🧭 Forging Spatial Intelligence: A Roadmap of Multi-Modal Data Pre-Training for Autonomous Systems（锻造空间智能：面向自主系统的多模态数据预训练路线图）[08:47] 🧠 Figure It Out: Improving the Frontier of Reasoning with Active Visual Thinking（图式求解：通过主动视觉思维提升推理前沿）[09:20] 🎯 Factorized Learning for Temporally Grounded Video-Language Models（面向时序定位视频语言模型的因子化学习）[09:59] 🎞 Pretraining Frame Preservation in Autoregressive Video Memory Compression（自回归视频记忆压缩中的预训练帧保留）【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

本期的 10 篇论文如下：[00:29] TOP1(🔥279) | 🧠 From Code Foundation Models to Agents and Applications: A Practical Guide to Code Intelligence（从代码基础模型到智能体与应用：代码智能实用指南）[02:22] TOP2(🔥242) | 🚀 DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models（DeepSeek-V3.2：推动开放大型语言模型前沿）[04:45] TOP3(🔥217) | 🚀 Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer（Z-Image：基于单流扩散Transformer的高效图像生成基础模型）[06:47] TOP4(🔥195) | ⚙ DataFlow: An LLM-Driven Framework for Unified Data Preparation and Workflow Automation in the Era of Data-Centric AI（DataFlow：面向数据为中心AI时代的统一数据准备与工作流自动化LLM驱动框架）[09:19] TOP5(🔥181) | 🎬 LongVT: Incentivizing "Thinking with Long Videos" via Native Tool Calling（LongVT：通过原生工具调用激励“长视频思考”）[11:39] TOP6(🔥167) | 🤖 Live Avatar: Streaming Real-time Audio-Driven Avatar Generation with Infinite Length（实时虚拟化身：基于无限时长的流式实时音频驱动化身生成）[13:15] TOP7(🔥163) | 🎬 Kling-Omni Technical Report（Kling-Omni技术报告）[15:12] TOP8(🔥149) | 📊 DAComp: Benchmarking Data Agents across the Full Data Intelligence Lifecycle（DAComp：跨全数据智能生命周期的数据智能体基准测试）[17:44] TOP9(🔥146) | 🧠 Qwen3-VL Technical Report（Qwen3-VL 技术报告）[20:34] TOP10(🔥128) | 🎬 Wan-Move: Motion-controllable Video Generation via Latent Trajectory Guidance（Wan-Move：通过潜在轨迹引导实现运动可控的视频生成）【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

本期的 6 篇论文如下：[00:24] 🧊 UltraShape 1.0: High-Fidelity 3D Shape Generation via Scalable Geometric Refinement（UltraShape 1.0：通过可扩展几何精化的高保真3D形状生成）[01:00] 🎨 DreamOmni3: Scribble-based Editing and Generation（DreamOmni3：基于涂鸦的编辑与生成）[01:34] 🧠 End-to-End Test-Time Training for Long Context（面向长上下文的端到端测试时训练）[02:18] 🔬 Evaluating Parameter Efficient Methods for RLVR（评估强化学习可验证奖励中的参数高效方法）[03:02] 🔍 GraphLocator: Graph-guided Causal Reasoning for Issue Localization（GraphLocator：基于图引导因果推理的缺陷定位方法）[03:35] ⚠ GateBreaker: Gate-Guided Attacks on Mixture-of-Expert LLMs（GateBreaker：针对专家混合大语言模型的基于门控的引导攻击）【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

本期的 15 篇论文如下：[00:24] 🔗 Coupling Experts and Routers in Mixture-of-Experts via an Auxiliary Loss（通过辅助损失耦合专家混合模型中的专家与路由器）[01:07] 🎬 LiveTalk: Real-Time Multimodal Interactive Video Diffusion via Improved On-Policy Distillation（LiveTalk：通过改进的策略内蒸馏实现实时多模态交互式视频扩散）[01:55] 🌍 Yume-1.5: A Text-Controlled Interactive World Generation Model（Yume-1.5：一种文本控制的交互式世界生成模型）[02:30] 🔍 SmartSnap: Proactive Evidence Seeking for Self-Verifying Agents（SmartSnap：自验证智能体的主动证据寻求范式）[02:59] 🔮 Diffusion Knows Transparency: Repurposing Video Diffusion for Transparent Object Depth and Normal Estimation（扩散模型知晓透明度：将视频扩散模型重新用于透明物体的深度与法线估计）[03:40] 🎯 SpotEdit: Selective Region Editing in Diffusion Transformers（SpotEdit：扩散变换器中的选择性区域编辑）[04:23] 🚀 Dream-VL & Dream-VLA: Open Vision-Language and Vision-Language-Action Models with Diffusion Language Model Backbone（Dream-VL与Dream-VLA：基于扩散语言模型骨干的开放视觉-语言与视觉-语言-动作模型）[05:09] 🔍 GRAN-TED: Generating Robust, Aligned, and Nuanced Text Embedding for Diffusion Models（GRAN-TED：为扩散模型生成鲁棒、对齐且细致的文本嵌入）[05:56] 🤖 Act2Goal: From World Model To General Goal-conditioned Policy（Act2Goal：从世界模型到通用目标条件策略）[06:31] ⚡ Stream-DiffVSR: Low-Latency Streamable Video Super-Resolution via Auto-Regressive Diffusion（Stream-DiffVSR：基于自回归扩散的低延迟可流式视频超分辨率）[06:59] 🌐 Web World Models（Web世界模型）[07:34] 🚀 DiRL: An Efficient Post-Training Framework for Diffusion Language Models（DiRL：一种高效的扩散语言模型后训练框架）[08:19] 🎬 Video-BrowseComp: Benchmarking Agentic Video Research on Open Web（Video-BrowseComp：面向开放网络的智能体视频研究基准测试）[09:02] 🧠 Training AI Co-Scientists Using Rubric Rewards（使用评分标准奖励训练AI科研助手）[09:39] 🧩 Monadic Context Engineering（单子上下文工程）【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

本期的 13 篇论文如下：[00:27] 🧠 Mindscape-Aware Retrieval Augmented Generation for Improved Long Context Understanding（面向提升长文本理解的思维景观感知检索增强生成）[01:07] 🎬 InsertAnywhere: Bridging 4D Scene Geometry and Diffusion Models for Realistic Video Object Insertion（InsertAnywhere：连接4D场景几何与扩散模型以实现逼真的视频对象插入）[01:46] 🤖 MAI-UI Technical Report: Real-World Centric Foundation GUI Agents（MAI-UI技术报告：面向真实世界的通用图形用户界面智能体）[02:22] 👁 UniPercept: Towards Unified Perceptual-Level Image Understanding across Aesthetics, Quality, Structure, and Texture（UniPercept：迈向跨美学、质量、结构与纹理的统一感知级图像理解）[03:04] 🎨 ProEdit: Inversion-based Editing From Prompts Done Right（ProEdit：基于反演的提示编辑的正确方法）[03:58] ⏱ TimeBill: Time-Budgeted Inference for Large Language Models（TimeBill：面向大语言模型的时间预算推理框架）[04:37] 🧠 See Less, See Right: Bi-directional Perceptual Shaping For Multimodal Reasoning（少看，看对：用于多模态推理的双向感知塑造）[05:16] 🌦 Omni-Weather: Unified Multimodal Foundation Model for Weather Generation and Understanding（Omni-Weather：用于天气生成与理解的多模态统一基础模型）[05:48] 🧠 SVBench: Evaluation of Video Generation Models on Social Reasoning（SVBench：视频生成模型在社会推理能力上的评估）[06:27] 🔍 InSight-o3: Empowering Multimodal Foundation Models with Generalized Visual Search（InSight-o3：赋能多模态基础模型实现广义视觉搜索）[07:15] 🎨 SlideTailor: Personalized Presentation Slide Generation for Scientific Papers（SlideTailor：面向科研论文的个性化演示文稿幻灯片生成）[08:11] 🤖 SWE-RM: Execution-free Feedback For Software Engineering Agents（SWE-RM：面向软件工程智能体的无执行反馈机制）[08:48] ⚡ A 58-Addition, Rank-23 Scheme for General 3x3 Matrix Multiplication（一种用于通用3x3矩阵乘法的58次加法、秩23方案）【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

本期的 5 篇论文如下：[00:42] TOP1(🔥188) | ⚙ DataFlow: An LLM-Driven Framework for Unified Data Preparation and Workflow Automation in the Era of Data-Centric AI（DataFlow：面向数据为中心AI时代的统一数据准备与工作流自动化LLM驱动框架）[02:34] TOP2(🔥105) | 🔬 Probing Scientific General Intelligence of LLMs with Scientist-Aligned Workflows（通过科学家对齐的工作流程探究大语言模型的科学通用智能）[05:04] TOP3(🔥85) | 🎬 SemanticGen: Video Generation in Semantic Space（SemanticGen：在语义空间中的视频生成）[07:03] TOP4(🔥73) | 🔍 Step-DeepResearch Technical Report（Step-DeepResearch技术报告）[09:31] TOP5(🔥71) | 🧠 PhysBrain: Human Egocentric Data as a Bridge from Vision Language Models to Physical Intelligence（PhysBrain：以人类第一人称数据为桥梁，从视觉语言模型迈向物理智能）【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

本期的 6 篇论文如下：[00:19] 🧠 Latent Implicit Visual Reasoning（潜在隐式视觉推理）[00:56] 🎬 Spatia: Video Generation with Updatable Spatial Memory（Spatia：基于可更新空间记忆的视频生成）[01:36] 🧠 Schoenfeld's Anatomy of Mathematical Reasoning by Language Models（基于舍恩菲尔德理论的语言模型数学推理解剖）[02:11] 🔍 How Much 3D Do Video Foundation Models Encode?（视频基础模型编码了多少3D信息？）[02:58] 🎯 VA-$π$: Variational Policy Alignment for Pixel-Aware Autoregressive Generation（VA-π：面向像素感知自回归生成的变分策略对齐）[03:36] 🚀 GTR-Turbo: Merged Checkpoint is Secretly a Free Teacher for Agentic VLM Training（GTR-Turbo：合并的检查点秘密成为智能体化视觉语言模型训练的免费教师）【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

本期的 14 篇论文如下：[00:20] 🧠 Learning to Reason in 4D: Dynamic Spatial Understanding for Vision Language Models（学习在四维空间中推理：视觉语言模型的动态空间理解）[01:11] ⚡ TurboDiffusion: Accelerating Video Diffusion Models by 100-200 Times（TurboDiffusion：将视频扩散模型加速100-200倍）[01:52] 🧭 T2AV-Compass: Towards Unified Evaluation for Text-to-Audio-Video Generation（T2AV-Compass：迈向文本到音视频生成的统一评估）[02:38] 🎬 DreaMontage: Arbitrary Frame-Guided One-Shot Video Generation（DreaMontage：基于任意帧引导的单镜头视频生成）[03:21] 🔍 Beyond Memorization: A Multi-Modal Ordinal Regression Benchmark to Expose Popularity Bias in Vision-Language Models（超越记忆：一个多模态序数回归基准揭示视觉语言模型中的流行度偏差）[04:07] 🎬 HiStream: Efficient High-Resolution Video Generation via Redundancy-Eliminated Streaming（HiStream：通过消除冗余的流式处理实现高效高分辨率视频生成）[04:52] 🚀 Nemotron 3 Nano: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning（Nemotron 3 Nano：用于智能体推理的开放、高效混合专家Mamba-Transformer模型）[05:38] 🔍 TokSuite: Measuring the Impact of Tokenizer Choice on Language Model Behavior（TokSuite：衡量分词器选择对语言模型行为的影响）[06:12] 🚀 NVIDIA Nemotron 3: Efficient and Open Intelligence（NVIDIA Nemotron 3：高效且开放的智能模型）[06:57] 🎬 Learning from Next-Frame Prediction: Autoregressive Video Modeling Encodes Effective Representations（基于下一帧预测的学习：自回归视频建模编码有效表示）[07:27] 🎬 Streaming Video Instruction Tuning（流式视频指令微调）[08:02] 🧠 Multi-hop Reasoning via Early Knowledge Alignment（通过早期知识对齐实现多跳推理）[08:43] 📊 SWE-EVO: Benchmarking Coding Agents in Long-Horizon Software Evolution Scenarios（SWE-EVO：在长周期软件演化场景中评估编码智能体的基准）[09:24] 🏆 LLM Swiss Round: Aggregating Multi-Benchmark Performance via Competitive Swiss-System Dynamics（LLM瑞士轮：通过竞争性瑞士制动态聚合多基准性能）【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

本期的 15 篇论文如下：[00:19] 🎬 SemanticGen: Video Generation in Semantic Space（SemanticGen：在语义空间中的视频生成）[01:01] 🔍 Bottom-up Policy Optimization: Your Language Model Policy Secretly Contains Internal Policies（自底向上策略优化：你的语言模型策略中暗含内部策略）[01:48] 🧠 SpatialTree: How Spatial Abilities Branch Out in MLLMs（SpatialTree：多模态大语言模型中的空间能力如何分支发展）[02:23] 🤖 LongVideoAgent: Multi-Agent Reasoning with Long Videos（LongVideoAgent：基于多智能体推理的长视频理解）[03:06] 🧠 MemEvolve: Meta-Evolution of Agent Memory Systems（MemEvolve：智能体记忆系统的元进化）[03:46] 🔍 Step-DeepResearch Technical Report（Step-DeepResearch技术报告）[04:22] 🎧 SAM Audio: Segment Anything in Audio（SAM Audio：音频中的任意分割）[05:00] 🚀 INTELLECT-3: Technical Report（INTELLECT-3：技术报告）[05:30] 🔍 FaithLens: Detecting and Explaining Faithfulness Hallucination（FaithLens：检测与解释忠实性幻觉）[06:07] 🧠 Reinforcement Learning for Self-Improving Agent with Skill Library（基于技能库与强化学习的自进化智能体研究）[06:53] 📊 QuantiPhy: A Quantitative Benchmark Evaluating Physical Reasoning Abilities of Vision-Language Models（QuantiPhy：评估视觉语言模型物理推理能力的定量基准）[07:38] 🔊 Simulstream: Open-Source Toolkit for Evaluation and Demonstration of Streaming Speech-to-Text Translation Systems（Simulstream：用于流式语音到文本翻译系统评估与演示的开源工具包）[08:18] 🧠 Active Intelligence in Video Avatars via Closed-loop World Modeling（通过闭环世界建模实现视频化身的主动智能）[08:55] 🔬 Multi-LLM Thematic Analysis with Dual Reliability Metrics: Combining Cohen's Kappa and Semantic Similarity for Qualitative Research Validation（基于多LLM与双重可靠性度量的主题分析：结合Cohen's Kappa与语义相似度进行定性研究验证）[09:32] ⚠ Toxicity Ahead: Forecasting Conversational Derailment on GitHub（毒性预警：预测GitHub对话中的脱轨行为）【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

本期的 15 篇论文如下：[00:22] ⚙ DataFlow: An LLM-Driven Framework for Unified Data Preparation and Workflow Automation in the Era of Data-Centric AI（DataFlow：面向数据为中心AI时代的统一数据准备与工作流自动化LLM驱动框架）[01:04] 🔍 The Prism Hypothesis: Harmonizing Semantic and Pixel Representations via Unified Autoencoding（棱镜假说：通过统一自编码协调语义与像素表示）[01:50] 🎬 Region-Constraint In-Context Generation for Instructional Video Editing（区域约束的上下文生成用于教学视频编辑）[02:33] 🎥 Infinite-Homography as Robust Conditioning for Camera-Controlled Video Generation（无限单应性变换作为相机控制视频生成的鲁棒条件）[03:08] 🔍 QuCo-RAG: Quantifying Uncertainty from the Pre-training Corpus for Dynamic Retrieval-Augmented Generation（QuCo-RAG：基于预训练语料的动态检索增强生成不确定性量化）[03:58] 🤔 Can LLMs Estimate Student Struggles? Human-AI Difficulty Alignment with Proficiency Simulation for Item Difficulty Prediction（大型语言模型能否评估学生困境？基于能力模拟的人机难度对齐用于试题难度预测）[04:35] 🧭 LoGoPlanner: Localization Grounded Navigation Policy with Metric-aware Visual Geometry（LoGoPlanner：基于定位与度量感知视觉几何的导航策略）[05:13] 🎬 WorldWarp: Propagating 3D Geometry with Asynchronous Video Diffusion（WorldWarp：利用异步视频扩散传播三维几何）[06:08] 🔍 UCoder: Unsupervised Code Generation by Internal Probing of Large Language Models（UCoder：通过内部探测大语言模型实现无监督代码生成）[06:45] 🧬 GenEnv: Difficulty-Aligned Co-Evolution Between LLM Agents and Environment Simulators（GenEnv：基于难度对齐的大语言模型智能体与环境模拟器协同进化框架）[07:22] 🎨 Reasoning Palette: Modulating Reasoning via Latent Contextualization for Controllable Exploration for (V)LMs（推理调色板：通过潜在情境化调节推理以实现（视觉）语言模型的可控探索）[07:56] ⚡ LoPA: Scaling dLLM Inference via Lookahead Parallel Decoding（LoPA：通过前瞻并行解码扩展扩散大语言模型推理）[08:38] 📱 MobileWorld: Benchmarking Autonomous Mobile Agents in Agent-User Interactive, and MCP-Augmented Environments（MobileWorld：在智能体-用户交互与MCP增强环境中评测自主移动智能体）[09:20] ⚖ Does It Tie Out? Towards Autonomous Legal Agents in Venture Capital（它能对上吗？迈向风险投资领域的自主法律智能体）[10:00] 🎬 StoryMem: Multi-shot Long Video Storytelling with Memory（StoryMem：基于记忆的多镜头长视频故事讲述）【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