HuggingFace 每日AI论文速递 - 节目列表

本期的 10 篇论文如下：[00:39] TOP1(🔥196) | 🤖 SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model（SmolLM2：当小型模型走向大型化——以数据为中心的小型语言模型训练）[02:32] TOP2(🔥183) | 🎥 OmniHuman-1: Rethinking the Scaling-Up of One-Stage Conditioned Human Animation Models（OmniHuman-1：重新思考一阶段条件人类动画模型的扩展）[05:02] TOP3(🔥182) | 🦜 The Stochastic Parrot on LLM's Shoulder: A Summative Assessment of Physical Concept Understanding（随机鹦鹉在大语言模型肩上：物理概念理解的总结性评估）[06:41] TOP4(🔥167) | 🧠 MLGym: A New Framework and Benchmark for Advancing AI Research Agents（MLGym：推进AI研究代理的新框架与基准）[09:03] TOP5(🔥152) | 🌐 Qwen2.5-VL Technical Report（Qwen2.5-VL 技术报告）[11:48] TOP6(🔥152) | 🔍 LLM-Microscope: Uncovering the Hidden Role of Punctuation in Context Memory of Transformers（LLM显微镜：揭示标点符号在Transformer上下文记忆中的隐藏作用）[13:41] TOP7(🔥142) | 🚀 InfiniteHiP: Extending Language Model Context Up to 3 Million Tokens on a Single GPU（InfiniteHiP：在单个GPU上扩展语言模型上下文至300万 tokens）[16:06] TOP8(🔥140) | 🤔 Can 1B LLM Surpass 405B LLM? Rethinking Compute-Optimal Test-Time Scaling（10亿参数LLM能否超越4050亿参数LLM？重新思考计算最优的测试时缩放）[18:40] TOP9(🔥137) | ⚡ Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention（原生稀疏注意力：硬件对齐与原生可训练的稀疏注意力）[20:46] TOP10(🔥125) | 💼 Expect the Unexpected: FailSafe Long Context QA for Finance（预料之外：金融领域长上下文问答的FailSafe）【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

本期的 5 篇论文如下：[00:50] TOP1(🔥152) | 🔍 LLM-Microscope: Uncovering the Hidden Role of Punctuation in Context Memory of Transformers（LLM显微镜：揭示标点符号在Transformer上下文记忆中的隐藏作用）[03:08] TOP2(🔥89) | 📚 SurveyX: Academic Survey Automation via Large Language Models（SurveyX 基于大型语言模型的学术调查自动化系统）[05:42] TOP3(🔥65) | 🎥 VideoGrain: Modulating Space-Time Attention for Multi-grained Video Editing（视频粒度：调节时空注意力实现多粒度视频编辑）[07:24] TOP4(🔥64) | 📖 Thus Spake Long-Context Large Language Model（长上下文大语言模型如是说）[09:35] TOP5(🔥61) | 🤖 OmniAlign-V: Towards Enhanced Alignment of MLLMs with Human Preference（OmniAlign-V：迈向多模态大语言模型与人类偏好增强对齐）【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

本期的 19 篇论文如下：[00:23] 🧠 Self-rewarding correction for mathematical reasoning（自我奖励的数学推理校正）[01:03] 🧠 MedVLM-R1: Incentivizing Medical Reasoning Capability of Vision-Language Models (VLMs) via Reinforcement Learning（MedVLM-R1：通过强化学习激励视觉语言模型的医疗推理能力）[01:53] 🧠 R2-T2: Re-Routing in Test-Time for Multimodal Mixture-of-Experts（R2-T2：测试时重路由在多模态专家混合模型中的应用）[02:34] 🧬 LongRoPE2: Near-Lossless LLM Context Window Scaling（LongRoPE2：近乎无损的LLM上下文窗口扩展）[03:11] 🧠 FINEREASON: Evaluating and Improving LLMs' Deliberate Reasoning through Reflective Puzzle Solving（FINEREASON：通过反思性谜题解决评估和改进大语言模型的深思熟虑推理）[04:02] 🤖 CODESYNC: Synchronizing Large Language Models with Dynamic Code Evolution at Scale（CODESYNC：大规模动态代码演化与大型语言模型同步）[04:48] 🚀 Lean and Mean: Decoupled Value Policy Optimization with Global Value Guidance（精简与高效：基于全局价值引导的解耦价值策略优化）[05:33] 🧩 UniTok: A Unified Tokenizer for Visual Generation and Understanding（UniTok：面向视觉生成与理解的统一分词器）[06:12] 🚀 NeoBERT: A Next-Generation BERT（NeoBERT：下一代BERT）[06:47] 🌀 FlexiDiT: Your Diffusion Transformer Can Easily Generate High-Quality Samples with Less Compute（FlexiDiT：让你的扩散Transformer轻松生成高质量样本，计算量更少）[07:30] 🛠 SoRFT: Issue Resolving with Subtask-oriented Reinforced Fine-Tuning（SoRFT：面向子任务的强化微调问题解决方法）[08:07] 🤖 Building Interactable Replicas of Complex Articulated Objects via Gaussian Splatting（基于高斯样条构建复杂 articulated 物体的交互式副本）[08:45] 🎨 Multimodal Representation Alignment for Image Generation: Text-Image Interleaved Control Is Easier Than You Think（多模态表示对齐用于图像生成：文本-图像交错控制比你想象的更简单）[09:30] 🎥 Mobius: Text to Seamless Looping Video Generation via Latent Shift（Mobius：通过潜在位移从文本生成无缝循环视频）[10:08] 🛡 Guardians of the Agentic System: Preventing Many Shots Jailbreak with Agentic System（代理系统守护者：通过代理系统防止多次越狱）[10:49] 🤖 R1-T1: Fully Incentivizing Translation Capability in LLMs via Reasoning Learning（通过推理学习全面激励大语言模型中的翻译能力）[11:29] 🧠 On Relation-Specific Neurons in Large Language Models（关于大型语言模型中的关系特定神经元）[12:05] 🔄 Training Consistency Models with Variational Noise Coupling（基于变分噪声耦合的训练一致性模型）[12:46] ⚡ Efficient Gaussian Splatting for Monocular Dynamic Scene Rendering via Sparse Time-Variant Attribute Modeling（通过稀疏时变属性建模实现单目动态场景渲染的高效高斯光栅化）【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

本期的 18 篇论文如下：[00:23] 🌐 Kanana: Compute-efficient Bilingual Language Models（Kanana：计算高效的双语语言模型）[00:54] 👤 GHOST 2.0: generative high-fidelity one shot transfer of heads（GHOST 2.0：生成高保真一次性头部转移）[01:43] 🎥 TheoremExplainAgent: Towards Multimodal Explanations for LLM Theorem Understanding（定理解释代理：面向大语言模型定理理解的多模态解释）[02:21] 🤖 Agentic Reward Modeling: Integrating Human Preferences with Verifiable Correctness Signals for Reliable Reward Systems（代理奖励建模：将人类偏好与可验证的正确性信号结合以构建可靠的奖励系统）[03:02] 🤖 Can Large Language Models Detect Errors in Long Chain-of-Thought Reasoning?（大型语言模型能否检测长链推理中的错误？）[03:47] 🌍 Language Models' Factuality Depends on the Language of Inquiry（语言模型的事实性依赖于查询语言）[04:27] 🧠 Can Language Models Falsify? Evaluating Algorithmic Reasoning with Counterexample Creation（语言模型能否证伪？评估算法推理中的反例创建）[05:11] 🤖 Towards an AI co-scientist（迈向人工智能合作科学家）[05:52] 🇬 Plutus: Benchmarking Large Language Models in Low-Resource Greek Finance（普鲁托斯：在低资源希腊金融环境中评估大型语言模型）[06:38] 🤖 VEM: Environment-Free Exploration for Training GUI Agent with Value Environment Model（VEM：利用价值环境模型训练GUI代理的无环境探索）[07:12] 📏 Distill Any Depth: Distillation Creates a Stronger Monocular Depth Estimator（蒸馏任意深度：蒸馏技术创造更强的单目深度估计器）[07:52] 📚 Project Alexandria: Towards Freeing Scientific Knowledge from Copyright Burdens via LLMs（亚历山大项目：通过大型语言模型解除科学知识的版权负担）[08:35] 🛡 AISafetyLab: A Comprehensive Framework for AI Safety Evaluation and Improvement（AISafetyLab：AI安全评估与改进的综合框架）[09:23] 🧠 BIG-Bench Extra Hard（BIG-Bench 超难版本）[10:07] 🔍 CritiQ: Mining Data Quality Criteria from Human Preferences（CritiQ：从人类偏好中挖掘数据质量标准）[10:44] 🔬 MolSpectra: Pre-training 3D Molecular Representation with Multi-modal Energy Spectra（MolSpectra：利用多模态能量光谱预训练三维分子表示）[11:28] 📄 PosterSum: A Multimodal Benchmark for Scientific Poster Summarization（PosterSum：科学海报摘要的多模态基准）[12:08] 🧠 DOEI: Dual Optimization of Embedding Information for Attention-Enhanced Class Activation Maps（双优化嵌入信息用于增强注意力类激活图）【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

本期的 14 篇论文如下：[00:23] 🤖 OmniAlign-V: Towards Enhanced Alignment of MLLMs with Human Preference（OmniAlign-V：迈向多模态大语言模型与人类偏好增强对齐）[01:06] ⚡ SpargeAttn: Accurate Sparse Attention Accelerating Any Model Inference（SpargeAttn：准确稀疏注意力加速任意模型推理）[01:53] 🖼 KV-Edit: Training-Free Image Editing for Precise Background Preservation（KV-编辑：无需训练的图像编辑方法，实现精确背景保留）[02:32] 🌈 ART: Anonymous Region Transformer for Variable Multi-Layer Transparent Image Generation（匿名区域变换器：可变多层透明图像生成）[03:08] 🤖 SWE-RL: Advancing LLM Reasoning via Reinforcement Learning on Open Software Evolution（SWE-RL：通过开源软件演化数据强化学习提升LLM推理能力）[03:51] 📊 Unveiling Downstream Performance Scaling of LLMs: A Clustering-Based Perspective（揭示大语言模型下游性能扩展：基于聚类的视角）[04:30] 🧠 Scale-Distribution Decoupling: Enabling Stable and Effective Training of Large Language Models（尺度分布解耦：实现大型语言模型稳定有效训练）[05:11] 🔄 K-LoRA: Unlocking Training-Free Fusion of Any Subject and Style LoRAs（K-LoRA：解锁无需训练的任意主题和风格LoRA融合）[05:51] 🌐 WebGames: Challenging General-Purpose Web-Browsing AI Agents（WebGames：挑战通用网页浏览AI代理）[06:29] 🧠 Introducing Visual Perception Token into Multimodal Large Language Model（引入视觉感知令牌的多模态大语言模型）[07:07] 🎰 The Lottery LLM Hypothesis, Rethinking What Abilities Should LLM Compression Preserve?（彩票LLM假说：重新思考LLM压缩应保留的能力）[07:47] 🧠 AAD-LLM: Neural Attention-Driven Auditory Scene Understanding（AAD-LLM：神经注意力驱动的听觉场景理解）[08:26] 🔍 LaTIM: Measuring Latent Token-to-Token Interactions in Mamba Models（LaTIM：测量Mamba模型中的潜在Token-to-Token交互）[09:07] 🧠 Shakti-VLMs: Scalable Vision-Language Models for Enterprise AI（Shakti-VLMs：企业级AI的可扩展视觉语言模型）【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

本期的 20 篇论文如下：[00:27] 📖 Thus Spake Long-Context Large Language Model（长上下文大语言模型如是说）[01:09] 🌈 DICEPTION: A Generalist Diffusion Model for Visual Perceptual Tasks（用于视觉感知任务的通用扩散模型）[01:48] 🚀 Slamming: Training a Speech Language Model on One GPU in a Day（撞击：在一天内使用单个GPU训练语音语言模型）[02:32] 🎥 VideoGrain: Modulating Space-Time Attention for Multi-grained Video Editing（视频粒度：调节时空注意力实现多粒度视频编辑）[03:11] 🎧 Audio-FLAN: A Preliminary Release（音频FLAN：初步发布）[03:43] 🧠 CodeCriticBench: A Holistic Code Critique Benchmark for Large Language Models（CodeCriticBench：面向大型语言模型的全面代码 critique 基准测试）[04:28] 🎨 GCC: Generative Color Constancy via Diffusing a Color Checker（GCC：通过扩散色卡生成颜色恒常性）[05:11] 📊 Linguistic Generalizability of Test-Time Scaling in Mathematical Reasoning（数学推理中测试时间扩展的语言通用性）[05:57] 🚀 Make LoRA Great Again: Boosting LoRA with Adaptive Singular Values and Mixture-of-Experts Optimization Alignment（让LoRA再次伟大：通过自适应奇异值和混合专家优化对齐提升LoRA性能）[06:38] 🧠 Multimodal Inconsistency Reasoning (MMIR): A New Benchmark for Multimodal Reasoning Models（多模态不一致性推理（MMIR）：多模态推理模型的新基准）[07:23] 🎥 RIFLEx: A Free Lunch for Length Extrapolation in Video Diffusion Transformers（RIFLEx：视频扩散Transformer中长度外推的免费午餐）[08:01] 📱 Mobile-Agent-V: Learning Mobile Device Operation Through Video-Guided Multi-Agent Collaboration（移动代理V：通过视频引导的多代理协作学习移动设备操作）[08:45] ⏳ Benchmarking Temporal Reasoning and Alignment Across Chinese Dynasties（中国朝代间的时间推理与对齐基准测试）[09:31] 🤖 Reflective Planning: Vision-Language Models for Multi-Stage Long-Horizon Robotic Manipulation（反射性规划：视觉语言模型在多阶段长时程机器人操作中的应用）[10:02] 🔄 Stable-SPAM: How to Train in 4-Bit More Stably than 16-Bit Adam（稳定-SPAM：如何在4位精度下比16位Adam更稳定地训练）[10:43] 📝 Can Community Notes Replace Professional Fact-Checkers?（社区笔记能替代专业事实核查员吗？）[11:24] 📈 Forecasting Open-Weight AI Model Growth on Hugging Face（预测Hugging Face上开放权重AI模型的增长）[12:08] 🔑 Beyond Release: Access Considerations for Generative AI Systems（超越发布：生成式人工智能系统的访问考量）[12:49] 🌐 TAG: A Decentralized Framework for Multi-Agent Hierarchical Reinforcement Learning（TAG：一种用于多智能体分层强化学习的去中心化框架）[13:30] 💃 X-Dancer: Expressive Music to Human Dance Video Generation（X-Dancer：从音乐生成生动舞蹈视频）【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

本期的 20 篇论文如下：[00:23] 📚 SurveyX: Academic Survey Automation via Large Language Models（基于大型语言模型的学术调查自动化）[01:10] 🔍 LLM-Microscope: Uncovering the Hidden Role of Punctuation in Context Memory of Transformers（LLM显微镜：揭示标点符号在Transformer上下文记忆中的隐藏作用）[01:50] 🚗 MaskGWM: A Generalizable Driving World Model with Video Mask Reconstruction（MaskGWM：结合视频掩码重建的通用驾驶世界模型）[02:28] 🧬 Mol-LLaMA: Towards General Understanding of Molecules in Large Molecular Language Model（Mol-LLaMA：面向大分子语言模型的分子通用理解）[03:12] 🎨 PhotoDoodle: Learning Artistic Image Editing from Few-Shot Pairwise Data（PhotoDoodle：从少量成对数据中学习艺术图像编辑）[03:55] 🔗 VLM$^2$-Bench: A Closer Look at How Well VLMs Implicitly Link Explicit Matching Visual Cues（VLM²-Bench：深入探究视觉语言模型在显式匹配视觉线索上的隐式链接能力）[04:42] 📌 SIFT: Grounding LLM Reasoning in Contexts via Stickers（SIFT：通过贴纸将大语言模型的推理扎根于上下文中）[05:27] 🧠 LightThinker: Thinking Step-by-Step Compression（光思者：逐步压缩推理）[05:59] 🗂 StructFlowBench: A Structured Flow Benchmark for Multi-turn Instruction Following（结构流基准：多轮指令跟随的结构流评估）[06:48] 🛡 Is Safety Standard Same for Everyone? User-Specific Safety Evaluation of Large Language Models（安全标准对所有人都一样吗？大型语言模型的用户特定安全评估）[07:40] 📚 KITAB-Bench: A Comprehensive Multi-Domain Benchmark for Arabic OCR and Document Understanding（KITAB-Bench：阿拉伯语OCR与文档理解的综合多领域基准）[08:30] 🧬 ReQFlow: Rectified Quaternion Flow for Efficient and High-Quality Protein Backbone Generation（ReQFlow：用于高效高质量蛋白质骨架生成的校正四元数流）[09:11] 🧠 MoBA: Mixture of Block Attention for Long-Context LLMs（MoBA：块注意力混合模型用于长上下文LLMs）[09:49] 🤖 InterFeedback: Unveiling Interactive Intelligence of Large Multimodal Models via Human Feedback（InterFeedback：通过人类反馈揭示大型多模态模型的交互智能）[10:37] 🧠 The Relationship Between Reasoning and Performance in Large Language Models -- o3 (mini) Thinks Harder, Not Longer（大语言模型中推理与性能的关系——o3（mini）通过更努力而非更长时间进行推理）[11:20] 📚 Evaluating Multimodal Generative AI with Korean Educational Standards（评估多模态生成式人工智能与韩国教育标准）[11:54] ⚠ Superintelligent Agents Pose Catastrophic Risks: Can Scientist AI Offer a Safer Path?（超级智能体带来灾难性风险：科学家AI能否提供更安全的路径？）[12:29] ⚡ One-step Diffusion Models with $f$-Divergence Distribution Matching（基于$f$-散度分布匹配的一步扩散模型）[13:09] 🧠 Think Inside the JSON: Reinforcement Strategy for Strict LLM Schema Adherence（在JSON内部思考：强化策略实现严格LLM模式遵循）[13:52] 🧠 MedHallu: A Comprehensive Benchmark for Detecting Medical Hallucinations in Large Language Models（MedHallu：检测大型语言模型中的医学幻觉的综合基准）【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

本期的 5 篇论文如下：[00:42] TOP1(🔥138) | 🧠 MLGym: A New Framework and Benchmark for Advancing AI Research Agents（MLGym：推进AI研究代理的新框架与基准）[03:30] TOP2(🔥131) | 🌐 Qwen2.5-VL Technical Report（Qwen2.5-VL 技术报告）[06:56] TOP3(🔥130) | ⚡ Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention（原生稀疏注意力：硬件对齐与原生可训练的稀疏注意力）[09:20] TOP4(🔥91) | 🌐 SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features（SigLIP 2：多语言视觉-语言编码器的语义理解、定位与密集特征改进）[11:50] TOP5(🔥86) | 📚 SuperGPQA: Scaling LLM Evaluation across 285 Graduate Disciplines（SuperGPQA：扩展LLM评估至285个研究生学科）【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

本期的 20 篇论文如下：[00:26] 🧠 MLGym: A New Framework and Benchmark for Advancing AI Research Agents（MLGym：推进AI研究代理的新框架与基准）[01:18] 📚 SuperGPQA: Scaling LLM Evaluation across 285 Graduate Disciplines（SuperGPQA：扩展LLM评估至285个研究生学科）[02:04] 🌐 SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features（SigLIP 2：多语言视觉-语言编码器的语义理解、定位与密集特征改进）[02:52] 🧠 How Much Knowledge Can You Pack into a LoRA Adapter without Harming LLM?（在不损害大型语言模型的情况下，LoRA适配器能容纳多少知识？）[03:49] 🚀 S*: Test Time Scaling for Code Generation（S*：代码生成中的测试时间缩放）[04:35] ⏳ Does Time Have Its Place? Temporal Heads: Where Language Models Recall Time-specific Information（时间是否有其位置？时间头：语言模型如何回忆时间特定信息）[05:28] 📄 LongWriter-V: Enabling Ultra-Long and High-Fidelity Generation in Vision-Language Models（LongWriter-V：在视觉-语言模型中实现超长和高保真生成）[06:17] 🧠 Logic-RL: Unleashing LLM Reasoning with Rule-Based Reinforcement Learning（逻辑-RL：通过基于规则的强化学习释放LLM推理能力）[07:13] 🖥 PC-Agent: A Hierarchical Multi-Agent Collaboration Framework for Complex Task Automation on PC（PC-Agent：一种用于复杂任务自动化在PC上的分层多智能体协作框架）[08:07] 🧠 S$^2$R: Teaching LLMs to Self-verify and Self-correct via Reinforcement Learning（S$^2$R：通过强化学习教导大语言模型自我验证与自我修正）[09:01] 🧠 Discovering highly efficient low-weight quantum error-correcting codes with reinforcement learning（利用强化学习发现高效低权重量子纠错码）[09:55] 🎥 Dynamic Concepts Personalization from Single Videos（单视频动态概念个性化）[10:38] 🖼 Scaling Text-Rich Image Understanding via Code-Guided Synthetic Multimodal Data Generation（通过代码引导的合成多模态数据生成扩展文本丰富的图像理解）[11:23] 🌍 NAVIG: Natural Language-guided Analysis with Vision Language Models for Image Geo-localization（NAVIG：基于自然语言引导的视觉语言模型用于图像地理定位分析）[12:13] 🧠 AlphaMaze: Enhancing Large Language Models' Spatial Intelligence via GRPO（AlphaMaze：通过GRPO提升大型语言模型的空间智能）[13:06] 🌍 How Much Do LLMs Hallucinate across Languages? On Multilingual Estimation of LLM Hallucination in the Wild（LLMs在多语言环境下的幻觉现象研究：在野外场景中的多语言幻觉估计）[13:52] 🌍 Geolocation with Real Human Gameplay Data: A Large-Scale Dataset and Human-Like Reasoning Framework（基于真实人类游戏数据的 geolocation：大规模数据集与人类推理框架）[14:55] 🌐 RelaCtrl: Relevance-Guided Efficient Control for Diffusion Transformers（RelaCtrl：引导相关性的高效控制扩散变换器）[15:54] 🧠 Enhancing Cognition and Explainability of Multimodal Foundation Models with Self-Synthesized Data（增强多模态基础模型的认知与可解释性通过自合成数据）[16:41] 🤖 LLM-based User Profile Management for Recommender System（基于大语言模型的推荐系统用户画像管理）【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速递在小宇宙查看该单集文稿

本期的 20 篇论文如下：[00:24] 🌐 Qwen2.5-VL Technical Report（Qwen2.5-VL 技术报告）[01:10] 🚗 RAD: Training an End-to-End Driving Policy via Large-Scale 3DGS-based Reinforcement Learning（RAD：基于大规模3DGS强化学习的端到端驾驶策略训练）[01:50] 🎶 SongGen: A Single Stage Auto-regressive Transformer for Text-to-Song Generation（SongGen：用于文本到歌曲生成的单阶段自回归Transformer）[02:38] 🧠 MoM: Linear Sequence Modeling with Mixture-of-Memories（MoM：结合记忆混合的线性序列建模）[03:15] 🌐 Craw4LLM: Efficient Web Crawling for LLM Pretraining（Craw4LLM：面向LLM预训练的高效网页爬取方法）[04:05] 🧠 LongPO: Long Context Self-Evolution of Large Language Models through Short-to-Long Preference Optimization（LongPO：通过短至长偏好优化实现大型语言模型的长上下文自进化）[04:45] 🤔 Small Models Struggle to Learn from Strong Reasoners（小型模型难以从强推理者中学习）[05:27] ⚙ Autellix: An Efficient Serving Engine for LLM Agents as General Programs（Autellix：一种用于LLM代理作为通用程序的高效服务引擎）[06:08] 🌍 Presumed Cultural Identity: How Names Shape LLM Responses（假定的文化身份：名字如何塑造LLM的回应）[06:53] 🚨 Why Safeguarded Ships Run Aground? Aligned Large Language Models' Safety Mechanisms Tend to Be Anchored in The Template Region（为什么安全保障的船只也会搁浅？对齐的大型语言模型安全机制往往锚定在模板区域）[07:38] 🩺 SearchRAG: Can Search Engines Be Helpful for LLM-based Medical Question Answering?（搜索RAG：搜索引擎能否助力基于LLM的医疗问答？）[08:21] 🧠 Thinking Preference Optimization（思考偏好优化）[08:59] 🧠 Is That Your Final Answer? Test-Time Scaling Improves Selective Question Answering（那是你的最终答案吗？测试时缩放提升选择性问答）[09:40] 🧠 AdaptiveStep: Automatically Dividing Reasoning Step through Model Confidence（自适应步骤：通过模型置信度自动划分推理步骤）[10:21] 🧬 NExT-Mol: 3D Diffusion Meets 1D Language Modeling for 3D Molecule Generation（NExT-Mol：3D扩散与1D语言建模结合的3D分子生成）[11:02] 🧩 ActionPiece: Contextually Tokenizing Action Sequences for Generative Recommendation（ActionPiece：面向生成推荐的上下文感知行为序列标记化）[11:44] 🧠 Train Small, Infer Large: Memory-Efficient LoRA Training for Large Language Models（小模型训练，大模型推理：用于大型语言模型的内存高效LoRA训练）[12:33] 🌍 GIMMICK -- Globally Inclusive Multimodal Multitask Cultural Knowledge Benchmarking（GIMMICK -- 全球包容性多模态多任务文化知识基准测试）[13:19] 🤖 InfiR : Crafting Effective Small Language Models and Multimodal Small Language Models in Reasoning（InfiR：构建高效的小型语言模型和多模态小型语言模型用于推理）[14:06] 🔊 Noise May Contain Transferable Knowledge: Understanding Semi-supervised Heterogeneous Domain Adaptation from an Empirical Perspective（噪声可能包含可转移的知识：从实证角度理解半监督异构域适应）【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

本期的 20 篇论文如下：[00:25] 🎙 Soundwave: Less is More for Speech-Text Alignment in LLMs（声波：减少数据需求，优化语音与文本对齐在LLMs中的应用）[01:05] 🔍 Cramming 1568 Tokens into a Single Vector and Back Again: Exploring the Limits of Embedding Space Capacity（将1568个Token压缩到一个向量并再次解压：探索嵌入空间容量的极限）[01:48] 🌊 Continuous Diffusion Model for Language Modeling（连续扩散模型用于语言建模）[02:30] 🎥 Phantom: Subject-consistent video generation via cross-modal alignment（幻影：通过跨模态对齐实现主体一致性视频生成）[03:12] 🧠 Rethinking Diverse Human Preference Learning through Principal Component Analysis（重新思考通过主成分分析进行多样化人类偏好学习）[04:00] 🤖 SoFar: Language-Grounded Orientation Bridges Spatial Reasoning and Object Manipulation（SoFar：语言引导的方向桥接空间推理与对象操作）[04:36] 🛡 SafeRoute: Adaptive Model Selection for Efficient and Accurate Safety Guardrails in Large Language Models（SafeRoute：大型语言模型中高效且准确的安全防护栏的自适应模型选择）[05:25] 🐍 Multimodal Mamba: Decoder-only Multimodal State Space Model via Quadratic to Linear Distillation（多模态Mamba：通过二次到线性蒸馏的解码器多模态状态空间模型）[06:08] 📚 You Do Not Fully Utilize Transformer's Representation Capacity（你没有充分利用Transformer的表示能力）[06:50] 🤖 Magma: A Foundation Model for Multimodal AI Agents（熔岩：多模态AI代理的基础模型）[07:23] 💹 FLAG-Trader: Fusion LLM-Agent with Gradient-based Reinforcement Learning for Financial Trading（FLAG-Trader：融合LLM与基于梯度的强化学习用于金融交易）[08:08] 📄 RealSyn: An Effective and Scalable Multimodal Interleaved Document Transformation Paradigm（RealSyn：一种有效且可扩展的多模态交错文档转换范式）[08:49] 🧠 PAFT: Prompt-Agnostic Fine-Tuning（PAFT：与提示无关的微调）[09:27] 🛠 OctoTools: An Agentic Framework with Extensible Tools for Complex Reasoning（OctoTools：一个具有扩展工具的复杂推理代理框架）[10:13] 📊 Revisiting the Test-Time Scaling of o1-like Models: Do they Truly Possess Test-Time Scaling Capabilities?（重新审视o1类模型的测试时缩放能力：它们是否真正具备测试时缩放能力？）[11:00] 🔄 MUDDFormer: Breaking Residual Bottlenecks in Transformers via Multiway Dynamic Dense Connections（MUDDFormer：通过多路动态密集连接打破Transformer中的残差瓶颈）[11:37] 🩺 HealthGPT: A Medical Large Vision-Language Model for Unifying Comprehension and Generation via Heterogeneous Knowledge Adaptation（HealthGPT：通过异构知识适应实现医疗大视觉语言模型的统一理解与生成）[12:12] 🧠 HeadInfer: Memory-Efficient LLM Inference by Head-wise Offloading（HeadInfer：通过分头卸载实现高效的LLM推理）[12:51] 🌍 Text2World: Benchmarking Large Language Models for Symbolic World Model Generation（文本到世界：大语言模型符号世界模型生成的基准测试）[13:32] 🧠 Atom of Thoughts for Markov LLM Test-Time Scaling（用于马尔可夫LLM测试时扩展的原子思维）【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

本期的 29 篇论文如下：[00:23] ⚡ Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention（原生稀疏注意力：硬件对齐与原生可训练的稀疏注意力）[01:10] 🤖 Learning Getting-Up Policies for Real-World Humanoid Robots（学习真实世界人形机器人起身策略）[01:55] 🧠 ReLearn: Unlearning via Learning for Large Language Models（ReLearn：通过学习实现大型语言模型的遗忘）[02:35] 💻 SWE-Lancer: Can Frontier LLMs Earn $1 Million from Real-World Freelance Software Engineering?（SWE-Lancer：前沿大语言模型能否从真实世界的自由软件工程中赚取100万美元？）[03:21] 🌐 HermesFlow: Seamlessly Closing the Gap in Multimodal Understanding and Generation（赫尔墨斯流：无缝衔接多模态理解和生成）[03:58] 🧠 How Do LLMs Acquire New Knowledge? A Knowledge Circuits Perspective on Continual Pre-Training（大型语言模型如何获取新知识？知识电路视角下的持续预训练）[04:33] 🤖 SURGE: On the Potential of Large Language Models as General-Purpose Surrogate Code Executors（SURGE：关于大型语言模型作为通用代理代码执行器的潜力）[05:12] 🔧 Diffusion-Sharpening: Fine-tuning Diffusion Models with Denoising Trajectory Sharpening（扩散锐化：利用去噪轨迹锐化优化扩散模型微调）[05:55] 🧠 I Think, Therefore I Diffuse: Enabling Multimodal In-Context Reasoning in Diffusion Models（我思故我扩散：在扩散模型中实现多模态上下文推理）[06:38] 🔧 SAFE-SQL: Self-Augmented In-Context Learning with Fine-grained Example Selection for Text-to-SQL（SAFE-SQL：基于细粒度示例选择的自增强上下文学习用于文本到SQL转换）[07:25] 🧠 CRANE: Reasoning with constrained LLM generation（CRANE：受限LLM生成的推理）[08:07] 🧠 Intuitive physics understanding emerges from self-supervised pretraining on natural videos（直觉物理理解从自然视频的自监督预训练中涌现）[08:46] 🐦 Cuckoo: An IE Free Rider Hatched by Massive Nutrition in LLM's Nest（杜鹃：在大型语言模型的巢中孵化出的信息抽取搭便车者）[09:22] 🧠 Dyve: Thinking Fast and Slow for Dynamic Process Verification（Dyve：动态过程验证中的快思与慢想）[10:06] 🧠 PhysReason: A Comprehensive Benchmark towards Physics-Based Reasoning（物理推理：基于物理推理的综合基准）[10:53] 🤖 System Message Generation for User Preferences using Open-Source Models（基于开源模型的用户偏好系统消息生成）[11:38] 🎥 video-SALMONN-o1: Reasoning-enhanced Audio-visual Large Language Model（视频-SALMONN-o1：推理增强的音视频大型语言模型）[12:33] 🧠 Building A Proof-Oriented Programmer That Is 64% Better Than GPT-4o Under Data Scarsity（构建一个在数据稀缺情况下比GPT-4o好64%的证明导向程序员）[13:11] 🤖 Memory, Benchmark & Robots: A Benchmark for Solving Complex Tasks with Reinforcement Learning（记忆、基准与机器人：一种用于强化学习解决复杂任务的基准）[13:52] 🤖 MagicArticulate: Make Your 3D Models Articulation-Ready（魔法清晰：让你的3D模型准备好关节动画）[14:37] 🤖 Talk Structurally, Act Hierarchically: A Collaborative Framework for LLM Multi-Agent Systems（结构化交流，层次化行动：LLM多智能体系统的协作框架）[15:21] 🧠 One Example Shown, Many Concepts Known! Counterexample-Driven Conceptual Reasoning in Mathematical LLMs（一个示例展示，多个概念知晓！数学大语言模型中的反例驱动概念推理）[16:03] 🤖 Can a Single Model Master Both Multi-turn Conversations and Tool Use? CALM: A Unified Conversational Agentic Language Model（单一模型能否同时掌握多轮对话与工具使用？CALM：一个统一的对话代理语言模型）[16:40] 🚀 Better Embeddings with Coupled Adam（结合Adam优化器的更好嵌入）[17:18] 🧐 Show Me the Work: Fact-Checkers' Requirements for Explainable Automated Fact-Checking（展示工作：事实核查员对可解释自动化事实核查的需求）[17:56] 🧪 Towards Data-Efficient Pretraining for Atomic Property Prediction（面向原子性质预测的数据高效预训练）[18:46] 🌀 The Mirage of Model Editing: Revisiting Evaluation in the Wild（模型编辑的幻象：重新审视实际应用中的评估）[19:31] 🧮 Large Language Models and Mathematical Reasoning Failures（大型语言模型与数学推理失败）[20:11] 📊 Language Complexity Measurement as a Noisy Zero-Shot Proxy for Evaluating LLM Performance（语言复杂度测量作为评估LLM性能的噪声零样本代理）【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速递在小宇宙查看该单集文稿