2025.07.08 | MemOS提升内存管理效率;MLM与CLM结合优化编码器训练。

HuggingFace 每日AI论文速递

本期的 15 篇论文如下: [00:21] 🧠 MemOS: A Memory OS for AI System(MemOS:面向人工智能系统的内存操作系统) [01:07] 🤔 Should We Still Pretrain Encoders with Masked Language Modeling?(我们是否还应该使用掩码语言模型预训练编码器?) [01:43] 🎥 4DSloMo: 4D Reconstruction for High Speed Scene with Asynchronous Capture(4DSloMo:基于异步捕获的高速场景4D重建) [02:22] 🤖 DreamVLA: A Vision-Language-Action Model Dreamed with Comprehensive World Knowledge(DreamVLA:一个基于综合世界知识构想的视觉-语言-动作模型) [03:02] 🤖 Pre-Trained Policy Discriminators are General Reward Models(预训练策略判别器是通用奖励模型) [03:38] 🧠 BMMR: A Large-Scale Bilingual Multimodal Multi-Discipline Reasoning Dataset(BMMR:一个大规模双语多模态多学科推理数据集) [04:23] 🤖 RoboBrain 2.0 Technical Report(RoboBrain 2.0 技术报告) [05:04] 🧩 Easy Dataset: A Unified and Extensible Framework for Synthesizing LLM Fine-Tuning Data from Unstructured Documents(Easy Dataset:一个从非结构化文档中合成LLM微调数据的统一且可扩展的框架) [05:42] ✨ RefineX: Learning to Refine Pre-training Data at Scale from Expert-Guided Programs(RefineX:通过专家指导的程序学习大规模优化预训练数据) [06:21] 🎬 StreamDiT: Real-Time Streaming Text-to-Video Generation(StreamDiT:实时流式文本到视频生成) [07:04] 📜 Reviving Cultural Heritage: A Novel Approach for Comprehensive Historical Document Restoration(复兴文化遗产:一种全面的历史文献修复新方法) [07:49] 💡 OmniDraft: A Cross-vocabulary, Online Adaptive Drafter for On-device Speculative Decoding(OmniDraft:一种用于端侧推测解码的跨词汇、在线自适应 Drafter) [08:35] 🎨 ArtifactsBench: Bridging the Visual-Interactive Gap in LLM Code Generation Evaluation(ArtifactsBench:弥合LLM代码生成评估中的视觉交互鸿沟) [09:16] 📊 On the rankability of visual embeddings(论视觉嵌入的可排序性) [09:59] 🖼 VLM2Vec-V2: Advancing Multimodal Embedding for Videos, Images, and Visual Documents(VLM2Vec-V2:推进视频、图像和视觉文档的多模态嵌入) 【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

11分钟
99+
3个月前

【月末特辑】6月最火AI论文 | LLM通过自我反思提升性能;MiniMax-M1高效扩展测试计算。

HuggingFace 每日AI论文速递

本期的 10 篇论文如下: [00:37] TOP1(🔥258) | 💡 Reflect, Retry, Reward: Self-Improving LLMs via Reinforcement Learning(反思、重试、奖励:通过强化学习实现LLM的自我提升) [02:51] TOP2(🔥249) | 💡 MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention(MiniMax-M1:利用闪电注意力高效扩展测试时计算) [05:24] TOP3(🔥240) | 🤖 Reinforcement Pre-Training(强化预训练) [07:54] TOP4(🔥165) | 🧠 Beyond the 80/20 Rule: High-Entropy Minority Tokens Drive Effective Reinforcement Learning for LLM Reasoning(超越80/20法则:高熵少数Token驱动LLM推理的有效强化学习) [09:53] TOP5(🔥134) | 🕰 Will It Still Be True Tomorrow? Multilingual Evergreen Question Classification to Improve Trustworthy QA(明日依旧为真吗?多语种常青问题分类以提升可信赖的问答系统) [12:24] TOP6(🔥132) | 🧠 ProRL: Prolonged Reinforcement Learning Expands Reasoning Boundaries in Large Language Models(ProRL:延长的强化学习拓展大型语言模型的推理边界) [14:50] TOP7(🔥126) | 🧠 Confidence Is All You Need: Few-Shot RL Fine-Tuning of Language Models(自信即全部:基于语言模型的小样本强化学习微调) [16:36] TOP8(🔥116) | 🧲 Drag-and-Drop LLMs: Zero-Shot Prompt-to-Weights(拖拽式大语言模型:零样本提示到权重) [18:34] TOP9(🔥108) | 🤖 SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics(SmolVLA:一种用于经济高效型机器人的视觉-语言-动作模型) [21:05] TOP10(🔥107) | 🩺 Lingshu: A Generalist Foundation Model for Unified Multimodal Medical Understanding and Reasoning(灵枢:用于统一多模态医学理解与推理的通用基础模型) 【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

24分钟
99+
4个月前

2025.07.04 | WebSailor提升LLM推理能力;LangScene-X优化3D场景重建。

HuggingFace 每日AI论文速递

本期的 15 篇论文如下: [00:22] 🧭 WebSailor: Navigating Super-human Reasoning for Web Agent(WebSailor:为Web Agent导航超人推理) [00:59] 🖼 LangScene-X: Reconstruct Generalizable 3D Language-Embedded Scenes with TriMap Video Diffusion(LangScene-X:通过TriMap视频扩散重建可泛化的3D语言嵌入场景) [01:44] 🧬 IntFold: A Controllable Foundation Model for General and Specialized Biomolecular Structure Prediction(IntFold:用于通用和专用生物分子结构预测的可控基础模型) [02:35] 👂 Heeding the Inner Voice: Aligning ControlNet Training via Intermediate Features Feedback(倾听内心的声音:通过中间特征反馈对齐ControlNet训练) [03:17] 🤝 Skywork-Reward-V2: Scaling Preference Data Curation via Human-AI Synergy(Skywork-Reward-V2:通过人机协同扩展偏好数据标注) [04:00] 🖼 Thinking with Images for Multimodal Reasoning: Foundations, Methods, and Future Frontiers(基于图像的多模态推理:基础、方法与未来前沿) [04:38] 🧠 Bourbaki: Self-Generated and Goal-Conditioned MDPs for Theorem Proving(布尔巴基:用于定理证明的自生成和目标条件MDP) [05:12] 🧠 Decoupled Planning and Execution: A Hierarchical Reasoning Framework for Deep Search(解耦规划与执行:一种用于深度搜索的分层推理框架) [05:47] 💡 Fast and Simplex: 2-Simplicial Attention in Triton(快速且简明:Triton中的2-单形注意力机制) [06:33] 🧐 Can LLMs Identify Critical Limitations within Scientific Research? A Systematic Evaluation on AI Research Papers(大型语言模型能否识别科学研究中的关键局限性?人工智能研究论文的系统性评估) [07:16] 🧩 Selecting and Merging: Towards Adaptable and Scalable Named Entity Recognition with Large Language Models(选择与合并:面向具有大型语言模型的可适应和可扩展的命名实体识别) [08:12] 🤖 Self-Correction Bench: Revealing and Addressing the Self-Correction Blind Spot in LLMs(自校正基准:揭示并解决大型语言模型中的自校正盲点) [08:51] 💡 Energy-Based Transformers are Scalable Learners and Thinkers(基于能量的Transformer是可扩展的学习者和思考者) [09:33] ⚙ AsyncFlow: An Asynchronous Streaming RL Framework for Efficient LLM Post-Training(AsyncFlow:用于高效大语言模型后训练的异步流式强化学习框架) [10:16] 🚀 ZeCO: Zero Communication Overhead Sequence Parallelism for Linear Attention(ZeCO:线性注意力机制的零通信开销序列并行) 【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

11分钟
73
4个月前

2025.07.03 | 多模态模型提升短视频理解;动画生成保持颜色一致。

HuggingFace 每日AI论文速递

本期的 9 篇论文如下: [00:21] 🎬 Kwai Keye-VL Technical Report(Kwai Keye-VL 技术报告) [01:02] 🎨 LongAnimation: Long Animation Generation with Dynamic Global-Local Memory(LongAnimation:基于动态全局-局部记忆的长期动画生成) [01:50] 👁 Depth Anything at Any Condition(任意条件下的深度感知) [02:28] 🤖 A Survey on Vision-Language-Action Models: An Action Tokenization Perspective(视觉-语言-动作模型综述:一种动作Token化的视角) [03:11] 🪄 FreeMorph: Tuning-Free Generalized Image Morphing with Diffusion Model(FreeMorph:基于扩散模型的免调参通用图像渐变) [03:51] 🖼 Locality-aware Parallel Decoding for Efficient Autoregressive Image Generation(面向高效自回归图像生成的局部感知并行解码) [04:33] 🎬 STR-Match: Matching SpatioTemporal Relevance Score for Training-Free Video Editing(STR-Match: 匹配时空相关性得分的免训练视频编辑方法) [05:14] 📊 MARVIS: Modality Adaptive Reasoning over VISualizations(MARVIS:基于可视化的模态自适应推理) [05:51] 🗣 JAM-Flow: Joint Audio-Motion Synthesis with Flow Matching(JAM-Flow:基于流匹配的联合音频-运动合成) 【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

6分钟
99+
4个月前

2025.07.02 | 多模态推理提升;双向嵌入优化

HuggingFace 每日AI论文速递

本期的 12 篇论文如下: [00:23] 💡 GLM-4.1V-Thinking: Towards Versatile Multimodal Reasoning with Scalable Reinforcement Learning(GLM-4.1V-Thinking:基于可扩展强化学习的通用多模态推理) [01:00] 🖼 MoCa: Modality-aware Continual Pre-training Makes Better Bidirectional Multimodal Embeddings(MoCa:模态感知持续预训练提升双向多模态嵌入效果) [01:35] 🔬 SciArena: An Open Evaluation Platform for Foundation Models in Scientific Literature Tasks(SciArena:科学文献任务中基础模型的开放评估平台) [02:19] 🤔 Does Math Reasoning Improve General LLM Capabilities? Understanding Transferability of LLM Reasoning(数学推理能力是否能提升通用大语言模型的能力?理解大语言模型推理的迁移性) [02:59] 🎬 Radial Attention: $O(n\log n)$ Sparse Attention with Energy Decay for Long Video Generation(径向注意力:用于长视频生成的具有能量衰减的O(n log n)稀疏注意力机制) [03:37] 🤖 DiffuCoder: Understanding and Improving Masked Diffusion Models for Code Generation(DiffuCoder:理解并改进用于代码生成的掩码扩散模型) [04:19] 🧠 HumanOmniV2: From Understanding to Omni-Modal Reasoning with Context(HumanOmniV2:基于上下文理解到全模态推理) [04:53] 🧠 Thinking Beyond Tokens: From Brain-Inspired Intelligence to Cognitive Foundations for Artificial General Intelligence and its Societal Impact(超越Token:从脑启发智能到通用人工智能的认知基础及其社会影响) [05:30] 💡 Data Efficacy for Language Model Training(语言模型训练中的数据效能) [06:05] 🎬 FreeLong++: Training-Free Long Video Generation via Multi-band SpectralFusion(FreeLong++:通过多频段频谱融合实现免训练长视频生成) [06:40] 🖼 IR3D-Bench: Evaluating Vision-Language Model Scene Understanding as Agentic Inverse Rendering(IR3D-Bench:评估视觉-语言模型作为智能体进行逆向渲染的场景理解能力) [07:28] 🛡 Peccavi: Visual Paraphrase Attack Safe and Distortion Free Image Watermarking Technique for AI-Generated Images(Peccavi:一种针对AI生成图像的视觉释义攻击安全且无失真的图像水印技术) 【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

8分钟
99+
4个月前

2025.07.01 | 多模态生成领先;视频扩散效率提升

HuggingFace 每日AI论文速递

本期的 15 篇论文如下: [00:21] 🖼 Ovis-U1 Technical Report(Ovis-U1 技术报告) [00:58] 🎬 VMoBA: Mixture-of-Block Attention for Video Diffusion Models(VMoBA:用于视频扩散模型的混合块注意力机制) [01:36] ✍ Calligrapher: Freestyle Text Image Customization(书法家:自由风格的文本图像定制) [02:21] 🖼 Listener-Rewarded Thinking in VLMs for Image Preferences(图像偏好:视觉语言模型中基于监听者奖励的思考) [03:04] 🧠 SPIRAL: Self-Play on Zero-Sum Games Incentivizes Reasoning via Multi-Agent Multi-Turn Reinforcement Learning(SPIRAL:基于零和博弈的自博弈通过多智能体多轮强化学习激励推理) [03:46] 📸 Consistent Time-of-Flight Depth Denoising via Graph-Informed Geometric Attention(基于图结构几何注意力机制的稳定ToF深度图像去噪) [04:29] 🧬 Evolving Prompts In-Context: An Open-ended, Self-replicating Perspective(上下文演化提示:一种开放式、自复制的视角) [05:09] 🤔 Aha Moment Revisited: Are VLMs Truly Capable of Self Verification in Inference-time Scaling?(“顿悟时刻”再探:视觉语言模型能否在推理时扩展中实现真正的自我验证?) [05:58] 💾 MEMFOF: High-Resolution Training for Memory-Efficient Multi-Frame Optical Flow Estimation(MEMFOF:面向内存高效多帧光流估计的高分辨率训练) [06:38] 🚀 SparseLoRA: Accelerating LLM Fine-Tuning with Contextual Sparsity(SparseLoRA:通过上下文稀疏性加速LLM微调) [07:23] 🏙 UrbanLLaVA: A Multi-modal Large Language Model for Urban Intelligence with Spatial Reasoning and Understanding(UrbanLLaVA:一个用于城市智能的、具备空间推理与理解能力的多模态大型语言模型) [08:01] 🧠 MARBLE: A Hard Benchmark for Multimodal Spatial Reasoning and Planning(MARBLE:一个用于多模态空间推理与规划的硬基准) [08:38] 🧰 Teaching a Language Model to Speak the Language of Tools(教语言模型说工具的语言) [09:16] ✂ VOCABTRIM: Vocabulary Pruning for Efficient Speculative Decoding in LLMs(VOCABTRIM:用于LLM高效推测解码的词汇表剪枝) [10:01] 🤖 RoboScape: Physics-informed Embodied World Model(RoboScape:物理信息驱动的具身世界模型) 【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

11分钟
99+
4个月前

2025.06.30 | 3D视觉编辑;视频令牌压缩

HuggingFace 每日AI论文速递

本期的 14 篇论文如下: [00:26] 🎨 BlenderFusion: 3D-Grounded Visual Editing and Generative Compositing(BlenderFusion:基于3D的视觉编辑和生成式合成) [00:59] ✂ LLaVA-Scissor: Token Compression with Semantic Connected Components for Video LLMs(LLaVA-Scissor:基于语义连通分量的视频LLM令牌压缩) [01:42] 🖼 XVerse: Consistent Multi-Subject Control of Identity and Semantic Attributes via DiT Modulation(XVerse:通过DiT调制实现对身份和语义属性的多主体一致性控制) [02:24] 🎬 ShotBench: Expert-Level Cinematic Understanding in Vision-Language Models(ShotBench:视觉-语言模型中专家级电影理解) [03:05] 🖼 From Ideal to Real: Unified and Data-Efficient Dense Prediction for Real-World Scenarios(从理想到现实:面向真实场景的统一且数据高效的密集预测) [03:44] 🖼 MiCo: Multi-image Contrast for Reinforcement Visual Reasoning(MiCo:用于增强视觉推理的多图像对比学习) [04:24] 🧮 Pangu Pro MoE: Mixture of Grouped Experts for Efficient Sparsity(Pangu Pro MoE:用于高效稀疏性的分组专家混合模型) [05:06] 🗺 Fine-Grained Preference Optimization Improves Spatial Reasoning in VLMs(细粒度偏好优化提升视觉语言模型中的空间推理能力) [05:52] 🤖 Ark: An Open-source Python-based Framework for Robot Learning(Ark:一个用于机器人学习的开源Python框架) [06:36] 🎨 Noise Consistency Training: A Native Approach for One-Step Generator in Learning Additional Controls(噪声一致性训练:一种在学习额外控制时用于单步生成器的原生方法) [07:20] 🏎 The Automated LLM Speedrunning Benchmark: Reproducing NanoGPT Improvements(自动化LLM竞速基准:复现NanoGPT的改进) [08:01] 🧠 Gazal-R1: Achieving State-of-the-Art Medical Reasoning with Parameter-Efficient Two-Stage Training(Gazal-R1:通过参数高效的两阶段训练实现最先进的医学推理) [08:45] 🧮 Confucius3-Math: A Lightweight High-Performance Reasoning LLM for Chinese K-12 Mathematics Learning(Confucius3-Math:一个用于中国K-12数学学习的轻量级高性能推理大语言模型) [09:39] 👁 RetFiner: A Vision-Language Refinement Scheme for Retinal Foundation Models(RetFiner:用于视网膜基础模型的视觉-语言精炼方案) 【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

10分钟
91
4个月前

2025.06.27 | 强化学习提升搜索效率;记忆增强生成逼真驾驶场景。

HuggingFace 每日AI论文速递

本期的 15 篇论文如下: [00:25] 🔍 MMSearch-R1: Incentivizing LMMs to Search(MMSearch-R1:激励大型多模态模型进行搜索) [00:59] 🚗 MADrive: Memory-Augmented Driving Scene Modeling(MADrive:基于记忆增强的驾驶场景建模) [01:43] 🤖 WorldVLA: Towards Autoregressive Action World Model(WorldVLA:面向自回归动作世界模型) [02:23] 💡 Where to find Grokking in LLM Pretraining? Monitor Memorization-to-Generalization without Test(大型语言模型预训练中Grokking现象 কোথায়? 无需测试,监测从记忆到泛化的过程) [03:14] 🤖 Mind2Web 2: Evaluating Agentic Search with Agent-as-a-Judge(Mind2Web 2:使用Agent-as-a-Judge评估自主搜索) [04:00] 🚗 SAM4D: Segment Anything in Camera and LiDAR Streams(SAM4D:相机和激光雷达流中的可分割一切) [04:40] 🎨 FaSTA$^*$: Fast-Slow Toolpath Agent with Subroutine Mining for Efficient Multi-turn Image Editing(FaSTA$^*$: 快速-慢速工具路径智能体,通过子程序挖掘实现高效的多轮图像编辑) [05:16] 🤖 Whole-Body Conditioned Egocentric Video Prediction(全身条件下的自我中心视频预测) [05:53] 🧠 Arch-Router: Aligning LLM Routing with Human Preferences(Arch-Router:将LLM路由与人类偏好对齐) [06:35] 🎨 FairyGen: Storied Cartoon Video from a Single Child-Drawn Character(FairyGen:从单张儿童绘画生成故事驱动的卡通视频) [07:12] 🌐 DiLoCoX: A Low-Communication Large-Scale Training Framework for Decentralized Cluster(DiLoCoX:一种用于去中心化集群的低通信大规模训练框架) [07:55] 🧬 An Agentic System for Rare Disease Diagnosis with Traceable Reasoning(基于Agent的罕见病诊断系统,具有可追溯的推理能力) [08:35] 🤖 HeurAgenix: Leveraging LLMs for Solving Complex Combinatorial Optimization Challenges(HeurAgenix:利用大型语言模型解决复杂组合优化难题) [09:18] 🦘 Learning to Skip the Middle Layers of Transformers(学习跳过Transformer的中间层) [09:57] 🎵 MuseControlLite: Multifunctional Music Generation with Lightweight Conditioners(MuseControlLite:基于轻量级调节器的多功能音乐生成) 【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

10分钟
99+
4个月前

2025.06.26 | 高质量多模态模型;4比特量化提升性能

HuggingFace 每日AI论文速递

本期的 14 篇论文如下: [00:23] 🖼 ShareGPT-4o-Image: Aligning Multimodal Models with GPT-4o-Level Image Generation(ShareGPT-4o-Image:通过GPT-4o级别的图像生成能力对齐多模态模型) [01:05] 🛡 Outlier-Safe Pre-Training for Robust 4-Bit Quantization of Large Language Models(面向稳健4比特量化的异常值安全预训练大语言模型) [01:49] 🎨 Inverse-and-Edit: Effective and Fast Image Editing by Cycle Consistency Models(逆向与编辑:基于循环一致性模型的高效快速图像编辑) [02:30] 🧠 OctoThinker: Mid-training Incentivizes Reinforcement Learning Scaling(OctoThinker:中期训练激励强化学习扩展) [03:13] 🤖 DualTHOR: A Dual-Arm Humanoid Simulation Platform for Contingency-Aware Planning(DualTHOR:一个用于情境感知规划的双臂人形机器人仿真平台) [03:49] 🦾 RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain Randomization for Robust Bimanual Robotic Manipulation(RoboTwin 2.0:一种可扩展的数据生成器和基准,具有强大的领域随机化,用于鲁棒的双臂机器人操作) [04:33] 🧪 Use Property-Based Testing to Bridge LLM Code Generation and Validation(利用基于属性的测试弥合LLM代码生成与验证之间的差距) [05:18] 🌍 When Life Gives You Samples: The Benefits of Scaling up Inference Compute for Multilingual LLMs(当生活给你样本时:扩展多语言LLM的推理计算的益处) [05:56] 🖼 HiWave: Training-Free High-Resolution Image Generation via Wavelet-Based Diffusion Sampling(HiWave:基于小波变换扩散采样的免训练高分辨率图像生成) [06:39] 🤖 ReCode: Updating Code API Knowledge with Reinforcement Learning(ReCode:利用强化学习更新代码API知识) [07:15] 💬 Is There a Case for Conversation Optimized Tokenizers in Large Language Models?(大型语言模型中,面向对话优化的分词器是否有意义?) [07:59] 🔬 Biomed-Enriched: A Biomedical Dataset Enriched with LLMs for Pretraining and Extracting Rare and Hidden Content(Biomed-Enriched:一个利用大型语言模型富集的生物医学数据集,用于预训练和提取稀有及隐藏内容) [08:47] 🤖 MATE: LLM-Powered Multi-Agent Translation Environment for Accessibility Applications(MATE:基于LLM的多智能体翻译环境,用于辅助应用) [09:28] 📉 The Debugging Decay Index: Rethinking Debugging Strategies for Code LLMs(调试衰减指数:重新思考代码大语言模型的调试策略) 【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

10分钟
99+
4个月前
EarsOnMe

加入我们的 Discord

与播客爱好者一起交流

立即加入

扫描微信二维码

添加微信好友,获取更多播客资讯

微信二维码

播放列表

自动播放下一个

播放列表还是空的

去找些喜欢的节目添加进来吧