{ "updatedAt": "2026-10-16T00:56:21.837Z", "date": "2026-10-16", "count": 60, "items": [ { "arxivId": "2610.20571", "title": "Any Time, Any Space: A World-State-Centric Framework for Agentic Video Generation", "summary": "ATAS improves long-form video consistency by compiling screenplays into referenced shot requests, retrieving spatial and temporal evidence from scene and event memories, and conditionally updating memory after generation.", "authors": [ "Weijie Wang", "Haomin Zhang", "Ye Li", "Akide Liu", "Zeyu Zhang", "Haoyu Zhao" ], "organization": { "_id": "61bac2af530e5c78d7b99667", "name": "zju", "fullname": "Zhejiang University", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/5e1058e9fcf41d740b69966d/7G1xjlxwCdMEmKcxNR0n5.png" }, "publishedAt": "2026-10-14T00:00:00.000Z", "submittedAt": "2026-10-15T00:00:00.000Z", "upvotes": 323, "comments": 1, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.20571.png", "arxivUrl": "https://arxiv.org/abs/2610.20571", "pdfUrl": "https://arxiv.org/pdf/2610.20571.pdf", "hfUrl": "https://huggingface.co/papers/2610.20571", "githubRepo": "https://github.com/ziplab/ATAS", "githubStars": 197, "keywords": [ "agentic video generation", "screenplay", "spatial consistency", "temporal consistency", "scene memory", "event memory" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Any Time, Any Space: A World-State-Centric Framework for Agentic Video Generation" }, "summary_i18n": { "en": "ATAS improves long-form video consistency by compiling screenplays into referenced shot requests, retrieving spatial and temporal evidence from scene and event memories, and conditionally updating memory after generation." } }, { "arxivId": "2610.20584", "title": "Player-1: Learning to Play from Synthetic Experience", "summary": "Player-1 improves game control by combining supervised fine-tuning on gameplay trajectories with reinforcement learning across synthetic environments, achieving substantial gains over demonstration-only baselines.", "authors": [ "Mingyu Ouyang", "Yiran Wang", "Siyuan Hu", "Xingyilang Yin", "Huiqiang Sun", "Guangzhi Wang" ], "organization": { "_id": "60e3f7f641ca131919975fe5", "name": "TencentARC", "fullname": "ARC Lab, Tencent", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/1625552871844-60e272ca6c78a8c122b12127.png" }, "publishedAt": "2026-10-14T00:00:00.000Z", "submittedAt": "2026-10-15T00:00:00.000Z", "upvotes": 270, "comments": 1, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.20584.png", "arxivUrl": "https://arxiv.org/abs/2610.20584", "pdfUrl": "https://arxiv.org/pdf/2610.20584.pdf", "hfUrl": "https://huggingface.co/papers/2610.20584", "githubRepo": "https://github.com/TencentARC/Player-1", "githubStars": 75, "keywords": [ "computer-use agents", "relative pointer motion", "sustained inputs", "composite actions", "supervised fine-tuning", "reinforcement learning" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Player-1: Learning to Play from Synthetic Experience" }, "summary_i18n": { "en": "Player-1 improves game control by combining supervised fine-tuning on gameplay trajectories with reinforcement learning across synthetic environments, achieving substantial gains over demonstration-only baselines." } }, { "arxivId": "2610.13527", "title": "S-Space: Exploring Spatial Workspace in Multimodal Language Models", "summary": "Multimodal language models use an intrinsic linear subspace called S-Space to encode object coordinates, supporting spatial perception, reasoning, and response generation, while also revealing bottlenecks in viewpoint transformation.", "authors": [ "Zhiyuan Shi", "Xiang Li", "Xingzhuo Guo", "Shangchen Miao", "Weiliang Chen", "Qiyu Dai" ], "organization": { "_id": "6a77dec42fad4e89f1ce346a", "name": "MirroS-Lab", "fullname": "MirroS", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/643b866bff50448bcfc7d1d1/pSoXy9fqBTNK3AUuisaH-.png" }, "publishedAt": "2026-10-09T12:48:41.000Z", "submittedAt": "2026-10-12T00:00:00.000Z", "upvotes": 294, "comments": 2, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.13527.png", "arxivUrl": "https://arxiv.org/abs/2610.13527", "pdfUrl": "https://arxiv.org/pdf/2610.13527.pdf", "hfUrl": "https://huggingface.co/papers/2610.13527", "githubRepo": "https://github.com/mirros-lab/s-space", "githubStars": 173, "keywords": [ "multimodal language models", "S-Space", "linear representation subspace", "object-token representations", "spatial reasoning", "causal interventions" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "S-Space: Exploring Spatial Workspace in Multimodal Language Models" }, "summary_i18n": { "en": "Multimodal language models use an intrinsic linear subspace called S-Space to encode object coordinates, supporting spatial perception, reasoning, and response generation, while also revealing bottlenecks in viewpoint transformation." } }, { "arxivId": "2609.39601", "title": "GroundingPI: A Grounding Foundation Model towards Physical Intelligence with Visual Primitives", "summary": "GroundingPI is a 4B grounding foundation model that improves precise object localization and downstream robotic and autonomous driving performance through multimodal pretraining and reinforcement learning.", "authors": [ "Qize Yu", "Lianrui Fan", "Boyu Chen", "Jiaqi Liang", "Xini Ding", "Yue Chen" ], "organization": { "_id": "6aa24a85fc727b9188a95572", "name": "XPENG-AI", "fullname": "XPENG AI", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/6406db5cd684369027166986/DAJtDU_HssW_R7WIw0USp.jpeg" }, "publishedAt": "2026-09-30T00:00:00.000Z", "submittedAt": "2026-10-13T00:00:00.000Z", "upvotes": 314, "comments": 2, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2609.39601.png", "arxivUrl": "https://arxiv.org/abs/2609.39601", "pdfUrl": "https://arxiv.org/pdf/2609.39601.pdf", "hfUrl": "https://huggingface.co/papers/2609.39601", "githubRepo": "https://github.com/groundingpi/GroundingPI", "githubStars": 93, "keywords": [ "vision-language-action", "world-action models", "grounding foundation model", "quantized coordinates", "multimodal pretraining", "supervised fine-tuning" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "GroundingPI: A Grounding Foundation Model towards Physical Intelligence with Visual Primitives" }, "summary_i18n": { "en": "GroundingPI is a 4B grounding foundation model that improves precise object localization and downstream robotic and autonomous driving performance through multimodal pretraining and reinforcement learning." } }, { "arxivId": "2610.18773", "title": "Jeva: System One Models from First Principles", "summary": "Jeva is a decision-native architecture that jointly enforces null-awareness, permutation invariance, option scalability, and cross-request caching for structured decision-making.", "authors": [ "Qi Li", "Xinchao Wang" ], "organization": { "_id": "6508ab2b349930913196378b", "name": "NationalUniversityofSingapore", "fullname": "National University of Singapore", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/630ca0817dacb93b33506ce7/ZYUmpSMsa5Whihw3me2Bw.png" }, "publishedAt": "2026-10-13T00:00:00.000Z", "submittedAt": "2026-10-14T00:00:00.000Z", "upvotes": 276, "comments": 4, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.18773.png", "arxivUrl": "https://arxiv.org/abs/2610.18773", "pdfUrl": "https://arxiv.org/pdf/2610.18773.pdf", "hfUrl": "https://huggingface.co/papers/2610.18773", "githubRepo": "https://github.com/LiQiiiii/Jeva", "githubStars": 76, "keywords": [ "encoder", "decision stack", "symmetric aggregation", "permutation-equivariant", "shared scoring", "coverage estimate" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Jeva: System One Models from First Principles" }, "summary_i18n": { "en": "Jeva is a decision-native architecture that jointly enforces null-awareness, permutation invariance, option scalability, and cross-request caching for structured decision-making." } }, { "arxivId": "2610.13361", "title": "AgentSWE: Can Coding Agents Build the Agent You Actually Want?", "summary": "AgentSWE evaluates whether coding agents build software that truly meets user requirements across creation, editing, and optimization stages, revealing widespread failures to deliver faithful behavior.", "authors": [ "Jiahao Wang", "Hongjin Qian", "Yuyang Hu", "Jiajun Zhang", "Zheng Liu" ], "organization": null, "publishedAt": "2026-10-09T00:00:00.000Z", "submittedAt": "2026-10-13T00:00:00.000Z", "upvotes": 227, "comments": 2, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.13361.png", "arxivUrl": "https://arxiv.org/abs/2610.13361", "pdfUrl": "https://arxiv.org/pdf/2610.13361.pdf", "hfUrl": "https://huggingface.co/papers/2610.13361", "githubRepo": "https://github.com/VectorSpaceLab/AgentSWE", "githubStars": 37, "keywords": [ "AgentSWE", "agent software engineering", "coding agents", "overfitting", "failure trajectories", "requirement-derived rubrics" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "AgentSWE: Can Coding Agents Build the Agent You Actually Want?" }, "summary_i18n": { "en": "AgentSWE evaluates whether coding agents build software that truly meets user requirements across creation, editing, and optimization stages, revealing widespread failures to deliver faithful behavior." } }, { "arxivId": "2610.12572", "title": "DiffusionART: Efficient, Stable and Scalable Reinforcement of Diffusion Models", "summary": "DiffusionART improves reinforcement learning fine-tuning of diffusion models by removing hidden consistency forces and enabling stable, efficient training without EMA tuning or differentiable rewards.", "authors": [ "Zhenglin Wan", "Jiayu Chen", "Zhengfeiyang Zhang", "Pengfei Zhou", "Wangbo Zhao", "Fangyuan Kong" ], "organization": { "_id": "662c559b322afcbae51b3c8b", "name": "KlingTeam", "fullname": "KlingAI Research", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/60e272ca6c78a8c122b12127/ZQV1aKLUDPf2rUcxxAqj6.jpeg" }, "publishedAt": "2026-10-08T00:00:00.000Z", "submittedAt": "2026-10-12T00:00:00.000Z", "upvotes": 148, "comments": 4, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.12572.png", "arxivUrl": "https://arxiv.org/abs/2610.12572", "pdfUrl": "https://arxiv.org/pdf/2610.12572.pdf", "hfUrl": "https://huggingface.co/papers/2610.12572", "githubRepo": "https://github.com/vanzll/DiffusionART", "githubStars": 130, "keywords": [ "online reinforcement learning", "diffusion models", "Flow-GRPO", "DiffusionNFT", "exponential-moving-average", "NFT-style objectives" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "DiffusionART: Efficient, Stable and Scalable Reinforcement of Diffusion Models" }, "summary_i18n": { "en": "DiffusionART improves reinforcement learning fine-tuning of diffusion models by removing hidden consistency forces and enabling stable, efficient training without EMA tuning or differentiable rewards." } }, { "arxivId": "2610.18615", "title": "VideoJev: System One Model for Video Understanding and Interaction", "summary": "VideoJev converts video observations into structured decisions by scoring temporal intervals and text options, enabling efficient reusable inference for visual agents.", "authors": [ "Yuhan Zhu", "Xiangyu Zeng", "Qingyi Si", "Chenxu Yang", "Yifei Li", "Luozheng Qin" ], "organization": { "_id": "62c77fde1e080b83746468bd", "name": "MCG-NJU", "fullname": "Multimedia Computing Group-Nanjing University", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/OIFtHl-mDBM_5SqyR9T8f.png" }, "publishedAt": "2026-10-13T00:00:00.000Z", "submittedAt": "2026-10-14T00:00:00.000Z", "upvotes": 145, "comments": 2, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.18615.png", "arxivUrl": "https://arxiv.org/abs/2610.18615", "pdfUrl": "https://arxiv.org/pdf/2610.18615.pdf", "hfUrl": "https://huggingface.co/papers/2610.18615", "githubRepo": "https://github.com/MCG-NJU/VideoJev", "githubStars": 123, "keywords": [ "multimodal large language models", "temporal grounding", "multiple-choice video question answering", "Cached inference", "Joint inference", "System One" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "VideoJev: System One Model for Video Understanding and Interaction" }, "summary_i18n": { "en": "VideoJev converts video observations into structured decisions by scoring temporal intervals and text options, enabling efficient reusable inference for visual agents." } }, { "arxivId": "2610.11765", "title": "Thinking Inertia: LLMs Keep Thinking When Told Not To", "summary": "Researchers propose multi-level metrics to evaluate LLM no-thinking behavior, finding that explicit reasoning persists despite controls and varies with task openness, establishing no-thinking as a distinct capability requiring systematic assessment.", "authors": [ "Dianqiao Lei", "Kevin Qinghong Lin", "Pan Lu", "Philip Torr", "James Zou" ], "organization": { "_id": "672c672dcf09d152f4da04c4", "name": "StanfordUniversity", "fullname": "Stanford University", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/68e396f2b5bb631e9b2fac9a/vJI0POlzGMXL2878t1vz2.jpeg" }, "publishedAt": "2026-10-08T00:00:00.000Z", "submittedAt": "2026-10-13T00:00:00.000Z", "upvotes": 244, "comments": 6, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.11765.png", "arxivUrl": "https://arxiv.org/abs/2610.11765", "pdfUrl": "https://arxiv.org/pdf/2610.11765.pdf", "hfUrl": "https://huggingface.co/papers/2610.11765", "githubRepo": "https://github.com/thinking-inertia/code", "githubStars": 6, "keywords": [ "no-thinking", "Empty-Thinking Rate", "Question-Pre-answer Relevance", "Explicit Inference Rate", "Thinking Inertia", "prompting interventions" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Thinking Inertia: LLMs Keep Thinking When Told Not To" }, "summary_i18n": { "en": "Researchers propose multi-level metrics to evaluate LLM no-thinking behavior, finding that explicit reasoning persists despite controls and varies with task openness, establishing no-thinking as a distinct capability requiring systematic assessment." } }, { "arxivId": "2610.13521", "title": "Beyond Domain-Routed Supervision: Cross-Teacher Guarding for Multi-Teacher On-Policy Distillation", "summary": "Road MOPD reduces gradient conflicts in multi-teacher distillation by using cross-teacher agreement for token-level credit assignment and auxiliary gradients to constrain conflicting updates.", "authors": [ "Shi-Yu Tian", "Yong Liu", "Ximan Liu", "Tianrui Zhu", "Sule Bai", "Xinwei Yang" ], "organization": { "_id": "682a97a154b087448a5504ee", "name": "jingdong1", "fullname": "jingdong", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/682a966182dc4fc3e373e3ed/9xUgXZCij6qGJpLqpR99M.png" }, "publishedAt": "2026-10-09T00:00:00.000Z", "submittedAt": "2026-10-15T00:00:00.000Z", "upvotes": 233, "comments": 1, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.13521.png", "arxivUrl": "https://arxiv.org/abs/2610.13521", "pdfUrl": "https://arxiv.org/pdf/2610.13521.pdf", "hfUrl": "https://huggingface.co/papers/2610.13521", "githubRepo": null, "githubStars": null, "keywords": [ "multi-teacher on-policy distillation", "on-policy distillation", "token-level supervision", "cross-teacher guarding", "credit assignment", "gradient projection" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Beyond Domain-Routed Supervision: Cross-Teacher Guarding for Multi-Teacher On-Policy Distillation" }, "summary_i18n": { "en": "Road MOPD reduces gradient conflicts in multi-teacher distillation by using cross-teacher agreement for token-level credit assignment and auxiliary gradients to constrain conflicting updates." } }, { "arxivId": "2610.20344", "title": "Doc2Harness: Benchmarking LLMs on Long-Horizon Harness Engineering of Coding Agents", "summary": "Doc2Harness is a deterministic two-stage benchmark that evaluates long-horizon harness engineering by having models build and run harnesses across isolated containers, revealing substantial performance gaps among frontier models.", "authors": [ "Yuyao Ge", "Shenghua Liu", "Peizhong Gao", "Chendi Ge", "Chenxu Liu", "Yiwei Wang" ], "organization": { "_id": "68ef0ab704f0f03d81964936", "name": "ict-cas", "fullname": "Institute of Computing Technology", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/656ad93853703dd78f3de7b8/MR5EsF33Ev3IdnMOHv1be.jpeg" }, "publishedAt": "2026-10-14T00:00:00.000Z", "submittedAt": "2026-10-15T00:00:00.000Z", "upvotes": 93, "comments": 1, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.20344.png", "arxivUrl": "https://arxiv.org/abs/2610.20344", "pdfUrl": "https://arxiv.org/pdf/2610.20344.pdf", "hfUrl": "https://huggingface.co/papers/2610.20344", "githubRepo": "https://github.com/YuyaoGe/doc2harness", "githubStars": 33, "keywords": [ "Doc2Harness", "harness engineering", "long-horizon agents", "build model", "interface-only contract skeleton", "mock model" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Doc2Harness: Benchmarking LLMs on Long-Horizon Harness Engineering of Coding Agents" }, "summary_i18n": { "en": "Doc2Harness is a deterministic two-stage benchmark that evaluates long-horizon harness engineering by having models build and run harnesses across isolated containers, revealing substantial performance gaps among frontier models." } }, { "arxivId": "2610.18356", "title": "Urbex: Agentic Spatial Grounding for City-Scale 3D Scenes", "summary": "Urbex reframes city-scale 3D grounding as an active search process using a vision-language agent with reinforcement learning, improving accuracy and speed without precomputed proposals.", "authors": [ "Yu Wang", "Rui Dai", "Wei Chen", "Yi Wang", "Bo Dang", "Kaikui Liu" ], "organization": { "_id": "68be41370a3fcebdcad6516a", "name": "alibabagroup", "fullname": "alibaba", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/68be3ab7e52df040b2cf80dc/li4G29u_EGswyTN1Sm_Kq.png" }, "publishedAt": "2026-10-13T00:00:00.000Z", "submittedAt": "2026-10-14T00:00:00.000Z", "upvotes": 116, "comments": 1, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.18356.png", "arxivUrl": "https://arxiv.org/abs/2610.18356", "pdfUrl": "https://arxiv.org/pdf/2610.18356.pdf", "hfUrl": "https://huggingface.co/papers/2610.18356", "githubRepo": "https://github.com/AMAP-ML/Urbex", "githubStars": 92, "keywords": [ "vision-language model", "VLM agent", "multi-view environment", "reinforcement learning", "bounding box", "cross-attention" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Urbex: Agentic Spatial Grounding for City-Scale 3D Scenes" }, "summary_i18n": { "en": "Urbex reframes city-scale 3D grounding as an active search process using a vision-language agent with reinforcement learning, improving accuracy and speed without precomputed proposals." } }, { "arxivId": "2610.19789", "title": "More Methods, Better Agents? A Reality Check on Self-Distillation in Agentic Reinforcement Learning", "summary": "Researchers benchmarked ten self-distillation methods against GRPO across multiple agent tasks and model sizes, finding limited consistent gains, hidden training instability, and no reliable cost-to-performance benefit.", "authors": [ "Linhao Zhong", "Bozhen Fang", "Xuehao Wang", "Duochao Shi", "Zekai Luo", "Guangming Yao" ], "organization": { "_id": "61bac2af530e5c78d7b99667", "name": "zju", "fullname": "Zhejiang University", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/5e1058e9fcf41d740b69966d/7G1xjlxwCdMEmKcxNR0n5.png" }, "publishedAt": "2026-10-14T00:00:00.000Z", "submittedAt": "2026-10-15T00:00:00.000Z", "upvotes": 72, "comments": 2, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.19789.png", "arxivUrl": "https://arxiv.org/abs/2610.19789", "pdfUrl": "https://arxiv.org/pdf/2610.19789.pdf", "hfUrl": "https://huggingface.co/papers/2610.19789", "githubRepo": "https://github.com/aim-uofa/Self-Distillation-AgentRL", "githubStars": 32, "keywords": [ "self-distillation", "agentic reinforcement learning", "GRPO", "ALFWorld", "ScienceWorld", "Search-QA" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "More Methods, Better Agents? A Reality Check on Self-Distillation in Agentic Reinforcement Learning" }, "summary_i18n": { "en": "Researchers benchmarked ten self-distillation methods against GRPO across multiple agent tasks and model sizes, finding limited consistent gains, hidden training instability, and no reliable cost-to-performance benefit." } }, { "arxivId": "2610.12818", "title": "LeWAM: End-to-End World and Action Modeling with JEPAs", "summary": "LeWAM is an end-to-end joint-embedding predictive architecture that jointly learns latent dynamics and action flow matching for efficient planning and direct control.", "authors": [ "Minghao Fu", "Tavis Siebert", "Eryk Halicki", "Randall Balestriero" ], "organization": { "_id": "6ac2f6069f39a66f240523a6", "name": "LeWAM", "fullname": "LeWAM", "avatar": "https://www.gravatar.com/avatar/13a0ac54a3ef634d52eeebaed58ebaeb?d=retro&size=100" }, "publishedAt": "2026-10-08T00:00:00.000Z", "submittedAt": "2026-10-14T00:00:00.000Z", "upvotes": 65, "comments": 2, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.12818.png", "arxivUrl": "https://arxiv.org/abs/2610.12818", "pdfUrl": "https://arxiv.org/pdf/2610.12818.pdf", "hfUrl": "https://huggingface.co/papers/2610.12818", "githubRepo": "https://github.com/MinghaoFu/lewam", "githubStars": 240, "keywords": [ "joint-embedding predictive architectures", "JEPA", "latent world models", "action flow matching", "isotropic Gaussian constraint", "latent dynamics" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "LeWAM: End-to-End World and Action Modeling with JEPAs" }, "summary_i18n": { "en": "LeWAM is an end-to-end joint-embedding predictive architecture that jointly learns latent dynamics and action flow matching for efficient planning and direct control." } }, { "arxivId": "2609.39600", "title": "GroundAnything: Reconciling Parallel Decoding with Precise Visual Grounding at Flash Speed", "summary": "GroundAnything is a 4B-parameter visual grounding model that uses bidirectional diffusion and blockwise denoising to enable fast parallel decoding with precise localization, outperforming autoregressive counterparts across 30 benchmarks.", "authors": [ "Qize Yu", "Lianrui Fan", "Bowen Ping", "Xini Ding", "Zetian Song", "Junbo Niu" ], "organization": { "_id": "6aa24a85fc727b9188a95572", "name": "XPENG-AI", "fullname": "XPENG AI", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/6406db5cd684369027166986/DAJtDU_HssW_R7WIw0USp.jpeg" }, "publishedAt": "2026-09-30T00:00:00.000Z", "submittedAt": "2026-10-13T00:00:00.000Z", "upvotes": 94, "comments": 2, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2609.39600.png", "arxivUrl": "https://arxiv.org/abs/2609.39600", "pdfUrl": "https://arxiv.org/pdf/2609.39600.pdf", "hfUrl": "https://huggingface.co/papers/2609.39600", "githubRepo": "https://github.com/groundingpi/GroundAnything", "githubStars": 46, "keywords": [ "autoregressive grounding", "bidirectional diffusion", "blockwise denoising", "diffusion models", "visual grounding", "entropy-guided decoding" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "GroundAnything: Reconciling Parallel Decoding with Precise Visual Grounding at Flash Speed" }, "summary_i18n": { "en": "GroundAnything is a 4B-parameter visual grounding model that uses bidirectional diffusion and blockwise denoising to enable fast parallel decoding with precise localization, outperforming autoregressive counterparts across 30 benchmarks." } }, { "arxivId": "2610.13912", "title": "RSI-Anything: Unlocking Recursive Self-Improvement with World Models", "summary": "RSI-Anything leverages predictive world models to guide action selection and adaptive learning, enabling recursive self-improvement across coding, interface navigation, laboratory planning, and scientific optimization tasks.", "authors": [ "Taoyong Cui", "Pinxian Zeng", "Wanghan Xu", "Hao Xiao", "Shuhan Xue", "Qiang Gao" ], "organization": { "_id": "6aab7c097e5721f5e005ff0a", "name": "PhAILabs", "fullname": "PhAI Labs", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/64fde4e252e82dd432b74ce9/5-hyYwAMko5JVJz5qWxEj.png" }, "publishedAt": "2026-10-09T00:00:00.000Z", "submittedAt": "2026-10-12T00:00:00.000Z", "upvotes": 46, "comments": 2, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.13912.png", "arxivUrl": "https://arxiv.org/abs/2610.13912", "pdfUrl": "https://arxiv.org/pdf/2610.13912.pdf", "hfUrl": "https://huggingface.co/papers/2610.13912", "githubRepo": "https://github.com/Gen-Verse/RSI-Anything", "githubStars": 751, "keywords": [ "recursive self-improvement", "predictive world models", "JEPA-Anything", "conditional world-model adaptation", "policy-model pairs", "code repair" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "RSI-Anything: Unlocking Recursive Self-Improvement with World Models" }, "summary_i18n": { "en": "RSI-Anything leverages predictive world models to guide action selection and adaptive learning, enabling recursive self-improvement across coding, interface navigation, laboratory planning, and scientific optimization tasks." } }, { "arxivId": "2610.12611", "title": "SoftwareGym: Scaling Computer-Use Agent Training with Diverse Synthesized Software Environments", "summary": "SoftwareGym automates synthesis of diverse, executable GUI applications to provide scalable training data that improves computer-use agent performance on out-of-domain benchmarks.", "authors": [ "Tian Jin", "Wenhao Wang", "Jing Kang", "Haoting Shi", "Rui Ye", "Yijun Lu" ], "organization": { "_id": "63e5ef7bf2e9a8f22c515654", "name": "SJTU", "fullname": "Shanghai Jiao Tong University", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/1676013394657-63e5ee22b6a40bf941da0928.png" }, "publishedAt": "2026-10-08T00:00:00.000Z", "submittedAt": "2026-10-12T00:00:00.000Z", "upvotes": 107, "comments": 2, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.12611.png", "arxivUrl": "https://arxiv.org/abs/2610.12611", "pdfUrl": "https://arxiv.org/pdf/2610.12611.pdf", "hfUrl": "https://huggingface.co/papers/2610.12611", "githubRepo": "https://github.com/MrKiing-Jin/SoftwareGym", "githubStars": 18, "keywords": [ "computer-use agents", "GUI software environments", "SoftwareGym", "GUI evaluation", "iterative repair", "SoftwareGym-Bench" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "SoftwareGym: Scaling Computer-Use Agent Training with Diverse Synthesized Software Environments" }, "summary_i18n": { "en": "SoftwareGym automates synthesis of diverse, executable GUI applications to provide scalable training data that improves computer-use agent performance on out-of-domain benchmarks." } }, { "arxivId": "2610.12772", "title": "Reflex: Visually Grounded Reactive Humanoid Control", "summary": "Reflex is a three-stage framework that enables humanoid robots to catch thrown boxes using egocentric vision by decomposing whole-body control, dynamics inference, and visual perception into sequentially trained stages.", "authors": [ "Taoyang Jia", "Weikai Huang", "Linxin Song", "Jared Darlington", "Jieyu Zhao", "Yue Wang" ], "organization": { "_id": "6315a1bb86b3db2ac420100e", "name": "UW", "fullname": "University of Washington", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/61ac8f8a00d01045fca0ad2f/gr5B_WVvbMr4kTox5UkwZ.jpeg" }, "publishedAt": "2026-10-08T00:00:00.000Z", "submittedAt": "2026-10-12T00:00:00.000Z", "upvotes": 88, "comments": 2, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.12772.png", "arxivUrl": "https://arxiv.org/abs/2610.12772", "pdfUrl": "https://arxiv.org/pdf/2610.12772.pdf", "hfUrl": "https://huggingface.co/papers/2610.12772", "githubRepo": "https://github.com/ToberJ/reflex", "githubStars": 30, "keywords": [ "reinforcement learning", "state-based policy", "egocentric RGB-D encoder", "object dynamics", "whole-body control", "visual reactive control" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Reflex: Visually Grounded Reactive Humanoid Control" }, "summary_i18n": { "en": "Reflex is a three-stage framework that enables humanoid robots to catch thrown boxes using egocentric vision by decomposing whole-body control, dynamics inference, and visual perception into sequentially trained stages." } }, { "arxivId": "2610.07987", "title": "VisionWeave: Weaving Elastic Visual Representations as a Native Capability of MLLMs", "summary": "VisionWeave introduces elastic visual representation weaving to adaptively allocate visual tokens by granularity, reducing token usage by 43% while preserving nearly all model performance across diverse tasks and serving conditions.", "authors": [ "Yuan Feng", "Qize Yang", "Ruizhe Chen", "Sibo Song", "Haolin He", "Muzhi Zhu" ], "organization": { "_id": "64c8b5837fe12ecd0a7e92eb", "name": "Qwen", "fullname": "Qwen", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/6215ca5692c0ecfba9186921/hrRM50-6XcdWgg2AKpENG.jpeg" }, "publishedAt": "2026-10-06T00:00:00.000Z", "submittedAt": "2026-10-12T00:00:00.000Z", "upvotes": 68, "comments": 2, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.07987.png", "arxivUrl": "https://arxiv.org/abs/2610.07987", "pdfUrl": "https://arxiv.org/pdf/2610.07987.pdf", "hfUrl": "https://huggingface.co/papers/2610.07987", "githubRepo": "https://github.com/FFY0/VisionWeave", "githubStars": 75, "keywords": [ "multimodal large language models", "visual tokens", "token pruning", "elastic visual representation weaving", "gated spatial pooler", "MRoPE" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "VisionWeave: Weaving Elastic Visual Representations as a Native Capability of MLLMs" }, "summary_i18n": { "en": "VisionWeave introduces elastic visual representation weaving to adaptively allocate visual tokens by granularity, reducing token usage by 43% while preserving nearly all model performance across diverse tasks and serving conditions." } }, { "arxivId": "2610.14224", "title": "Harness^2: Recursive Agent Harnessing for an Open World", "summary": "Harness² improves agent execution harnesses through recursive task-level refinement and continual harness-level learning, boosting performance across benchmarks without ground-truth verifiers.", "authors": [ "Ruiyao Xu", "Yanfei Chen", "Zhongying CuiZhu", "Bhavana Dalvi Mishra", "Yifei Ming", "Han Yu" ], "organization": { "_id": "5e6aca39878b8b2bf9806447", "name": "google", "fullname": "Google", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/5dd96eb166059660ed1ee413/WtA3YYitedOr9n02eHfJe.png" }, "publishedAt": "2026-10-09T00:00:00.000Z", "submittedAt": "2026-10-14T00:00:00.000Z", "upvotes": 90, "comments": 1, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.14224.png", "arxivUrl": "https://arxiv.org/abs/2610.14224", "pdfUrl": "https://arxiv.org/pdf/2610.14224.pdf", "hfUrl": "https://huggingface.co/papers/2610.14224", "githubRepo": "https://github.com/google-research/harness2", "githubStars": 24, "keywords": [ "recursive self-improvement", "LLM agents", "execution harnesses", "Harness²", "Propose-Probe-Compose", "task-level recursion" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Harness^2: Recursive Agent Harnessing for an Open World" }, "summary_i18n": { "en": "Harness² improves agent execution harnesses through recursive task-level refinement and continual harness-level learning, boosting performance across benchmarks without ground-truth verifiers." } }, { "arxivId": "2610.19975", "title": "LatentLoop: Scaling Looped MoE Models with Asymmetric Compute Allocation", "summary": "LatentLoop improves looped Mixture-of-Experts by compressing the first loop visit, reallocating saved computation to larger hidden dimensions, and using separate attention parameters, yielding lower validation loss and better benchmark scores.", "authors": [ "Zefeng He", "Yunkai Chai", "Xiaoyang Chen", "Tong Zhu", "Zongliang Wu", "Yun Luo" ], "organization": null, "publishedAt": "2026-10-14T00:00:00.000Z", "submittedAt": "2026-10-15T00:00:00.000Z", "upvotes": 64, "comments": 1, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.19975.png", "arxivUrl": "https://arxiv.org/abs/2610.19975", "pdfUrl": "https://arxiv.org/pdf/2610.19975.pdf", "hfUrl": "https://huggingface.co/papers/2610.19975", "githubRepo": "https://github.com/lcqysl/LatentLoop", "githubStars": 10, "keywords": [ "Mixture-of-Experts", "Looped MoE", "residual coefficients", "LatentLoop", "asymmetric compute allocation", "cross-visit weight sharing" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "LatentLoop: Scaling Looped MoE Models with Asymmetric Compute Allocation" }, "summary_i18n": { "en": "LatentLoop improves looped Mixture-of-Experts by compressing the first loop visit, reallocating saved computation to larger hidden dimensions, and using separate attention parameters, yielding lower validation loss and better benchmark scores." } }, { "arxivId": "2610.19366", "title": "Seeing, Maintaining, and Learning: The Evolution of Multimodal Agent Memory", "summary": "This work proposes a two-axis framework for multimodal agent memory covering information formation and adaptive management to support lifelong learning.", "authors": [ "Minghao Guo", "Qingyue Jiao", "Xiujin Liu", "Jianhong Pang", "Ruimeng Yang", "Naihao Xu" ], "organization": { "_id": "643cb0625fcffe09fb6ca688", "name": "Fudan-University", "fullname": "Fudan University", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/6437eca0819f3ab20d162e14/kWv0cGlAhAG3iNWVxowkJ.png" }, "publishedAt": "2026-10-14T01:11:57.000Z", "submittedAt": "2026-10-15T00:00:00.000Z", "upvotes": 45, "comments": 1, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.19366.png", "arxivUrl": "https://arxiv.org/abs/2610.19366", "pdfUrl": "https://arxiv.org/pdf/2610.19366.pdf", "hfUrl": "https://huggingface.co/papers/2610.19366", "githubRepo": "https://github.com/Seeing-Maintaining-Learning/multimodal-agent-memory-evolution", "githubStars": 31, "keywords": [ "multimodal agents", "memory mechanisms", "modality-specific embeddings", "latent representations", "parametric representations", "static memory" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Seeing, Maintaining, and Learning: The Evolution of Multimodal Agent Memory" }, "summary_i18n": { "en": "This work proposes a two-axis framework for multimodal agent memory covering information formation and adaptive management to support lifelong learning." } }, { "arxivId": "2610.17201", "title": "Multics: Generative Multi-Player World Model with Hierarchical Communication", "summary": "Multics introduces hierarchical group-level communication and dynamic grouping to scale multi-player world models efficiently while reducing interference and maintaining consistency.", "authors": [ "Kaiyi Huang", "Wangbo Yu", "Yukun Huang", "Jiwen Yu", "Kunhao Liu", "Donglin Yang" ], "organization": { "_id": "60e3f7f641ca131919975fe5", "name": "TencentARC", "fullname": "ARC Lab, Tencent", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/1625552871844-60e272ca6c78a8c122b12127.png" }, "publishedAt": "2026-10-12T00:00:00.000Z", "submittedAt": "2026-10-14T00:00:00.000Z", "upvotes": 57, "comments": 2, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.17201.png", "arxivUrl": "https://arxiv.org/abs/2610.17201", "pdfUrl": "https://arxiv.org/pdf/2610.17201.pdf", "hfUrl": "https://huggingface.co/papers/2610.17201", "githubRepo": "https://github.com/TencentARC/Multics", "githubStars": 68, "keywords": [ "multi-player interactive world models", "hierarchical communication", "group tokens", "self-attention", "global tokens", "budget-constrained dynamic grouping" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Multics: Generative Multi-Player World Model with Hierarchical Communication" }, "summary_i18n": { "en": "Multics introduces hierarchical group-level communication and dynamic grouping to scale multi-player world models efficiently while reducing interference and maintaining consistency." } }, { "arxivId": "2610.13990", "title": "DoSA: Dots Sparse Attention", "summary": "DoSA accelerates sparse attention by using a certificate to conditionally score only eligible tokens, preserving exact selection while reducing computation.", "authors": [ "Fanxu Meng", "Taosong Fang" ], "organization": { "_id": "68246a0a98117c02df67a547", "name": "dots-studio", "fullname": "dots studio", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/683f0a13218449ec8f707047/K8V6vNs42cvpaTq_s6snK.png" }, "publishedAt": "2026-09-26T00:00:00.000Z", "submittedAt": "2026-10-14T00:00:00.000Z", "upvotes": 71, "comments": 1, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.13990.png", "arxivUrl": "https://arxiv.org/abs/2610.13990", "pdfUrl": "https://arxiv.org/pdf/2610.13990.pdf", "hfUrl": "https://huggingface.co/papers/2610.13990", "githubRepo": "https://github.com/llm-sys/flash-indexer", "githubStars": 26, "keywords": [ "DoSA", "sparse attention", "Top-K selection", "indexer head", "seed set", "nonnegative head projections" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "DoSA: Dots Sparse Attention" }, "summary_i18n": { "en": "DoSA accelerates sparse attention by using a certificate to conditionally score only eligible tokens, preserving exact selection while reducing computation." } }, { "arxivId": "2610.18072", "title": "LensErase: A Foundation Model for Erase Anything", "summary": "LensErase is a mask-based erasure foundation model pretrained on a large paired dataset and refined with reinforcement learning to handle diverse real-world removal tasks while preserving surrounding content.", "authors": [ "Yunuo Chen", "Xiuyu Wu", "Qinhong Yang", "Jinjing Zhao", "Yanchen Dong", "Sirui Zhang" ], "organization": { "_id": "5e6485f787403103f9f1055e", "name": "microsoft", "fullname": "Microsoft", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/1583646260758-5e64858c87403103f9f1055d.png" }, "publishedAt": "2026-10-13T09:56:02.000Z", "submittedAt": "2026-10-15T00:00:00.000Z", "upvotes": 36, "comments": 1, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.18072.png", "arxivUrl": "https://arxiv.org/abs/2610.18072", "pdfUrl": "https://arxiv.org/pdf/2610.18072.pdf", "hfUrl": "https://huggingface.co/papers/2610.18072", "githubRepo": "https://github.com/LensFamily/LensErase", "githubStars": 53, "keywords": [ "Mask-Aware RoPE", "3D rotary positional encoding", "text-to-image model", "supervised fine-tuning", "reinforcement learning", "contract-guided rewards" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "LensErase: A Foundation Model for Erase Anything" }, "summary_i18n": { "en": "LensErase is a mask-based erasure foundation model pretrained on a large paired dataset and refined with reinforcement learning to handle diverse real-world removal tasks while preserving surrounding content." } }, { "arxivId": "2610.13166", "title": "Scaling Agentic Data for Long-Horizon Terminal Intelligence", "summary": "TerminalHorizon automates long-horizon terminal agent data synthesis by reconstructing environments around omitted work and progressively expanding tasks, yielding substantial gains across coding, production, and reasoning benchmarks.", "authors": [ "Shun Zou", "Ziyu Ma", "Yi Zou", "Yong Wang", "Lin Chen", "Zehui Chen" ], "organization": { "_id": "68be41370a3fcebdcad6516a", "name": "alibabagroup", "fullname": "alibaba", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/68be3ab7e52df040b2cf80dc/li4G29u_EGswyTN1Sm_Kq.png" }, "publishedAt": "2026-10-09T00:00:00.000Z", "submittedAt": "2026-10-13T00:00:00.000Z", "upvotes": 160, "comments": 2, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.13166.png", "arxivUrl": "https://arxiv.org/abs/2610.13166", "pdfUrl": "https://arxiv.org/pdf/2610.13166.pdf", "hfUrl": "https://huggingface.co/papers/2610.13166", "githubRepo": null, "githubStars": null, "keywords": [ "long-horizon terminal agents", "environment synthesis", "supervised fine-tuning", "Qwen3.5-27B", "TerminalHorizon-27B" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Scaling Agentic Data for Long-Horizon Terminal Intelligence" }, "summary_i18n": { "en": "TerminalHorizon automates long-horizon terminal agent data synthesis by reconstructing environments around omitted work and progressively expanding tasks, yielding substantial gains across coding, production, and reasoning benchmarks." } }, { "arxivId": "2610.16794", "title": "TPPO: Tracking-Value PPO For Language Model Reasoning", "summary": "Tracking-Value PPO improves reinforcement learning for reasoning by modeling value evolution through cumulative log-odds updates, yielding better accuracy and generalization.", "authors": [ "Futing Wang", "Chenxi Li", "Jianhao Yan", "Yun Luo", "Zhi Wang", "Yizhuo Li" ], "organization": null, "publishedAt": "2026-10-12T00:00:00.000Z", "submittedAt": "2026-10-14T00:00:00.000Z", "upvotes": 78, "comments": 3, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.16794.png", "arxivUrl": "https://arxiv.org/abs/2610.16794", "pdfUrl": "https://arxiv.org/pdf/2610.16794.pdf", "hfUrl": "https://huggingface.co/papers/2610.16794", "githubRepo": "https://github.com/LINs-lab/TPPO", "githubStars": 14, "keywords": [ "Proximal Policy Optimization", "value evolution", "sparse terminal rewards", "Bayesian decomposition", "log-odds", "Tracking-Value PPO" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "TPPO: Tracking-Value PPO For Language Model Reasoning" }, "summary_i18n": { "en": "Tracking-Value PPO improves reinforcement learning for reasoning by modeling value evolution through cumulative log-odds updates, yielding better accuracy and generalization." } }, { "arxivId": "2610.20572", "title": "Recovering Reliable Supervision via Critique-Guided On-Policy Distillation", "summary": "CriticOPD improves on-policy distillation by using teacher critiques to restore recoverability, enabling students to revise incorrect reasoning and achieve stronger mathematical reasoning performance.", "authors": [ "Kaixiang Zhao", "Tianrun Yu", "Shangzhe Li", "Yuxiao Yang", "Amanda Hughes", "Weitong Zhang" ], "organization": null, "publishedAt": "2026-10-14T00:00:00.000Z", "submittedAt": "2026-10-15T00:00:00.000Z", "upvotes": 71, "comments": 1, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.20572.png", "arxivUrl": "https://arxiv.org/abs/2610.20572", "pdfUrl": "https://arxiv.org/pdf/2610.20572.pdf", "hfUrl": "https://huggingface.co/papers/2610.20572", "githubRepo": "https://github.com/kzhao5/CriticOPD", "githubStars": 2, "keywords": [ "on-policy distillation", "recoverability", "CriticOPD", "token-level supervision", "privileged information" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Recovering Reliable Supervision via Critique-Guided On-Policy Distillation" }, "summary_i18n": { "en": "CriticOPD improves on-policy distillation by using teacher critiques to restore recoverability, enabling students to revise incorrect reasoning and achieve stronger mathematical reasoning performance." } }, { "arxivId": "2610.13021", "title": "EmbodiedCLBench: Evaluating Continual Learning for Self-Evolving Embodied Agents", "summary": "EmbodiedCLBench evaluates continual learning in embodied agents by testing compositional skill reuse across basic and complex tasks, revealing stable but non-scaling experience benefits.", "authors": [ "Jie Huang", "Yanan Chen", "Ruixun Liu", "Kaichen He", "Xinyi He", "Rui Huang" ], "organization": { "_id": "6a07ff936781d8b803c28343", "name": "PKU-VaLuE-Lab", "fullname": "PKU-VaLuE-Lab", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/62b4fe7ab25cb80fcf2ffd66/velgm3kH_oJ2b0aoapBBM.png" }, "publishedAt": "2026-10-09T00:00:00.000Z", "submittedAt": "2026-10-12T00:00:00.000Z", "upvotes": 73, "comments": 3, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.13021.png", "arxivUrl": "https://arxiv.org/abs/2610.13021", "pdfUrl": "https://arxiv.org/pdf/2610.13021.pdf", "hfUrl": "https://huggingface.co/papers/2610.13021", "githubRepo": "https://github.com/PKU-VaLuE-Lab/EmbodiedCLBench", "githubStars": 16, "keywords": [ "continual learning", "embodied environments", "compositional generalization", "self-evolving agents", "EmbodiedCLBench" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "EmbodiedCLBench: Evaluating Continual Learning for Self-Evolving Embodied Agents" }, "summary_i18n": { "en": "EmbodiedCLBench evaluates continual learning in embodied agents by testing compositional skill reuse across basic and complex tasks, revealing stable but non-scaling experience benefits." } }, { "arxivId": "2610.14709", "title": "Vela: User Behavior as Supervision Sparks LLM Intelligence", "summary": "Vela converts real-world user behavior records into training tasks and trajectories, improving large language model performance on user-related and general benchmarks through supervised fine-tuning and reinforcement learning.", "authors": [ "Yang Xiao", "Yusong Sun", "Haoyi Wu", "Wenyang Hui", "Weiqiang Hang", "Luqing Wang" ], "organization": { "_id": "6a96bf5bae44840b913e27b8", "name": "AllSpark-Research", "fullname": "AllSpark Research", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/6a96be78642bb1433bbbf002/e9W8yeoVttMzcdc_Z_VrR.png" }, "publishedAt": "2026-10-10T00:00:00.000Z", "submittedAt": "2026-10-14T00:00:00.000Z", "upvotes": 71, "comments": 6, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.14709.png", "arxivUrl": "https://arxiv.org/abs/2610.14709", "pdfUrl": "https://arxiv.org/pdf/2610.14709.pdf", "hfUrl": "https://huggingface.co/papers/2610.14709", "githubRepo": "https://github.com/AllSpark-Research/Vela", "githubStars": 17, "keywords": [ "large language models", "user behavior supervision", "Vela", "fine-tuning", "reinforcement learning", "Qwen3.6-35B-A3B" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Vela: User Behavior as Supervision Sparks LLM Intelligence" }, "summary_i18n": { "en": "Vela converts real-world user behavior records into training tasks and trajectories, improving large language model performance on user-related and general benchmarks through supervised fine-tuning and reinforcement learning." } }, { "arxivId": "2610.20606v1", "title": "Modeling and Understanding Neural Weights of Humanoid Policies with Diffusion", "summary": "Recent advances in robot learning have driven rapid progress in humanoid control, enabling increasingly agile behaviors in the real world. However, humanoid control networks are typically obtained through conventional training procedures, while comparatively little attention has been given to directly modeling the distribution of their learned weights and generating functional networks in weight space. In this work, we investigate whether high-performing humanoid control networks can be generatively modeled directly through their weights. We adopt a latent diffusion framework that couples a linear autoencoder with a Diffusion Transformer. The autoencoder compresses actor network parameters into a compact latent space, and the Diffusion Transformer models the distribution of these latents. Trained on networks collected from reinforcement learning for the Unitree G1 robot, our approach generates deployable networks that perform various motions in the real world, including walking, dancing, and spin kick. Beyond generation, we further study the learned weight space to better understand the structure captured by these humanoid control networks. We identify a structural principle in the first-layer weights that consistently reflects the robot's kinematic prior. Notably, this kinematic prior emerges under both Reinforcement Learning (RL) and Imitation Learning (IL) training, suggesting that it is not specific to a particular optimization process but may instead reflect common knowledge associated with humanoid embodiment. Together, these results demonstrate the feasibility of generative modeling in humanoid network weight space, while providing initial insights into the information encoded by their learned weights. Additional results are available on our project website: https://humanoid-weight-diffusion.github.io/", "authors": [ "Huixuan Cheng", "Bike Zhang", "Shankar Sastry", "Koushil Sreenath" ], "organization": null, "publishedAt": "2026-10-14T17:59:52Z", "submittedAt": "2026-10-14T17:59:52Z", "upvotes": 0, "comments": 0, "thumbnail": null, "arxivUrl": "https://arxiv.org/abs/2610.20606v1", "pdfUrl": "https://arxiv.org/pdf/2610.20606v1", "hfUrl": "https://arxiv.org/abs/2610.20606v1", "githubRepo": null, "githubStars": null, "keywords": [ "cs.RO", "cs.AI", "cs.LG" ], "source": "arXiv cs.AI", "sourceType": "arxiv", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Modeling and Understanding Neural Weights of Humanoid Policies with Diffusion" }, "summary_i18n": { "en": "Recent advances in robot learning have driven rapid progress in humanoid control, enabling increasingly agile behaviors in the real world. However, humanoid control networks are typically obtained through conventional training procedures, while comparatively little attention has been given to directly modeling the distribution of their learned weights and generating functional networks in weight space. In this work, we investigate whether high-performing humanoid control networks can be generatively modeled directly through their weights. We adopt a latent diffusion framework that couples a linear autoencoder with a Diffusion Transformer. The autoencoder compresses actor network parameters into a compact latent space, and the Diffusion Transformer models the distribution of these latents. Trained on networks collected from reinforcement learning for the Unitree G1 robot, our approach generates deployable networks that perform various motions in the real world, including walking, dancing, and spin kick. Beyond generation, we further study the learned weight space to better understand the structure captured by these humanoid control networks. We identify a structural principle in the first-layer weights that consistently reflects the robot's kinematic prior. Notably, this kinematic prior emerges under both Reinforcement Learning (RL) and Imitation Learning (IL) training, suggesting that it is not specific to a particular optimization process but may instead reflect common knowledge associated with humanoid embodiment. Together, these results demonstrate the feasibility of generative modeling in humanoid network weight space, while providing initial insights into the information encoded by their learned weights. Additional results are available on our project website: https://humanoid-weight-diffusion.github.io/" } }, { "arxivId": "2610.20607v1", "title": "Hierarchical Residual Policy Learning for Real-World Mobile Manipulation with Sparse Human Guidance", "summary": "Mobile Manipulation (MoMa) is a cornerstone of autonomous robotic systems, yet real-world fine-tuning remains a formidable challenge because of high-dimensional action spaces and intricate coupling between heterogeneous modules. In this paper, we propose HiRe-MoMa (Hierarchical Residual Mobile Manipulation), a systematic framework for stable and sample-efficient real-world fine-tuning of complex MoMa systems. At its core, HiRe-MoMa adopts a Hierarchical Residual Architecture that fine-tunes a frozen base policy. A Policy-Agnostic Residual learns additive corrections atop the base policy, while Hierarchical Decoupling splits the residual into mobile-base and manipulator-arm heads with unidirectional conditioning, mitigating gradient interference and aligning corrections with the platform's physical hierarchy. To improve sample efficiency in data-expensive real-world settings, we propose Expert-Guided Value Alignment, which leverages sparse human interventions with under 20% training coverage to reshape the critic's value landscape, effectively guiding autonomous exploration through expert priors. Extensive real-world experiments on the TIAGo++ platform demonstrate that HiRe-MoMa consistently enhances task performance with only around 30 minutes of online interaction, regardless of the base policy architecture. To our knowledge, this is among the first frameworks to enable efficient on-robot fine-tuning for mobile manipulation in the real world, providing a practical paradigm for deploying complex robotic systems in dynamic, human-centric environments. Further details can be found on our project website: https://hire-moma.github.io/", "authors": [ "Zhefei Gong", "Hefeifei Jiang", "Zehao Wang", "Jiachen Li" ], "organization": null, "publishedAt": "2026-10-14T17:59:52Z", "submittedAt": "2026-10-14T17:59:52Z", "upvotes": 0, "comments": 0, "thumbnail": null, "arxivUrl": "https://arxiv.org/abs/2610.20607v1", "pdfUrl": "https://arxiv.org/pdf/2610.20607v1", "hfUrl": "https://arxiv.org/abs/2610.20607v1", "githubRepo": null, "githubStars": null, "keywords": [ "cs.RO", "cs.AI", "cs.LG" ], "source": "arXiv cs.AI", "sourceType": "arxiv", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Hierarchical Residual Policy Learning for Real-World Mobile Manipulation with Sparse Human Guidance" }, "summary_i18n": { "en": "Mobile Manipulation (MoMa) is a cornerstone of autonomous robotic systems, yet real-world fine-tuning remains a formidable challenge because of high-dimensional action spaces and intricate coupling between heterogeneous modules. In this paper, we propose HiRe-MoMa (Hierarchical Residual Mobile Manipulation), a systematic framework for stable and sample-efficient real-world fine-tuning of complex MoMa systems. At its core, HiRe-MoMa adopts a Hierarchical Residual Architecture that fine-tunes a frozen base policy. A Policy-Agnostic Residual learns additive corrections atop the base policy, while Hierarchical Decoupling splits the residual into mobile-base and manipulator-arm heads with unidirectional conditioning, mitigating gradient interference and aligning corrections with the platform's physical hierarchy. To improve sample efficiency in data-expensive real-world settings, we propose Expert-Guided Value Alignment, which leverages sparse human interventions with under 20% training coverage to reshape the critic's value landscape, effectively guiding autonomous exploration through expert priors. Extensive real-world experiments on the TIAGo++ platform demonstrate that HiRe-MoMa consistently enhances task performance with only around 30 minutes of online interaction, regardless of the base policy architecture. To our knowledge, this is among the first frameworks to enable efficient on-robot fine-tuning for mobile manipulation in the real world, providing a practical paradigm for deploying complex robotic systems in dynamic, human-centric environments. Further details can be found on our project website: https://hire-moma.github.io/" } }, { "arxivId": "2610.20597v1", "title": "Co-Evolving Acting and Forecasting for Language Agents", "summary": "Training language agents is bottlenecked by the high cost of long-horizon rollouts, yet standard agentic reinforcement learning (RL) learns only from a terminal scalar reward, leaving the supervision in intermediate steps largely unused. Prior work exploits this supervision through value targets that estimate per-step success probabilities or observation targets that fit environment response. Both are hard to construct well: (1) value targets come from the same single scalar signal for each step in the trajectory, and (2) observation targets, such as web pages and terminal logs, are long and noisy and reach only one step ahead. We propose \\name, which adds a third kind of multi-step action targets. On successful on-policy trajectories, \\name trains the policy backbone to generate the action sequences that follows each history. Each action is also supervised when forecast from the preceding histories, forming a temporal ensemble that the shared backbone absorbs during training, and this enables acting and forecasting to co-evolve. The policy still acts one step at a time at inference, with no overhead. Across five benchmarks (ALFWorld, ScienceWorld, WebShop, $τ^2$-bench, and AppWorld), \\name (1) achieves stronger aggregate performance than GRPO and RL baselines with predictive auxiliary targets across simulated environments, while improving overall performance on both realistic applications; (2) delivers up to $1.9\\times$ and $2.0\\times$ higher sample efficiency than ECHO and GRPO, respectively, on ScienceWorld; and (3) achieves these gains with a single policy backbone, without additional models, separate pretraining, or inference overhead. The codebase is available at https://github.com/ulab-uiuc/CoPE.", "authors": [ "Haofei Yu", "Yixing Chen", "Yining Zhao", "Xuyan Ye", "Yuxiang Liu", "Zhigang Hua" ], "organization": null, "publishedAt": "2026-10-14T17:59:28Z", "submittedAt": "2026-10-14T17:59:28Z", "upvotes": 0, "comments": 0, "thumbnail": null, "arxivUrl": "https://arxiv.org/abs/2610.20597v1", "pdfUrl": "https://arxiv.org/pdf/2610.20597v1", "hfUrl": "https://arxiv.org/abs/2610.20597v1", "githubRepo": null, "githubStars": null, "keywords": [ "cs.AI", "cs.LG" ], "source": "arXiv cs.AI", "sourceType": "arxiv", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Co-Evolving Acting and Forecasting for Language Agents" }, "summary_i18n": { "en": "Training language agents is bottlenecked by the high cost of long-horizon rollouts, yet standard agentic reinforcement learning (RL) learns only from a terminal scalar reward, leaving the supervision in intermediate steps largely unused. Prior work exploits this supervision through value targets that estimate per-step success probabilities or observation targets that fit environment response. Both are hard to construct well: (1) value targets come from the same single scalar signal for each step in the trajectory, and (2) observation targets, such as web pages and terminal logs, are long and noisy and reach only one step ahead. We propose \\name, which adds a third kind of multi-step action targets. On successful on-policy trajectories, \\name trains the policy backbone to generate the action sequences that follows each history. Each action is also supervised when forecast from the preceding histories, forming a temporal ensemble that the shared backbone absorbs during training, and this enables acting and forecasting to co-evolve. The policy still acts one step at a time at inference, with no overhead. Across five benchmarks (ALFWorld, ScienceWorld, WebShop, $τ^2$-bench, and AppWorld), \\name (1) achieves stronger aggregate performance than GRPO and RL baselines with predictive auxiliary targets across simulated environments, while improving overall performance on both realistic applications; (2) delivers up to $1.9\\times$ and $2.0\\times$ higher sample efficiency than ECHO and GRPO, respectively, on ScienceWorld; and (3) achieves these gains with a single policy backbone, without additional models, separate pretraining, or inference overhead. The codebase is available at https://github.com/ulab-uiuc/CoPE." } }, { "arxivId": "2610.20588v1", "title": "Your Transformer Is Secretly Elastic", "summary": "Scaling up transformers has produced models with broad capabilities, deployed across heterogeneous hardware and fluctuating workloads. Yet, model families offer only a few fixed sizes (e.g., S, B, L), forcing users to commit to a model based on hardware and latency constraints. This rigidity is often suboptimal under varying query loads. We introduce a retraining-free method that elastifies pre-trained transformers into a single model that adapts on-the-fly across a continuum of computational budgets. Our key contribution is a differentiable relaxation of the functional unit-ranking problem: unlike prior work, which optimizes a single importance weight per MLP block or attention head, we rank every individual MLP neuron and attention head jointly, on a common scale, by minimizing the model's loss under soft pruning across many sparsity targets, all at once. Across 27 vision and text transformer encoders ranging from 22M to 8B parameters, our method yields smooth degradation up to 60% sparsity and outperforms one-shot baselines under deep pruning. For example, pruning DINOv3 ViT-H+/16 to 50% sparsity removes 427M parameters while reducing ImageNet-1k linear-probing accuracy by only 0.8%, without the need for weight correction or retraining. At 60% sparsity, a pruned AugReg ViT-B/16 model achieves up to $1.80\\times$ higher throughput and $1.26\\times$ lower single-query latency on an A100 GPU. Elastic models are produced in minutes, with construction time scaling sub-linearly with parameter count.", "authors": [ "Walter Simoncini", "Dheeraj Varghese", "Gertjan J. Burghouts", "Cees G. M. Snoek" ], "organization": null, "publishedAt": "2026-10-14T17:58:08Z", "submittedAt": "2026-10-14T17:58:08Z", "upvotes": 0, "comments": 0, "thumbnail": null, "arxivUrl": "https://arxiv.org/abs/2610.20588v1", "pdfUrl": "https://arxiv.org/pdf/2610.20588v1", "hfUrl": "https://arxiv.org/abs/2610.20588v1", "githubRepo": null, "githubStars": null, "keywords": [ "cs.LG", "cs.AI" ], "source": "arXiv cs.AI", "sourceType": "arxiv", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Your Transformer Is Secretly Elastic" }, "summary_i18n": { "en": "Scaling up transformers has produced models with broad capabilities, deployed across heterogeneous hardware and fluctuating workloads. Yet, model families offer only a few fixed sizes (e.g., S, B, L), forcing users to commit to a model based on hardware and latency constraints. This rigidity is often suboptimal under varying query loads. We introduce a retraining-free method that elastifies pre-trained transformers into a single model that adapts on-the-fly across a continuum of computational budgets. Our key contribution is a differentiable relaxation of the functional unit-ranking problem: unlike prior work, which optimizes a single importance weight per MLP block or attention head, we rank every individual MLP neuron and attention head jointly, on a common scale, by minimizing the model's loss under soft pruning across many sparsity targets, all at once. Across 27 vision and text transformer encoders ranging from 22M to 8B parameters, our method yields smooth degradation up to 60% sparsity and outperforms one-shot baselines under deep pruning. For example, pruning DINOv3 ViT-H+/16 to 50% sparsity removes 427M parameters while reducing ImageNet-1k linear-probing accuracy by only 0.8%, without the need for weight correction or retraining. At 60% sparsity, a pruned AugReg ViT-B/16 model achieves up to $1.80\\times$ higher throughput and $1.26\\times$ lower single-query latency on an A100 GPU. Elastic models are produced in minutes, with construction time scaling sub-linearly with parameter count." } }, { "arxivId": "2610.20582v1", "title": "Latent Compaction with Test-Time Training", "summary": "Long contexts enlarge key--value (KV) caches, increasing memory use and attention costs during generation. Compressing these caches must preserve information for unseen questions. We introduce **Latent Compaction**, which learns a compact document-specific cache through test-time training. We initialize cache entries using reconstruction importance, teacher attention, and key-space diversity. Training first fits attention outputs and log-normalizers to those of the full cache, with reconstruction-selected entries held fixed. It then jointly optimizes the entire cache through end-to-end distillation from the full-cache model. Model weights remain frozen throughout, and training uses only generic prompts about the document, without access to downstream questions or answers. Experiments on RULER, LongBench, and Phonebook Lookup demonstrate substantial gains over budget-matched compression baselines. On a RULER, for example, our method achieves 96.7% accuracy at 32$\\times$ compression, compared with 40.0% for KVZip+. The method also transfers across model scales and families, with retrieval gains on both Qwen3 and MiniCPM5. These results show that treating the KV cache as a neural network and optimizing it with test-time training can preserve retrieval accuracy under aggressive compression. The code is released at https://www.github.com/THUNLP/latent-compaction", "authors": [ "Yingfa Chen", "Xingyu Shen", "Chaojun Xiao", "Xu Han" ], "organization": null, "publishedAt": "2026-10-14T17:57:01Z", "submittedAt": "2026-10-14T17:57:01Z", "upvotes": 0, "comments": 0, "thumbnail": null, "arxivUrl": "https://arxiv.org/abs/2610.20582v1", "pdfUrl": "https://arxiv.org/pdf/2610.20582v1", "hfUrl": "https://arxiv.org/abs/2610.20582v1", "githubRepo": null, "githubStars": null, "keywords": [ "cs.CL", "cs.LG" ], "source": "arXiv cs.CL", "sourceType": "arxiv", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Latent Compaction with Test-Time Training" }, "summary_i18n": { "en": "Long contexts enlarge key--value (KV) caches, increasing memory use and attention costs during generation. Compressing these caches must preserve information for unseen questions. We introduce **Latent Compaction**, which learns a compact document-specific cache through test-time training. We initialize cache entries using reconstruction importance, teacher attention, and key-space diversity. Training first fits attention outputs and log-normalizers to those of the full cache, with reconstruction-selected entries held fixed. It then jointly optimizes the entire cache through end-to-end distillation from the full-cache model. Model weights remain frozen throughout, and training uses only generic prompts about the document, without access to downstream questions or answers. Experiments on RULER, LongBench, and Phonebook Lookup demonstrate substantial gains over budget-matched compression baselines. On a RULER, for example, our method achieves 96.7% accuracy at 32$\\times$ compression, compared with 40.0% for KVZip+. The method also transfers across model scales and families, with retrieval gains on both Qwen3 and MiniCPM5. These results show that treating the KV cache as a neural network and optimizing it with test-time training can preserve retrieval accuracy under aggressive compression. The code is released at https://www.github.com/THUNLP/latent-compaction" } }, { "arxivId": "2610.20570v1", "title": "Manacá-1B-Instruct-2: Reproducible Safety Alignment for a Native Brazilian-Portuguese LLM. Negative DPO Results, a Controlled Composition Scan, and a Silent Failure Mode of the LLM Judge", "summary": "Brazilian Portuguese remains underserved by open-source language models, and the few that are safety-aligned rarely document what did not work or validate comparisons with statistical significance. We present Manacá-1B-Instruct-2, the second-generation post-training alignment of Manacá-1B (1.72B parameters, trained from scratch for PT-BR), with an SFT corpus of 291,223 examples. We document three DPO attempts, all negative, including a case where the training reward margin reaches 1.0 while free generation does not change at all, and show that safety came from a dedicated safety-SFT (LLM-jp AnswerCarefully style), refined over three iterations until closing the safety-helpfulness frontier. A controlled scan of the fraction of general data in the safety-SFT mix (10% to 60%) reveals a non-monotonic over-refusal curve with a knee point at 40-50%. The final model is compared against four external PT-BR models on the same items, with Wilson confidence intervals and paired McNemar tests with Holm-Bonferroni correction: the lower bound of its safety interval (94.1%) lies above the upper bound of all four baselines. We also report a methodological finding for LLM-judge evaluation: on two of six safety items in MT-Bench-PT, the judge categorically refuses to assign a score regardless of the evaluated response, silently reducing the effective N of that category in every prior report of the project. We complement this with an n-gram decontamination check and a Chinchilla-anchored compute-budget table that attributes the gap in general capability to pre-training scale rather than alignment methodology. All code, raw data, logs, and scripts that produce every reported number are publicly released.", "authors": [ "Bruno Leonardo Santos Menezes", "Carlos Leonardo Souza Cardoso", "Fabio André Machado Porto" ], "organization": null, "publishedAt": "2026-10-14T17:53:46Z", "submittedAt": "2026-10-14T17:53:46Z", "upvotes": 0, "comments": 0, "thumbnail": null, "arxivUrl": "https://arxiv.org/abs/2610.20570v1", "pdfUrl": "https://arxiv.org/pdf/2610.20570v1", "hfUrl": "https://arxiv.org/abs/2610.20570v1", "githubRepo": null, "githubStars": null, "keywords": [ "cs.CL" ], "source": "arXiv cs.CL", "sourceType": "arxiv", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Manacá-1B-Instruct-2: Reproducible Safety Alignment for a Native Brazilian-Portuguese LLM. Negative DPO Results, a Controlled Composition Scan, and a Silent Failure Mode of the LLM Judge" }, "summary_i18n": { "en": "Brazilian Portuguese remains underserved by open-source language models, and the few that are safety-aligned rarely document what did not work or validate comparisons with statistical significance. We present Manacá-1B-Instruct-2, the second-generation post-training alignment of Manacá-1B (1.72B parameters, trained from scratch for PT-BR), with an SFT corpus of 291,223 examples. We document three DPO attempts, all negative, including a case where the training reward margin reaches 1.0 while free generation does not change at all, and show that safety came from a dedicated safety-SFT (LLM-jp AnswerCarefully style), refined over three iterations until closing the safety-helpfulness frontier. A controlled scan of the fraction of general data in the safety-SFT mix (10% to 60%) reveals a non-monotonic over-refusal curve with a knee point at 40-50%. The final model is compared against four external PT-BR models on the same items, with Wilson confidence intervals and paired McNemar tests with Holm-Bonferroni correction: the lower bound of its safety interval (94.1%) lies above the upper bound of all four baselines. We also report a methodological finding for LLM-judge evaluation: on two of six safety items in MT-Bench-PT, the judge categorically refuses to assign a score regardless of the evaluated response, silently reducing the effective N of that category in every prior report of the project. We complement this with an n-gram decontamination check and a Chinchilla-anchored compute-budget table that attributes the gap in general capability to pre-training scale rather than alignment methodology. All code, raw data, logs, and scripts that produce every reported number are publicly released." } }, { "arxivId": "2610.20503v1", "title": "ReaDy: Recurrent States as Dynamic Context for Speculative Drafting in Hybrid Linear Attention Large Language Models", "summary": "Speculative decoding accelerates autoregressive generation by drafting multiple tokens per step, yet feature-based drafters rely on target hidden features whose contextual emphasis may not align with the needs of subsequent draft positions. Hybrid linear attention (HLA) models maintain compact recurrent states that accumulate historical associations, offering an efficient source of complementary context for speculative drafting. We introduce ReaDy, the first speculative decoding method, to the best of our knowledge, to leverage HLA recurrent states for position-specific context retrieval during drafting. ReaDy equips each draft position with a learned query to retrieve supplementary context directly from the target's committed recurrent states. A lightweight state reader integrates this context into draft representations through residual updates, while low-rank state reads and fused operators keep additional computation and I/O overhead low. Building on DSpark's parallel backbone and sequential prediction head, ReaDy employs a lightweight two-stage adaptation strategy, with reader capacity, state sources, and residual placement examined through ablation studies. Across seven benchmarks spanning mathematics, code generation, and dialogue, ReaDy achieves an average $3.16\\times$ decoding speedup over autoregressive decoding with Qwen3.8-27B on a single NVIDIA B200 GPU. It outperforms DSpark on every benchmark, improving throughput by 7.97% and acceptance length by 8.50% on average. Our code is available at https://github.com/StartLuxLabs/ReaDy.", "authors": [ "Bohan Sun", "Zunhai Su", "Xialie Zhuang", "Jianjin Li", "Ngai Wong", "Chuan-Wei Kuo" ], "organization": null, "publishedAt": "2026-10-14T17:08:52Z", "submittedAt": "2026-10-14T17:08:52Z", "upvotes": 0, "comments": 0, "thumbnail": null, "arxivUrl": "https://arxiv.org/abs/2610.20503v1", "pdfUrl": "https://arxiv.org/pdf/2610.20503v1", "hfUrl": "https://arxiv.org/abs/2610.20503v1", "githubRepo": null, "githubStars": null, "keywords": [ "cs.CL" ], "source": "arXiv cs.CL", "sourceType": "arxiv", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "ReaDy: Recurrent States as Dynamic Context for Speculative Drafting in Hybrid Linear Attention Large Language Models" }, "summary_i18n": { "en": "Speculative decoding accelerates autoregressive generation by drafting multiple tokens per step, yet feature-based drafters rely on target hidden features whose contextual emphasis may not align with the needs of subsequent draft positions. Hybrid linear attention (HLA) models maintain compact recurrent states that accumulate historical associations, offering an efficient source of complementary context for speculative drafting. We introduce ReaDy, the first speculative decoding method, to the best of our knowledge, to leverage HLA recurrent states for position-specific context retrieval during drafting. ReaDy equips each draft position with a learned query to retrieve supplementary context directly from the target's committed recurrent states. A lightweight state reader integrates this context into draft representations through residual updates, while low-rank state reads and fused operators keep additional computation and I/O overhead low. Building on DSpark's parallel backbone and sequential prediction head, ReaDy employs a lightweight two-stage adaptation strategy, with reader capacity, state sources, and residual placement examined through ablation studies. Across seven benchmarks spanning mathematics, code generation, and dialogue, ReaDy achieves an average $3.16\\times$ decoding speedup over autoregressive decoding with Qwen3.8-27B on a single NVIDIA B200 GPU. It outperforms DSpark on every benchmark, improving throughput by 7.97% and acceptance length by 8.50% on average. Our code is available at https://github.com/StartLuxLabs/ReaDy." } }, { "arxivId": "2610.20474v1", "title": "AdaRSI: On-Demand Recursive Self-Improvement for Embodied Agents", "summary": "Pretrained vision-language-action models (VLAs) often fail midway through unseen long-horizon tasks. Recursive self-improvement (RSI) offers a remedy, but embodied RSI methods either refine the harness around a frozen executor and so cannot add a missing motion, or train the executor on persistent failures without asking whether a failure lies in how the executor is called or in what it can do. We ask when a self-improving embodied system should change the weights of its executor, and propose AdaRSI, an on-demand RSI framework for embodied agents. A coding agent first improves its skills and harness, then attributes persistent failures by retrying failing sub-steps from saved states, and fine-tunes only the sub-steps that remain unreliable, with demonstrations it collects itself and experience carried over from earlier tasks. On all 50 target tasks of RoboCasa365, with GR00T N1.5 as the executor and no human demonstrations, AdaRSI improves the VLA on all three suites and raises its success on unseen composite tasks from 4.6% to 13.8%. Compared with fine-tuning every sub-step, it reaches slightly higher success while using only 28% of the GPU time and 22% of the self-collected demonstrations, and fine-tuning without a harness trains more yet gains far less.", "authors": [ "Peidong Wang", "Ying Chang", "Zhiming Ma", "Xiaocui Yang", "Shi Feng", "Daqi Liu" ], "organization": null, "publishedAt": "2026-10-14T16:51:09Z", "submittedAt": "2026-10-14T16:51:09Z", "upvotes": 0, "comments": 0, "thumbnail": null, "arxivUrl": "https://arxiv.org/abs/2610.20474v1", "pdfUrl": "https://arxiv.org/pdf/2610.20474v1", "hfUrl": "https://arxiv.org/abs/2610.20474v1", "githubRepo": null, "githubStars": null, "keywords": [ "cs.CL", "cs.RO" ], "source": "arXiv cs.CL", "sourceType": "arxiv", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "AdaRSI: On-Demand Recursive Self-Improvement for Embodied Agents" }, "summary_i18n": { "en": "Pretrained vision-language-action models (VLAs) often fail midway through unseen long-horizon tasks. Recursive self-improvement (RSI) offers a remedy, but embodied RSI methods either refine the harness around a frozen executor and so cannot add a missing motion, or train the executor on persistent failures without asking whether a failure lies in how the executor is called or in what it can do. We ask when a self-improving embodied system should change the weights of its executor, and propose AdaRSI, an on-demand RSI framework for embodied agents. A coding agent first improves its skills and harness, then attributes persistent failures by retrying failing sub-steps from saved states, and fine-tunes only the sub-steps that remain unreliable, with demonstrations it collects itself and experience carried over from earlier tasks. On all 50 target tasks of RoboCasa365, with GR00T N1.5 as the executor and no human demonstrations, AdaRSI improves the VLA on all three suites and raises its success on unseen composite tasks from 4.6% to 13.8%. Compared with fine-tuning every sub-step, it reaches slightly higher success while using only 28% of the GPU time and 22% of the self-collected demonstrations, and fine-tuning without a harness trains more yet gains far less." } }, { "arxivId": "2610.20610v1", "title": "Can Learning to Answer Spatial Questions Improve Novel-View Generation?", "summary": "Generating novel views from limited observations requires preserving both the appearance and spatial structure of a scene. Although recent approaches exploit fine-grained geometric cues, they do not explicitly model high-level semantic and spatial properties such as object identity, attributes, and inter-object relationships. In this work, we investigate whether 3D language understanding can provide complementary supervision for novel-view generation. We introduce Ask4NV, built on a pretrained unified multimodal model with a Mixture-of-Transformers (MoT) architecture. Our model incorporates camera geometry into both VAE and ViT representations and leverages understanding features during generation through joint self-attention with a modality-aware attention masking strategy. Experiments demonstrate strong image quality and geometric consistency in novel-view generation, alongside competitive 3D language understanding. Further analysis shows that joint training improves cross-view correspondence in VAE representations and similarity in a 3D-aware feature space.", "authors": [ "Jahyeok Koo", "Hyeonseo Yu", "Seungho Jang", "Jaewoo Jung", "Jiwon Kang", "Seokju Cho" ], "organization": null, "publishedAt": "2026-10-14T17:59:58Z", "submittedAt": "2026-10-14T17:59:58Z", "upvotes": 0, "comments": 0, "thumbnail": null, "arxivUrl": "https://arxiv.org/abs/2610.20610v1", "pdfUrl": "https://arxiv.org/pdf/2610.20610v1", "hfUrl": "https://arxiv.org/abs/2610.20610v1", "githubRepo": null, "githubStars": null, "keywords": [ "cs.CV" ], "source": "arXiv cs.CV", "sourceType": "arxiv", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Can Learning to Answer Spatial Questions Improve Novel-View Generation?" }, "summary_i18n": { "en": "Generating novel views from limited observations requires preserving both the appearance and spatial structure of a scene. Although recent approaches exploit fine-grained geometric cues, they do not explicitly model high-level semantic and spatial properties such as object identity, attributes, and inter-object relationships. In this work, we investigate whether 3D language understanding can provide complementary supervision for novel-view generation. We introduce Ask4NV, built on a pretrained unified multimodal model with a Mixture-of-Transformers (MoT) architecture. Our model incorporates camera geometry into both VAE and ViT representations and leverages understanding features during generation through joint self-attention with a modality-aware attention masking strategy. Experiments demonstrate strong image quality and geometric consistency in novel-view generation, alongside competitive 3D language understanding. Further analysis shows that joint training improves cross-view correspondence in VAE representations and similarity in a 3D-aware feature space." } }, { "arxivId": "2610.20608v1", "title": "Representation Conditioning for Generation", "summary": "Representation-alignment methods effectively accelerate the diffusion transformer training. REPA, for example, uses a pretrained vision encoder to supervise an intermediate layer through an MLP projection head which is discarded after training. It aims to import semantics from the vision encoder to the diffusion backbone. While this objective is achieved at early layers by back-propagating the alignment loss, there is no guarantee that later layers, supervised by the velocity prediction loss alone, can well access such semantics. We propose a simple solution, representation conditioning (RECON), to improve semantics in later layers. During training, we feed the pretrained vision encoder tokens into later layers; during inference, because we no longer have the clean image, we reuse the MLP head (as a proxy to the vision encoder) and project its output tokens to the later layers. Our method adds no parameters and leads to much faster convergence: RECON is $63\\times$ faster than SiT and $25\\times$ faster than REPA. Code is available at https://github.com/saeidrazavi/RECON", "authors": [ "Saeed Razavi", "Xingjian Leng", "Liang Zheng" ], "organization": null, "publishedAt": "2026-10-14T17:59:52Z", "submittedAt": "2026-10-14T17:59:52Z", "upvotes": 0, "comments": 0, "thumbnail": null, "arxivUrl": "https://arxiv.org/abs/2610.20608v1", "pdfUrl": "https://arxiv.org/pdf/2610.20608v1", "hfUrl": "https://arxiv.org/abs/2610.20608v1", "githubRepo": null, "githubStars": null, "keywords": [ "cs.CV" ], "source": "arXiv cs.CV", "sourceType": "arxiv", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Representation Conditioning for Generation" }, "summary_i18n": { "en": "Representation-alignment methods effectively accelerate the diffusion transformer training. REPA, for example, uses a pretrained vision encoder to supervise an intermediate layer through an MLP projection head which is discarded after training. It aims to import semantics from the vision encoder to the diffusion backbone. While this objective is achieved at early layers by back-propagating the alignment loss, there is no guarantee that later layers, supervised by the velocity prediction loss alone, can well access such semantics. We propose a simple solution, representation conditioning (RECON), to improve semantics in later layers. During training, we feed the pretrained vision encoder tokens into later layers; during inference, because we no longer have the clean image, we reuse the MLP head (as a proxy to the vision encoder) and project its output tokens to the later layers. Our method adds no parameters and leads to much faster convergence: RECON is $63\\times$ faster than SiT and $25\\times$ faster than REPA. Code is available at https://github.com/saeidrazavi/RECON" } }, { "arxivId": "2610.20604v1", "title": "VITAL: From Visual Motion to Precise Action", "summary": "World-action models aim to turn visual anticipation into driving decisions, yet typically train action experts on trajectory representations defined independently of visual dynamics. This motivates deriving action targets directly from visual motion. We first examine a simpler setting in which trajectories are reconstructed from recorded future images. Even without the need for future prediction, reconstruction captures the dominant maneuver but leaves centimeter-scale errors. We therefore introduce VITAL (Vision-Induced Tokens for Action Learning), an action tokenizer that couples a visually derived motion scaffold with a compact residual encoding the trajectory error left by that scaffold. This representation makes visual motion the principal prediction target and gives small metric corrections a separate, normalized learning scale. Using flow matching, the action expert jointly predicts both components from the current image, ego state, and navigation command, without generating future images at inference. The same residual coordinates support bounded reinforcement post-training that refines execution while keeping the predicted visual scaffold fixed. With only one action denoising step, VITAL achieves 90.8 PDMS on NAVSIM v1 navtest, within 0.13 points of its ten-step score at 3.29x the end-to-end inference speed. Without retraining, it achieves 91.1 EPDMS on NAVSIM v2 navtest and 38.6 on navhard. Project page: https://worldbench.github.io/vital", "authors": [ "Ao Liang", "Youquan Liu", "Mingxiao Li", "Wei Chow", "Lingdong Kong", "Yue Liao" ], "organization": null, "publishedAt": "2026-10-14T17:59:49Z", "submittedAt": "2026-10-14T17:59:49Z", "upvotes": 0, "comments": 0, "thumbnail": null, "arxivUrl": "https://arxiv.org/abs/2610.20604v1", "pdfUrl": "https://arxiv.org/pdf/2610.20604v1", "hfUrl": "https://arxiv.org/abs/2610.20604v1", "githubRepo": null, "githubStars": null, "keywords": [ "cs.CV" ], "source": "arXiv cs.CV", "sourceType": "arxiv", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "VITAL: From Visual Motion to Precise Action" }, "summary_i18n": { "en": "World-action models aim to turn visual anticipation into driving decisions, yet typically train action experts on trajectory representations defined independently of visual dynamics. This motivates deriving action targets directly from visual motion. We first examine a simpler setting in which trajectories are reconstructed from recorded future images. Even without the need for future prediction, reconstruction captures the dominant maneuver but leaves centimeter-scale errors. We therefore introduce VITAL (Vision-Induced Tokens for Action Learning), an action tokenizer that couples a visually derived motion scaffold with a compact residual encoding the trajectory error left by that scaffold. This representation makes visual motion the principal prediction target and gives small metric corrections a separate, normalized learning scale. Using flow matching, the action expert jointly predicts both components from the current image, ego state, and navigation command, without generating future images at inference. The same residual coordinates support bounded reinforcement post-training that refines execution while keeping the predicted visual scaffold fixed. With only one action denoising step, VITAL achieves 90.8 PDMS on NAVSIM v1 navtest, within 0.13 points of its ten-step score at 3.29x the end-to-end inference speed. Without retraining, it achieves 91.1 EPDMS on NAVSIM v2 navtest and 38.6 on navhard. Project page: https://worldbench.github.io/vital" } }, { "arxivId": "2610.20603v1", "title": "CommerceGen-Studio: An End-to-End Framework for Commercial-Grade Main-Image Generation", "summary": "We present CommerceGen-Studio, an end-to-end framework for generating commercial-grade main images for e-commerce Stock Keeping Units (SKUs). Our framework is organized along a three-stage pipeline of information extraction, semantic rewriting, and multi-condition generation: Prompt-Eval-Refine-Loop Agent (PERL-Agent) distills heterogeneous SKU metadata into concise selling-point text lists through an iterative closed loop of VLM and LLM roles scored by a reference-free rubric. Layout Thinker Optimization (LTO) aligns the rewriter with both caption and rendered-image quality via alternating domain-specific policy optimization. SKU-DiT, trained under Reinforce-Distill-Reinforce (RDR) progressive training, conditions jointly on the white-background product image, the trademark crop, and the structured layout to hold subject fidelity, trademark integrity, and character-level text accuracy together. The framework runs in production at a scale of millions of main images per day with a latency of a few seconds per image, which we believe makes it among the first systems of this kind to operate at commercial scale. We also release CommerceGen-Bench, a benchmark of hard cases collected from live traffic across all product categories. On it, our framework achieves the highest overall scores among the evaluated open-source baselines on both editing tracks, with and without prompt enhancement. It also matches much larger closed-source models, and beats several of them at Instruction Following and Product Integration - the two properties that determine whether an image can be published.", "authors": [ "Xiaoli Xu", "Zhengtao Wu", "Zhengze Xu", "Tao Liu", "Yongchao Du", "Zuan Gao" ], "organization": null, "publishedAt": "2026-10-14T17:59:48Z", "submittedAt": "2026-10-14T17:59:48Z", "upvotes": 0, "comments": 0, "thumbnail": null, "arxivUrl": "https://arxiv.org/abs/2610.20603v1", "pdfUrl": "https://arxiv.org/pdf/2610.20603v1", "hfUrl": "https://arxiv.org/abs/2610.20603v1", "githubRepo": null, "githubStars": null, "keywords": [ "cs.CV" ], "source": "arXiv cs.CV", "sourceType": "arxiv", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "CommerceGen-Studio: An End-to-End Framework for Commercial-Grade Main-Image Generation" }, "summary_i18n": { "en": "We present CommerceGen-Studio, an end-to-end framework for generating commercial-grade main images for e-commerce Stock Keeping Units (SKUs). Our framework is organized along a three-stage pipeline of information extraction, semantic rewriting, and multi-condition generation: Prompt-Eval-Refine-Loop Agent (PERL-Agent) distills heterogeneous SKU metadata into concise selling-point text lists through an iterative closed loop of VLM and LLM roles scored by a reference-free rubric. Layout Thinker Optimization (LTO) aligns the rewriter with both caption and rendered-image quality via alternating domain-specific policy optimization. SKU-DiT, trained under Reinforce-Distill-Reinforce (RDR) progressive training, conditions jointly on the white-background product image, the trademark crop, and the structured layout to hold subject fidelity, trademark integrity, and character-level text accuracy together. The framework runs in production at a scale of millions of main images per day with a latency of a few seconds per image, which we believe makes it among the first systems of this kind to operate at commercial scale. We also release CommerceGen-Bench, a benchmark of hard cases collected from live traffic across all product categories. On it, our framework achieves the highest overall scores among the evaluated open-source baselines on both editing tracks, with and without prompt enhancement. It also matches much larger closed-source models, and beats several of them at Instruction Following and Product Integration - the two properties that determine whether an image can be published." } }, { "arxivId": "2610.20534v1", "title": "Revisiting Simple Baselines for Relational Deep Learning", "summary": "Relational deep learning has emerged as a powerful paradigm for learning end-to-end predictive models over relational databases by representing them as temporal heterogeneous graphs. Recent advances based on graph neural networks and Transformers have achieved strong performance across diverse domains. However, the effectiveness of simpler relational baselines remains underexplored, making it difficult to assess how much architectural complexity is needed to achieve strong predictive performance. Motivated by the observation that predictive signals arise from target attributes, relational attributes, and historical labels, we revisit how effectively these signals can be exploited using simple propagation methods. We adapt fixed low-pass feature propagation to relational entity graphs and train an MLP for prediction. The resulting predictions are further refined using Correct-and-Smooth, a graph-based label propagation algorithm that incorporates historical labels over task-specific meta-graphs connecting target entities. Across classification and regression tasks on RelBench, an MLP trained on neighborhood-filtered features outperforms one using only target-entity features on most tasks and recovers a substantial portion of the performance of heterogeneous GNNs. Correct-and-Smooth provides further gains from historical labels, although applying it requires a separately constructed meta-graph because connections among target entities are not directly specified by the relational entity graph. Together, our studies revisit the effectiveness of established propagation methods as simple RDL baselines while identifying canonical meta-graph construction as a key challenge for effective label propagation.", "authors": [ "Viet The Nguyen", "Vijay Prakash Dwivedi", "Gunther Gust" ], "organization": null, "publishedAt": "2026-10-14T17:30:17Z", "submittedAt": "2026-10-14T17:30:17Z", "upvotes": 0, "comments": 0, "thumbnail": null, "arxivUrl": "https://arxiv.org/abs/2610.20534v1", "pdfUrl": "https://arxiv.org/pdf/2610.20534v1", "hfUrl": "https://arxiv.org/abs/2610.20534v1", "githubRepo": null, "githubStars": null, "keywords": [ "cs.LG" ], "source": "arXiv cs.LG", "sourceType": "arxiv", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Revisiting Simple Baselines for Relational Deep Learning" }, "summary_i18n": { "en": "Relational deep learning has emerged as a powerful paradigm for learning end-to-end predictive models over relational databases by representing them as temporal heterogeneous graphs. Recent advances based on graph neural networks and Transformers have achieved strong performance across diverse domains. However, the effectiveness of simpler relational baselines remains underexplored, making it difficult to assess how much architectural complexity is needed to achieve strong predictive performance. Motivated by the observation that predictive signals arise from target attributes, relational attributes, and historical labels, we revisit how effectively these signals can be exploited using simple propagation methods. We adapt fixed low-pass feature propagation to relational entity graphs and train an MLP for prediction. The resulting predictions are further refined using Correct-and-Smooth, a graph-based label propagation algorithm that incorporates historical labels over task-specific meta-graphs connecting target entities. Across classification and regression tasks on RelBench, an MLP trained on neighborhood-filtered features outperforms one using only target-entity features on most tasks and recovers a substantial portion of the performance of heterogeneous GNNs. Correct-and-Smooth provides further gains from historical labels, although applying it requires a separately constructed meta-graph because connections among target entities are not directly specified by the relational entity graph. Together, our studies revisit the effectiveness of established propagation methods as simple RDL baselines while identifying canonical meta-graph construction as a key challenge for effective label propagation." } }, { "arxivId": "2610.20525v1", "title": "The Physics Is the Network: Fast, Training-Free Neural Emulation", "summary": "Neural emulation need not require learning a substitute for known physics. We investigate PolyNets that couple a continuous-time neural representation of polynomial dynamics with a simple recurrent network for discrete-time integration. When the governing coefficients are known, the resulting tensor graph requires no training and preserves the prescribed numerical scheme while enabling efficient hardware execution and parallel simulation. Using the one-dimensional Korteweg-de Vries equation, we compare PolyNet with CPU and GPU finite-difference Runge-Kutta solvers and three trained neural models: a state-map emulator, a physics-informed predictor, and a learned-tendency integrator. Analytical soliton solutions provide a reference for amplitude and phase errors across grid resolutions, 10^6 dispersively scaled time steps, ensembles of up to 10^4 trajectories, and varying initial amplitudes. PolyNet achieves accuracy comparable to the numerical solvers, including at amplitudes outside the learned models' training conditions. It combines this fidelity with high GPU throughput, approximately constant over the tested resolution range, and favorable ensemble scaling.", "authors": [ "Luka Todorovic", "Kaushal Gianchandani", "Sai Ravela" ], "organization": null, "publishedAt": "2026-10-14T17:24:48Z", "submittedAt": "2026-10-14T17:24:48Z", "upvotes": 0, "comments": 0, "thumbnail": null, "arxivUrl": "https://arxiv.org/abs/2610.20525v1", "pdfUrl": "https://arxiv.org/pdf/2610.20525v1", "hfUrl": "https://arxiv.org/abs/2610.20525v1", "githubRepo": null, "githubStars": null, "keywords": [ "cs.LG", "eess.SY", "nlin.CD", "stat.AP" ], "source": "arXiv cs.LG", "sourceType": "arxiv", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "The Physics Is the Network: Fast, Training-Free Neural Emulation" }, "summary_i18n": { "en": "Neural emulation need not require learning a substitute for known physics. We investigate PolyNets that couple a continuous-time neural representation of polynomial dynamics with a simple recurrent network for discrete-time integration. When the governing coefficients are known, the resulting tensor graph requires no training and preserves the prescribed numerical scheme while enabling efficient hardware execution and parallel simulation. Using the one-dimensional Korteweg-de Vries equation, we compare PolyNet with CPU and GPU finite-difference Runge-Kutta solvers and three trained neural models: a state-map emulator, a physics-informed predictor, and a learned-tendency integrator. Analytical soliton solutions provide a reference for amplitude and phase errors across grid resolutions, 10^6 dispersively scaled time steps, ensembles of up to 10^4 trajectories, and varying initial amplitudes. PolyNet achieves accuracy comparable to the numerical solvers, including at amplitudes outside the learned models' training conditions. It combines this fidelity with high GPU throughput, approximately constant over the tested resolution range, and favorable ensemble scaling." } }, { "arxivId": "2610.20514v1", "title": "Bayesian Learning to Select Fair Classifiers", "summary": "Evaluations of classifiers (AI or human or agents) often assume a batch setting in which every agent produces labels on every example in a test set, which can be expensive or impractical in many contexts. This motivates online policies that adaptively select agents to query as examples arrive and update policies as outcomes are observed. We study the problem of selecting, in an online stream, the agent that best balances (i) predictive performance and (ii) subgroup fairness of a sensitive attribute, and that allows for delayed observations of true class labels. We formulate this as a Bayesian multi-armed bandit problem using Thompson Sampling, explicitly balancing accuracy and fairness. We conduct empirical evaluations on real-world datasets spanning problems involving tabular classifiers, large language models, and human labelers, demonstrating substantial reductions in regret relative to baselines. Our theoretical contributions include guarantees on the performance of the selected agent that incorporate both the accuracy and fairness aspects of the problem.", "authors": [ "Saatvik Kher", "Hyungrok Do", "Judy Zhong", "Padhraic Smyth" ], "organization": null, "publishedAt": "2026-10-14T17:16:28Z", "submittedAt": "2026-10-14T17:16:28Z", "upvotes": 0, "comments": 0, "thumbnail": null, "arxivUrl": "https://arxiv.org/abs/2610.20514v1", "pdfUrl": "https://arxiv.org/pdf/2610.20514v1", "hfUrl": "https://arxiv.org/abs/2610.20514v1", "githubRepo": null, "githubStars": null, "keywords": [ "cs.LG" ], "source": "arXiv cs.LG", "sourceType": "arxiv", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Bayesian Learning to Select Fair Classifiers" }, "summary_i18n": { "en": "Evaluations of classifiers (AI or human or agents) often assume a batch setting in which every agent produces labels on every example in a test set, which can be expensive or impractical in many contexts. This motivates online policies that adaptively select agents to query as examples arrive and update policies as outcomes are observed. We study the problem of selecting, in an online stream, the agent that best balances (i) predictive performance and (ii) subgroup fairness of a sensitive attribute, and that allows for delayed observations of true class labels. We formulate this as a Bayesian multi-armed bandit problem using Thompson Sampling, explicitly balancing accuracy and fairness. We conduct empirical evaluations on real-world datasets spanning problems involving tabular classifiers, large language models, and human labelers, demonstrating substantial reductions in regret relative to baselines. Our theoretical contributions include guarantees on the performance of the selected agent that incorporate both the accuracy and fairness aspects of the problem." } }, { "arxivId": "2610.20447v1", "title": "Testing Conditional Distribution Equality Using Generative Models", "summary": "Conditional distribution equality testing is a core problem with applications in causal inference, model diagnostics, and transfer learning. Existing methods often lose power when the conditioning variable is high-dimensional. We propose a new test that leverages generative models to compare conditional distributions using both observed data and synthetic samples. A sample-splitting and studentization scheme is introduced to control estimation error. Under the null, the test statistic is asymptotically normal, and we establish local power against contiguous alternatives. Our theory under the null and fixed alternative requires only consistency of the learned conditional distribution, allowing for a broad class of generative models. Empirical results demonstrate accurate size control and strong power in high-dimensional settings compared to representative kernel-based conditional and joint two-sample tests.", "authors": [ "Zekai Jin", "Yi Zhang", "Yun Yang", "Xiaofeng Shao" ], "organization": null, "publishedAt": "2026-10-14T16:33:59Z", "submittedAt": "2026-10-14T16:33:59Z", "upvotes": 0, "comments": 0, "thumbnail": null, "arxivUrl": "https://arxiv.org/abs/2610.20447v1", "pdfUrl": "https://arxiv.org/pdf/2610.20447v1", "hfUrl": "https://arxiv.org/abs/2610.20447v1", "githubRepo": null, "githubStars": null, "keywords": [ "stat.ME", "stat.ML" ], "source": "arXiv stat.ML", "sourceType": "arxiv", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Testing Conditional Distribution Equality Using Generative Models" }, "summary_i18n": { "en": "Conditional distribution equality testing is a core problem with applications in causal inference, model diagnostics, and transfer learning. Existing methods often lose power when the conditioning variable is high-dimensional. We propose a new test that leverages generative models to compare conditional distributions using both observed data and synthetic samples. A sample-splitting and studentization scheme is introduced to control estimation error. Under the null, the test statistic is asymptotically normal, and we establish local power against contiguous alternatives. Our theory under the null and fixed alternative requires only consistency of the learned conditional distribution, allowing for a broad class of generative models. Empirical results demonstrate accurate size control and strong power in high-dimensional settings compared to representative kernel-based conditional and joint two-sample tests." } }, { "arxivId": "2610.20413v1", "title": "When Does Momentum Help in Reinforcement Learning with Verifiable Rewards?", "summary": "Momentum is a classical tool for accelerating convex optimization and a standard component of modern large language model training. Yet recent empirical work has questioned the usefulness of first-moment momentum in reinforcement learning with verifiable rewards (RLVR). We study this question through a bandit formulation of RLVR with bias-corrected exponential moving average (EMA) momentum, where the EMA coefficient $β$ controls how quickly the weights on past gradients decay. When policy gradients are evaluated exactly, we prove that increasing $β$ within a range that prevents overshoot strictly accelerates optimization. When gradients are estimated from sampled responses, a positive expected gain can be obscured by sampling fluctuations. We show that as the learning rate $η$ decreases, the expected gain with fixed $β$ vanishes faster than these fluctuations, whereas increasing $β$ to keep $η/(1-β)$ within the safe regime preserves a nonvanishing gain that eventually dominates them. Bandit simulations quantitatively validate these predictions. RLVR experiments on Qwen3-4B-Base using SGD with EMA and AdamW support our principle of choosing larger coefficients in practical LLM post-training. On Qwen2.5-32B with AdamW, simply increasing $β_1$ improves mean validation accuracy by nearly $2$ percentage points over the conventional choice, demonstrating the principle's effectiveness at a larger model scale.", "authors": [ "Kaifei Wang", "Zikang Shan", "Liwei Wang", "Yinyu Ye", "Han Zhong" ], "organization": null, "publishedAt": "2026-10-14T16:11:14Z", "submittedAt": "2026-10-14T16:11:14Z", "upvotes": 0, "comments": 0, "thumbnail": null, "arxivUrl": "https://arxiv.org/abs/2610.20413v1", "pdfUrl": "https://arxiv.org/pdf/2610.20413v1", "hfUrl": "https://arxiv.org/abs/2610.20413v1", "githubRepo": null, "githubStars": null, "keywords": [ "cs.LG", "cs.AI", "math.OC", "stat.ML" ], "source": "arXiv stat.ML", "sourceType": "arxiv", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "When Does Momentum Help in Reinforcement Learning with Verifiable Rewards?" }, "summary_i18n": { "en": "Momentum is a classical tool for accelerating convex optimization and a standard component of modern large language model training. Yet recent empirical work has questioned the usefulness of first-moment momentum in reinforcement learning with verifiable rewards (RLVR). We study this question through a bandit formulation of RLVR with bias-corrected exponential moving average (EMA) momentum, where the EMA coefficient $β$ controls how quickly the weights on past gradients decay. When policy gradients are evaluated exactly, we prove that increasing $β$ within a range that prevents overshoot strictly accelerates optimization. When gradients are estimated from sampled responses, a positive expected gain can be obscured by sampling fluctuations. We show that as the learning rate $η$ decreases, the expected gain with fixed $β$ vanishes faster than these fluctuations, whereas increasing $β$ to keep $η/(1-β)$ within the safe regime preserves a nonvanishing gain that eventually dominates them. Bandit simulations quantitatively validate these predictions. RLVR experiments on Qwen3-4B-Base using SGD with EMA and AdamW support our principle of choosing larger coefficients in practical LLM post-training. On Qwen2.5-32B with AdamW, simply increasing $β_1$ improves mean validation accuracy by nearly $2$ percentage points over the conventional choice, demonstrating the principle's effectiveness at a larger model scale." } }, { "arxivId": "2610.20363v1", "title": "Fractional is Better: Learnable Derivative Orders in Neural Operator Learning", "summary": "Neural operators learn mappings between function spaces, enabling fast surrogate solutions to partial differential equations. Despite remarkable architectural diversity, these methods often share a common input representation: raw coordinate-value pairs $(x, u(x))$. We ask whether inputs aligned with PDE differential structure can improve learning. Through Picard iteration on mild solutions, we show that derivatives of the input appear explicitly in the solution operator, suggesting that providing derivative features should reduce the network's implicit differentiation burden. We prove this intuition: providing derivative features improves approximation bounds from $O(W^{-(s-m)/d})$ to $O(W^{-s/d})$, where $W$ is parameter count, $s$ is input regularity, $m$ is the PDE order, and $d$ is spatial dimension. Our central finding, however, is a surprise: the optimal derivative order $β^*$ is strictly less than the PDE order $m$. This gap arises from a bias-variance tradeoff in spectral space that we characterize in closed form. Learning $β$ from data achieves automatic spectral regularization. We introduce $\\partial$-NO (del-NO), for derivative-augmented neural operators, an architecture-agnostic augmentation that provides learnable fractional derivative features to any neural operator backbone. Across benchmark problems and architectures, $\\partial$-NO consistently improves prediction accuracy, with learned orders that reflect a representation of known physics modulated by noise and finite-sample constraints.", "authors": [ "Fares B. Mehouachi", "Saif Eddin Jabari" ], "organization": null, "publishedAt": "2026-10-14T15:37:48Z", "submittedAt": "2026-10-14T15:37:48Z", "upvotes": 0, "comments": 0, "thumbnail": null, "arxivUrl": "https://arxiv.org/abs/2610.20363v1", "pdfUrl": "https://arxiv.org/pdf/2610.20363v1", "hfUrl": "https://arxiv.org/abs/2610.20363v1", "githubRepo": null, "githubStars": null, "keywords": [ "cs.LG", "stat.ML" ], "source": "arXiv stat.ML", "sourceType": "arxiv", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Fractional is Better: Learnable Derivative Orders in Neural Operator Learning" }, "summary_i18n": { "en": "Neural operators learn mappings between function spaces, enabling fast surrogate solutions to partial differential equations. Despite remarkable architectural diversity, these methods often share a common input representation: raw coordinate-value pairs $(x, u(x))$. We ask whether inputs aligned with PDE differential structure can improve learning. Through Picard iteration on mild solutions, we show that derivatives of the input appear explicitly in the solution operator, suggesting that providing derivative features should reduce the network's implicit differentiation burden. We prove this intuition: providing derivative features improves approximation bounds from $O(W^{-(s-m)/d})$ to $O(W^{-s/d})$, where $W$ is parameter count, $s$ is input regularity, $m$ is the PDE order, and $d$ is spatial dimension. Our central finding, however, is a surprise: the optimal derivative order $β^*$ is strictly less than the PDE order $m$. This gap arises from a bias-variance tradeoff in spectral space that we characterize in closed form. Learning $β$ from data achieves automatic spectral regularization. We introduce $\\partial$-NO (del-NO), for derivative-augmented neural operators, an architecture-agnostic augmentation that provides learnable fractional derivative features to any neural operator backbone. Across benchmark problems and architectures, $\\partial$-NO consistently improves prediction accuracy, with learned orders that reflect a representation of known physics modulated by noise and finite-sample constraints." } }, { "arxivId": "2610.20321v1", "title": "Hierarchical Gaussian Processes for Survey Weighting: Local Equivalent Weights in Structured Populations", "summary": "Regression-based survey weighting connects model-based population estimation with design-based inference through equivalent weights. For nonlinear models, equivalent weights can be defined locally through the sensitivity of a poststratified population estimator. However, conventional regression models may inadequately capture nonlinear relationships and complex interactions across sparse poststratification cells. We develop a survey-weighting framework based on hierarchical Gaussian process (HGP) regression, combining structured group-level effects with kernel-based smoothing to share information across related cells. We derive Gaussian posterior prediction formulas that exploit the structured covariance representation to reduce computational costs. We also establish a uniform convergence result for the posterior predictive mean under stated regularity conditions and construct local equivalent weights by matching the first-order sensitivity of a plug-in poststratified estimand to that of a weighted average of fitted response means, with both derivatives taken with respect to the latent mean-scale vector. Simulations show that HGP achieves lower population-weighted prediction error under nonparametric structure, whereas logistic regression performs well under correct parametric specification. Additional analyses examine sample size, frame coverage, subgroup prediction, and population-mean agreement. An application to the Future of Families and Child Wellbeing Study illustrates the method. Overall, this work extends local equivalent-weight construction to nonparametric modeling and provides a computational framework and conditional theoretical results for survey weighting in structured, sparse populations.", "authors": [ "Seonghun Lee" ], "organization": null, "publishedAt": "2026-10-14T15:08:06Z", "submittedAt": "2026-10-14T15:08:06Z", "upvotes": 0, "comments": 0, "thumbnail": null, "arxivUrl": "https://arxiv.org/abs/2610.20321v1", "pdfUrl": "https://arxiv.org/pdf/2610.20321v1", "hfUrl": "https://arxiv.org/abs/2610.20321v1", "githubRepo": null, "githubStars": null, "keywords": [ "stat.ME", "stat.ML" ], "source": "arXiv stat.ML", "sourceType": "arxiv", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Hierarchical Gaussian Processes for Survey Weighting: Local Equivalent Weights in Structured Populations" }, "summary_i18n": { "en": "Regression-based survey weighting connects model-based population estimation with design-based inference through equivalent weights. For nonlinear models, equivalent weights can be defined locally through the sensitivity of a poststratified population estimator. However, conventional regression models may inadequately capture nonlinear relationships and complex interactions across sparse poststratification cells. We develop a survey-weighting framework based on hierarchical Gaussian process (HGP) regression, combining structured group-level effects with kernel-based smoothing to share information across related cells. We derive Gaussian posterior prediction formulas that exploit the structured covariance representation to reduce computational costs. We also establish a uniform convergence result for the posterior predictive mean under stated regularity conditions and construct local equivalent weights by matching the first-order sensitivity of a plug-in poststratified estimand to that of a weighted average of fitted response means, with both derivatives taken with respect to the latent mean-scale vector. Simulations show that HGP achieves lower population-weighted prediction error under nonparametric structure, whereas logistic regression performs well under correct parametric specification. Additional analyses examine sample size, frame coverage, subgroup prediction, and population-mean agreement. An application to the Future of Families and Child Wellbeing Study illustrates the method. Overall, this work extends local equivalent-weight construction to nonparametric modeling and provides a computational framework and conditional theoretical results for survey weighting in structured, sparse populations." } }, { "arxivId": "2610.19820", "title": "ActionOPD: Distilling Decisions, Not Tokens", "summary": "ActionOPD aligns smaller language-model agents with teachers via semantic action-level on-policy distillation, improving out-of-distribution tool-use performance without requiring shared tokenizers or teacher probabilities.", "authors": [ "Huining Yuan", "Xiangmin Yi", "Zelai Xu", "Zixu Wang", "Xinlei Chen", "Xiao-Ping Zhang" ], "organization": { "_id": "64b74b5fb727f8771ab887f9", "name": "nics-efc", "fullname": "Tsinghua-NICS-EFC", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/641031b1a78453b8d96b8420/vmgxct2WsyHcKT2x2NAYT.jpeg" }, "publishedAt": "2026-10-14T00:00:00.000Z", "submittedAt": "2026-10-15T00:00:00.000Z", "upvotes": 53, "comments": 1, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.19820.png", "arxivUrl": "https://arxiv.org/abs/2610.19820", "pdfUrl": "https://arxiv.org/pdf/2610.19820.pdf", "hfUrl": "https://huggingface.co/papers/2610.19820", "githubRepo": "https://github.com/thu-nics/ActionOPD", "githubStars": 7, "keywords": [ "agent distillation", "on-policy distillation", "token-level OPD", "semantic actions", "ActionOPD", "policy optimization" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "ActionOPD: Distilling Decisions, Not Tokens" }, "summary_i18n": { "en": "ActionOPD aligns smaller language-model agents with teachers via semantic action-level on-policy distillation, improving out-of-distribution tool-use performance without requiring shared tokenizers or teacher probabilities." } }, { "arxivId": "2610.17953", "title": "Agent-Reflex: Teaching Language Agents to Act Reflectively", "summary": "Agent-Reflex trains interactive agents to reflect and recover from noisy tool feedback and user corrections through skill discovery, supervised fine-tuning, and on-policy optimization.", "authors": [ "Guanting Dong", "Xinyu Yang", "Jiajie Jin", "Yuyang Hu", "Xiaoshuai Song", "Yifei Chen" ], "organization": { "_id": "6695ed048765c1560ce56423", "name": "RUC-NLPIR", "fullname": "NLPIR Lab @ RUC", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/625e62452a7279d3c77b5c38/CBwmyPCRzm4rHTGWhiCzR.jpeg" }, "publishedAt": "2026-10-13T00:00:00.000Z", "submittedAt": "2026-10-14T00:00:00.000Z", "upvotes": 49, "comments": 3, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.17953.png", "arxivUrl": "https://arxiv.org/abs/2610.17953", "pdfUrl": "https://arxiv.org/pdf/2610.17953.pdf", "hfUrl": "https://huggingface.co/papers/2610.17953", "githubRepo": "https://github.com/RUC-NLPIR/Agent-Reflex", "githubStars": 75, "keywords": [ "Agent-Reflex", "Reflection Skill Discovery", "Reflective Supervised Fine-Tuning", "On-Policy Reflective Policy Optimization", "long-horizon reasoning", "tool-use" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Agent-Reflex: Teaching Language Agents to Act Reflectively" }, "summary_i18n": { "en": "Agent-Reflex trains interactive agents to reflect and recover from noisy tool feedback and user corrections through skill discovery, supervised fine-tuning, and on-policy optimization." } }, { "arxivId": "2610.20442", "title": "Towards Agentic Robots: Model, Data, Environment, and Harness", "summary": "This survey unifies research on agentic robots by tracing the evolution from robot policies to integrated systems using vision-language-action models, world models, and structured functional domains.", "authors": [ "Chonghan Qin", "Dongming Wu", "Kaituo Feng", "Jiaxu Wang", "Xinlei Yu", "Peiwen Sun" ], "organization": null, "publishedAt": "2026-10-14T00:00:00.000Z", "submittedAt": "2026-10-15T00:00:00.000Z", "upvotes": 29, "comments": 1, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.20442.png", "arxivUrl": "https://arxiv.org/abs/2610.20442", "pdfUrl": "https://arxiv.org/pdf/2610.20442.pdf", "hfUrl": "https://huggingface.co/papers/2610.20442", "githubRepo": "https://github.com/chonghanqin/Awesome-Agentic-Robots", "githubStars": 66, "keywords": [ "vision-language-action models", "world models", "world-action models" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Towards Agentic Robots: Model, Data, Environment, and Harness" }, "summary_i18n": { "en": "This survey unifies research on agentic robots by tracing the evolution from robot policies to integrated systems using vision-language-action models, world models, and structured functional domains." } }, { "arxivId": "2610.18774", "title": "GreenWAM: A World Foundation Model for Robotics", "summary": "GreenWAM is a unified world foundation model for robotics that uses a shared anatomical action space and joint training to serve as both a predictive world model and a control policy across diverse robot bodies.", "authors": [ "Gennady Sidorov", "Valerii Kornilov", "Denis Grankin", "Polina Fedotova", "Aleksey Postnikov" ], "organization": { "_id": "6973998bee83f4964edef012", "name": "SberRoboticsCenter", "fullname": "Sber Robotics Center", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/642694c721d5f027bec07c71/LkuEJI3abphK4MFbq8tPf.png" }, "publishedAt": "2026-10-13T00:00:00.000Z", "submittedAt": "2026-10-14T00:00:00.000Z", "upvotes": 88, "comments": 2, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.18774.png", "arxivUrl": "https://arxiv.org/abs/2610.18774", "pdfUrl": "https://arxiv.org/pdf/2610.18774.pdf", "hfUrl": "https://huggingface.co/papers/2610.18774", "githubRepo": "https://github.com/greenvla/GreenWAM", "githubStars": 5, "keywords": [ "world foundation model", "anatomical action space", "joint training", "world-action modes", "action expert", "vision-language-action" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "GreenWAM: A World Foundation Model for Robotics" }, "summary_i18n": { "en": "GreenWAM is a unified world foundation model for robotics that uses a shared anatomical action space and joint training to serve as both a predictive world model and a control policy across diverse robot bodies." } }, { "arxivId": "2609.38748", "title": "Hear the World in Stereo: Learning Dynamic Spatial Correspondence for Immersive Joint Video-Audio Generation", "summary": "StereoBind binds visual motion to stereo sound generation via motion tracks and specialized encoders, improving spatial alignment in joint video-audio synthesis.", "authors": [ "Hanmo Chen", "Chengcheng Liu", "Tianxiao Chen", "Zheyu Zhang", "Siming Zheng", "Jinwei Chen" ], "organization": null, "publishedAt": "2026-10-08T00:00:00.000Z", "submittedAt": "2026-10-12T00:00:00.000Z", "upvotes": 65, "comments": 3, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2609.38748.png", "arxivUrl": "https://arxiv.org/abs/2609.38748", "pdfUrl": "https://arxiv.org/pdf/2609.38748.pdf", "hfUrl": "https://huggingface.co/papers/2609.38748", "githubRepo": "https://github.com/vivoCameraResearch/StereoBind", "githubStars": 13, "keywords": [ "joint video-audio generation", "stereo audio", "Dynamic Spatial Correspondence", "StereoBind", "motion tracks", "Visual Motion Binding" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Hear the World in Stereo: Learning Dynamic Spatial Correspondence for Immersive Joint Video-Audio Generation" }, "summary_i18n": { "en": "StereoBind binds visual motion to stereo sound generation via motion tracks and specialized encoders, improving spatial alignment in joint video-audio synthesis." } }, { "arxivId": "2610.12984", "title": "Work2Skill: Capturing Human GUI Experience for Reusable Agent Skill Induction", "summary": "Work2Skill captures human GUI demonstrations to induce reusable agent skills, improving execution performance on desktop tasks through workflow rules and step dependencies.", "authors": [ "Yuqi Huang", "Hui Li", "Yifei Cao", "Zikun Cai", "Tianci Xie", "Yangfan Gao" ], "organization": { "_id": "6907441c72f7d95376e910a5", "name": "nex-agi", "fullname": "Nex AGI", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/65435cad429b80b14922ab8d/a_O9jT_daz_NXTfxtcw6S.png" }, "publishedAt": "2026-10-09T00:00:00.000Z", "submittedAt": "2026-10-12T00:00:00.000Z", "upvotes": 51, "comments": 2, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.12984.png", "arxivUrl": "https://arxiv.org/abs/2610.12984", "pdfUrl": "https://arxiv.org/pdf/2610.12984.pdf", "hfUrl": "https://huggingface.co/papers/2610.12984", "githubRepo": "https://github.com/china-qijizhifeng/Work2Skill", "githubStars": 29, "keywords": [ "Work2Skill", "WorkRecorder", "WorkDemo-350", "accessibility-tree", "multimodal dataset", "agentic skill-induction" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Work2Skill: Capturing Human GUI Experience for Reusable Agent Skill Induction" }, "summary_i18n": { "en": "Work2Skill captures human GUI demonstrations to induce reusable agent skills, improving execution performance on desktop tasks through workflow rules and step dependencies." } }, { "arxivId": "2610.17227", "title": "AReA: Adaptive Receptive-Field Attention for Efficient High-Resolution Diffusion Transformers", "summary": "AReA improves high-resolution image generation with diffusion transformers by adapting attention from global to local across denoising steps without extra training.", "authors": [ "Javad Rajabi", "Koorosh Roohi", "David B. Lindell", "Babak Taati" ], "organization": { "_id": "6591d63868d0b76331f82b3b", "name": "uoft-cs", "fullname": "University of Toronto Computer Science", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/5e77ccfa3d77a72421292d0c/8cyh-p0m5Hbz8xhOxjfuI.jpeg" }, "publishedAt": "2026-10-12T00:00:00.000Z", "submittedAt": "2026-10-14T00:00:00.000Z", "upvotes": 49, "comments": 1, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.17227.png", "arxivUrl": "https://arxiv.org/abs/2610.17227", "pdfUrl": "https://arxiv.org/pdf/2610.17227.pdf", "hfUrl": "https://huggingface.co/papers/2610.17227", "githubRepo": "https://github.com/KooroshRH/AReA", "githubStars": 33, "keywords": [ "Diffusion Transformers", "DiTs", "full attention", "attention receptive field", "cross-scale coarse attention", "local attention" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "AReA: Adaptive Receptive-Field Attention for Efficient High-Resolution Diffusion Transformers" }, "summary_i18n": { "en": "AReA improves high-resolution image generation with diffusion transformers by adapting attention from global to local across denoising steps without extra training." } }, { "arxivId": "2610.13630", "title": "Multi-Stream Attention Residuals", "summary": "Multi-Stream Attention Residuals extend depth-wise attention to multi-head retrieval with adaptive mixing, improving efficiency and performance in large-scale pre-training.", "authors": [ "Xiaohan Qin", "Xiangdong Zhang", "Tuo Dai", "Sunan Zou", "Huaijin Wu", "Yebin Yang" ], "organization": { "_id": "63e5ef7bf2e9a8f22c515654", "name": "SJTU", "fullname": "Shanghai Jiao Tong University", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/1676013394657-63e5ee22b6a40bf941da0928.png" }, "publishedAt": "2026-10-09T00:00:00.000Z", "submittedAt": "2026-10-13T00:00:00.000Z", "upvotes": 47, "comments": 4, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.13630.png", "arxivUrl": "https://arxiv.org/abs/2610.13630", "pdfUrl": "https://arxiv.org/pdf/2610.13630.pdf", "hfUrl": "https://huggingface.co/papers/2610.13630", "githubRepo": "https://github.com/jianke0604/MSAR", "githubStars": 34, "keywords": [ "Attention Residuals", "multi-head depth attention", "multi-scale causal convolutions", "adaptive mixing", "MoE", "scaling laws" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Multi-Stream Attention Residuals" }, "summary_i18n": { "en": "Multi-Stream Attention Residuals extend depth-wise attention to multi-head retrieval with adaptive mixing, improving efficiency and performance in large-scale pre-training." } }, { "arxivId": "2610.13395", "title": "ProgramGym: Scaling Training Environments for Long-Horizon Program Reproduction Agents", "summary": "ProgramGym introduces scalable verifiable training environments and a role-isolated agent workflow to improve long-horizon program reproduction from scratch.", "authors": [ "Chengyi Yang", "Renxu Bao", "Jue Hou", "Zongpei Teng", "Ziheng Wang", "Yujie Lin" ], "organization": { "_id": "6a71a45f584d1d33007a4f56", "name": "ProgramGym", "fullname": "ProgramGym", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/68676f9642df3fdd87e5d368/PZQwVyqoKPrpaVSqYk1E7.png" }, "publishedAt": "2026-10-09T00:00:00.000Z", "submittedAt": "2026-10-12T00:00:00.000Z", "upvotes": 54, "comments": 2, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.13395.png", "arxivUrl": "https://arxiv.org/abs/2610.13395", "pdfUrl": "https://arxiv.org/pdf/2610.13395.pdf", "hfUrl": "https://huggingface.co/papers/2610.13395", "githubRepo": "https://github.com/Acm1011/ProgramGym", "githubStars": 13, "keywords": [ "ProgramGym", "Program-Env", "Program-Agent", "long-horizon program reproduction", "feedback-driven loop", "black-box tests" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "ProgramGym: Scaling Training Environments for Long-Horizon Program Reproduction Agents" }, "summary_i18n": { "en": "ProgramGym introduces scalable verifiable training environments and a role-isolated agent workflow to improve long-horizon program reproduction from scratch." } }, { "arxivId": "2610.20399", "title": "Cortex: Consolidating Context into Parametric Memory", "summary": "Cortex consolidates context-induced improvements into mergeable low-rank or structural updates, preserving non-target capabilities across sequential edits and diverse model scales.", "authors": [ "Jintian Zhang", "Feng He", "Xinnian Liang", "Ziwen Xu", "Haoming Xu", "Haojing Huang" ], "organization": { "_id": "67d1140985ea0644e2f14b99", "name": "ByteDance-Seed", "fullname": "ByteDance Seed", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/6535c9e88bde2fae19b6fb25/flkDUqd_YEuFsjeNET3r-.png" }, "publishedAt": "2026-10-14T00:00:00.000Z", "submittedAt": "2026-10-15T00:00:00.000Z", "upvotes": 63, "comments": 1, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.20399.png", "arxivUrl": "https://arxiv.org/abs/2610.20399", "pdfUrl": "https://arxiv.org/pdf/2610.20399.pdf", "hfUrl": "https://huggingface.co/papers/2610.20399", "githubRepo": null, "githubStars": null, "keywords": [ "Memory Consolidation", "context-specified editing", "skill learning", "behavioral steering", "prospective queries", "low-rank updates" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "Cortex: Consolidating Context into Parametric Memory" }, "summary_i18n": { "en": "Cortex consolidates context-induced improvements into mergeable low-rank or structural updates, preserving non-target capabilities across sequential edits and diverse model scales." } }, { "arxivId": "2609.37712", "title": "PolyOCR-Venus: Unified OCR Foundation Models for Text-Centric Visual Intelligence", "summary": "PolyOCR is a family of unified OCR foundation models that integrates instruction following, large-scale data synthesis, and competence-guided policy optimization to achieve strong recognition, parsing, and reasoning across diverse visual text tasks.", "authors": [ "GuangJian Team", "Kaili Huang", "Yongshuo Zhang", "Bingtao Fu", "Changjiang Jiang", "Chenfan Qu" ], "organization": { "_id": "67aea5c8f086ab0f70ed97c9", "name": "inclusionAI", "fullname": "inclusionAI", "avatar": "https://cdn-avatars.huggingface.co/v1/production/uploads/662e1f9da266499277937d33/fyKuazRifqiaIO34xrhhm.jpeg" }, "publishedAt": "2026-09-29T00:00:00.000Z", "submittedAt": "2026-10-12T00:00:00.000Z", "upvotes": 36, "comments": 2, "thumbnail": "https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2609.37712.png", "arxivUrl": "https://arxiv.org/abs/2609.37712", "pdfUrl": "https://arxiv.org/pdf/2609.37712.pdf", "hfUrl": "https://huggingface.co/papers/2609.37712", "githubRepo": "https://github.com/inclusionAI/PolyOCR-Venus", "githubStars": 59, "keywords": [ "PolyOCR", "instruction-following framework", "Competence-Guided Policy Optimization", "Group Relative Policy Optimization", "on-policy distillation", "sample-wise routing" ], "source": "HF Daily Papers", "sourceType": "hf", "title_zh": null, "summary_zh": null, "title_i18n": { "en": "PolyOCR-Venus: Unified OCR Foundation Models for Text-Centric Visual Intelligence" }, "summary_i18n": { "en": "PolyOCR is a family of unified OCR foundation models that integrates instruction following, large-scale data synthesis, and competence-guided policy optimization to achieve strong recognition, parsing, and reasoning across diverse visual text tasks." } } ], "i18nUpdatedAt": "2026-10-16T00:56:58.286Z", "locales": [ "en", "en", "zh-CN", "ja", "ko", "es", "de" ] }