{"meta":{"schema":"aimap.ai-papers.latest/v1","generatedAt":"2026-10-22T10:04:53+08:00","window":["2026-09-22","2026-10-21"],"dataAsOf":"2026-10-21","title":{"zh":"AI 最新论文","en":"Latest AI Papers"},"counts":{"papers":307,"labPosts":48,"labRepos":17,"hfCandidates":968},"tags":[{"key":"llm","zh":"大语言模型","en":"LLMs","count":27,"primary":16},{"key":"reasoning","zh":"推理","en":"Reasoning","count":28,"primary":8},{"key":"alignment","zh":"对齐","en":"Alignment","count":8,"primary":4},{"key":"agent","zh":"智能体","en":"Agents","count":119,"primary":72},{"key":"multimodal","zh":"多模态","en":"Multimodal","count":51,"primary":34},{"key":"visual-gen","zh":"视觉生成","en":"Visual generation","count":67,"primary":52},{"key":"vision","zh":"视觉理解与 3D","en":"Vision & 3D","count":21,"primary":10},{"key":"robotics","zh":"机器人与具身","en":"Robotics & embodied AI","count":40,"primary":37},{"key":"efficiency","zh":"高效训练与推理","en":"Efficient training & inference","count":69,"primary":23},{"key":"safety","zh":"安全与可解释","en":"Safety & interpretability","count":11,"primary":7},{"key":"ai4science","zh":"AI for Science","en":"AI for Science","count":19,"primary":9},{"key":"code","zh":"代码与软件工程","en":"Code & SWE","count":21,"primary":8},{"key":"benchmark","zh":"评测与数据","en":"Benchmarks & data","count":60,"primary":14},{"key":"rl","zh":"强化学习","en":"Reinforcement learning","count":48,"primary":13}],"hotFormula":"hot = 10·ln(1+HF点赞) + 2.5·ln(1+GitHub star) + 4·ln(1+引用) + 1.5·ln(1+HF评论) + 4（知名机构）","notes":{"zh":"候选来自 Hugging Face Daily Papers 30 天（社区精选，偏向有 HF 页面的论文）；标题/作者/日期以 arXiv API 为准；引用数来自 Semantic Scholar（新论文多为 0–数十）。要点为本站根据摘要自写。","en":"Candidates come from 30 days of Hugging Face Daily Papers (community-curated); titles/authors/dates from the arXiv API; citations from Semantic Scholar. One-line takeaways are written by AIMap from abstracts."},"sources":{"huggingface_daily_papers":{"ok":true,"note":"30 天可取，968 篇去重后","checked":"2026-10-22T02:00Z"},"arxiv_api":{"ok":true,"note":"元数据 600/600","checked":"2026-10-22T02:01Z"},"semantic_scholar":{"ok":true,"note":"命中 589/600（无 API key，公共限流）","checked":"2026-10-22T02:01Z"},"github_api":{"ok":true,"note":"补 star 0 次","checked":"2026-10-22T02:01Z"},"arxiv_rss":{"ok":true,"note":"{\"cs.AI\": {\"items\": 473, \"new\": 133}, \"cs.CL\": {\"items\": 435, \"new\": 173}, \"cs.LG\": {\"items\": 817, \"new\": 308}, \"cs.CV\": {\"items\": 287, \"new\": 154}, \"cs.RO\": {\"items\": 126, \"new\": 72}, \"stat.ML\": {\"items\": 71, \"new\": 16}, \"cs.CR\": {\"items\": 83, \"new\": 44}, \"cs.MA\": {\"items\": 19, \"new\": 3}}","checked":"2026-10-22T02:02Z"},"blog:OpenAI":{"ok":true,"note":"72 篇在窗口内","checked":"2026-10-22T02:02Z"},"blog:Google DeepMind":{"ok":true,"note":"11 篇在窗口内","checked":"2026-10-22T02:02Z"},"blog:Google Research":{"ok":true,"note":"11 篇在窗口内","checked":"2026-10-22T02:02Z"},"blog:Microsoft Research":{"ok":true,"note":"10 篇在窗口内","checked":"2026-10-22T02:02Z"},"blog:BAIR":{"ok":true,"note":"0 篇在窗口内","checked":"2026-10-22T02:02Z"},"blog:Hugging Face":{"ok":true,"note":"26 篇在窗口内","checked":"2026-10-22T02:02Z"},"blog:Qwen":{"ok":true,"note":"0 篇在窗口内","checked":"2026-10-22T02:02Z"},"blog:Anthropic":{"ok":true,"note":"HTML 解析，13 篇在窗口内（21 个链接）","checked":"2026-10-22T02:02Z"},"github_org_repos":{"ok":true,"note":"17/18 个组织可取，新仓库 22","checked":"2026-10-22T02:02Z"},"openreview_api":{"ok":false,"note":"HTTP 403 {\"name\":\"ChallengeRequiredError\",\"message\":\"Challenge verification required (2026-10-22-1385947)\",\"status\":403,\"details\"","checked":"2026-10-22T02:02Z"}},"note":"AIMap latest AI papers (30-day window). Takeaways written by AIMap; abstracts not reproduced. Archive: /api/papers/archive/index.json"},"papers":[{"id":"arxiv-2609.33439","arxiv":"2609.33439","title":"Raven: The Harness of Harnesses for Composable Agentic Intelligence","authors":["EverMind AI"],"authorCount":1,"org":"EverMind","affiliations":[],"orgNote":null,"date":"2026-09-27","hfDaily":"2026-09-30","primaryCategory":"cs.AI","categories":["cs.AI","cs.CL","cs.GT","cs.MA","cs.NE"],"links":{"abs":"https://arxiv.org/abs/2609.33439","pdf":"https://arxiv.org/pdf/2609.33439","hf":"https://huggingface.co/papers/2609.33439","code":"https://github.com/EverMind-AI/Raven","project":"https://raven.evermind.ai/"},"metrics":{"hfUpvotes":748,"hfComments":6,"citations":1,"citationsSource":"semantic_scholar","githubStars":5521},"hot":93.42,"tags":["agent"],"summary":{"zh":"开源多智能体生态 Raven，自动搭建、演化并跨领域编排各类专用 Agent 框架（harness）。","en":"Raven is an open multi-agent ecosystem that automatically builds, evolves and orchestrates specialised agent harnesses across domains."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":1},{"id":"arxiv-2609.24972","arxiv":"2609.24972","title":"RRSI: Regularized Recursive Self-Improvement of Agent Harnesses","authors":["Peng Xia","Rujun Han","Zifeng Wang","Yanfei Chen","Yufan Zhuang","Yoonho Lee","Chengsong Huang","Han Yu"],"authorCount":14,"org":"Google","affiliations":[],"orgNote":null,"date":"2026-09-21","hfDaily":"2026-09-22","primaryCategory":"cs.LG","categories":["cs.LG","cs.AI","cs.CL"],"links":{"abs":"https://arxiv.org/abs/2609.24972","pdf":"https://arxiv.org/pdf/2609.24972","hf":"https://huggingface.co/papers/2609.24972","code":"https://github.com/google-research/rrsi","project":"https://regularized-rsi.com/"},"metrics":{"hfUpvotes":220,"hfComments":3,"citations":22,"citationsSource":"semantic_scholar","githubStars":1563},"hot":90.99,"tags":["agent","benchmark"],"summary":{"zh":"给 Agent 框架的递归自我改进加上正则化，防止只在训练任务上涨分、换分布就失效。","en":"Adds regularisation to recursive self-improvement of agent harnesses so gains survive out-of-distribution tests."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":2},{"id":"arxiv-2609.33757","arxiv":"2609.33757","title":"YuE2: Unifying Symbolic and Audio Music Generation at Frontier Quality","authors":["Ruibin Yuan","Jiahao Pan","Junyan Jiang","Zhiyue Wu","Ziya Zhou","Jiankai Sun","Yizhi Li","Ge Zhang"],"authorCount":35,"org":"Multimodal Art Projection","affiliations":[],"orgNote":null,"date":"2026-09-27","hfDaily":"2026-09-29","primaryCategory":"eess.AS","categories":["eess.AS","cs.LG","cs.SD"],"links":{"abs":"https://arxiv.org/abs/2609.33757","pdf":"https://arxiv.org/pdf/2609.33757","hf":"https://huggingface.co/papers/2609.33757","code":"https://github.com/multimodal-art-projection/YuE","project":"https://map-yue2.github.io/"},"metrics":{"hfUpvotes":223,"hfComments":3,"citations":1,"citationsSource":"semantic_scholar","githubStars":12349},"hot":82.52,"tags":["multimodal","visual-gen"],"summary":{"zh":"YuE2 先写可读乐谱再生成整首歌，把符号作曲和音频生成统一到一个模型里。","en":"YuE2 plans a readable score first and then renders full songs, unifying symbolic and audio music generation."},"disciplines":["music"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":3},{"id":"arxiv-2610.14224","arxiv":"2610.14224","title":"Harness^2: Recursive Agent Harnessing for an Open World","authors":["Ruiyao Xu","Yanfei Chen","Zhongying CuiZhu","Bhavana Dalvi Mishra","Yifei Ming","Han Yu","Rujun Han","Chen-Yu Lee"],"authorCount":9,"org":"Google","affiliations":[],"orgNote":null,"date":"2026-10-09","hfDaily":"2026-10-14","primaryCategory":"cs.AI","categories":["cs.AI","cs.CL","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2610.14224","pdf":"https://arxiv.org/pdf/2610.14224","hf":"https://huggingface.co/papers/2610.14224","code":"https://github.com/google-research/harness2","project":"https://harness2.github.io/"},"metrics":{"hfUpvotes":115,"hfComments":1,"citations":1,"citationsSource":"semantic_scholar","githubStars":13969},"hot":79.21,"tags":["agent"],"summary":{"zh":"Harness² 让 Agent 在没有标准答案的开放任务里也能递归改进自己的执行框架。","en":"Harness² lets agents recursively improve their own harness on unseen open-world tasks without ground-truth verifiers."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":4},{"id":"arxiv-2610.18773","arxiv":"2610.18773","title":"Jeva: System One Models from First Principles","authors":["Qi Li","Xinchao Wang"],"authorCount":2,"org":"National University of Singapore","affiliations":[],"orgNote":null,"date":"2026-10-13","hfDaily":"2026-10-14","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.18773","pdf":"https://arxiv.org/pdf/2610.18773","hf":"https://huggingface.co/papers/2610.18773","code":"https://github.com/LiQiiiii/Jeva","project":"https://liqiiiii.github.io/Jeva/"},"metrics":{"hfUpvotes":371,"hfComments":4,"citations":2,"citationsSource":"semantic_scholar","githubStars":115},"hot":77.88,"tags":["llm","efficiency"],"summary":{"zh":"提出“决策原生”的 System 1 模型 Jeva：直接给候选项打分，而不是逐字生成答案，更快且可缓存。","en":"Jeva is a decision-native System 1 model that scores candidate options directly instead of generating text answers, making it fast and cacheable."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":5},{"id":"arxiv-2610.20584","arxiv":"2610.20584","title":"Player-1: Learning to Play from Synthetic Experience","authors":["Mingyu Ouyang","Yiran Wang","Siyuan Hu","Xingyilang Yin","Huiqiang Sun","Guangzhi Wang","Junfu Pu","Cheng Cheng"],"authorCount":12,"org":"ARC Lab, Tencent","affiliations":[],"orgNote":null,"date":"2026-10-14","hfDaily":"2026-10-15","primaryCategory":"cs.CV","categories":["cs.CV","cs.AI","cs.HC"],"links":{"abs":"https://arxiv.org/abs/2610.20584","pdf":"https://arxiv.org/pdf/2610.20584","hf":"https://huggingface.co/papers/2610.20584","code":"https://github.com/TencentARC/Player-1","project":"https://player-1-project.github.io/"},"metrics":{"hfUpvotes":348,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":187},"hot":77.72,"tags":["agent","rl"],"summary":{"zh":"Player-1 用合成交互经验做后训练，让电脑操作型 Agent 学会玩游戏。","en":"Player-1 post-trains computer-use agents on synthetic experience so they learn to control games beyond demonstrations."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":6},{"id":"arxiv-2609.36012","arxiv":"2609.36012","title":"In-Context Learning for Robots: Methods and Applications","authors":["Haojian Huang","Zexi Li","Junhao Guo","Yehang Zhang","Wenxuan Peng","Bohan Zhou","Weilin Ruan","Leyi Wu"],"authorCount":39,"org":"Knowin AI","affiliations":[],"orgNote":null,"date":"2026-09-28","hfDaily":"2026-09-30","primaryCategory":"cs.RO","categories":["cs.RO","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2609.36012","pdf":"https://arxiv.org/pdf/2609.36012","hf":"https://huggingface.co/papers/2609.36012","code":"https://github.com/JethroJames/awesome-robots-icl","project":"https://jethrojames.github.io/awesome-robots-icl/"},"metrics":{"hfUpvotes":388,"hfComments":2,"citations":2,"citationsSource":"semantic_scholar","githubStars":116},"hot":77.58,"tags":["robotics"],"summary":{"zh":"机器人上下文学习综述：按“上下文如何转成动作”把方法分成四大类并对比。","en":"A survey of in-context learning for robots, grouping methods into four families by how context is turned into action."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":7},{"id":"arxiv-2610.18615","arxiv":"2610.18615","title":"VideoJev: System One Model for Video Understanding and Interaction","authors":["Yuhan Zhu","Xiangyu Zeng","Qingyi Si","Chenxu Yang","Yifei Li","Luozheng Qin","Junhao Zhou","Dingyu Yao"],"authorCount":18,"org":"Multimedia Computing Group-Nanjing University","affiliations":[],"orgNote":null,"date":"2026-10-13","hfDaily":"2026-10-14","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.18615","pdf":"https://arxiv.org/pdf/2610.18615","hf":"https://huggingface.co/papers/2610.18615","code":"https://github.com/MCG-NJU/VideoJev","project":"https://mcg-nju.github.io/VideoJev/"},"metrics":{"hfUpvotes":375,"hfComments":3,"citations":1,"citationsSource":"semantic_scholar","githubStars":210},"hot":77.53,"tags":["multimodal","efficiency"],"summary":{"zh":"VideoJev 把视频理解改成对时间段和选项直接打分，统一时序定位和视频问答。","en":"VideoJev scores time spans and answer options directly, unifying temporal grounding and video QA for fast decisions."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":8},{"id":"arxiv-2609.31847","arxiv":"2609.31847","title":"Omni-IO Skills: Harnessing Your Agent Omni-Native","authors":["Yanlin Li","Mingyang Hao","Shengqiong Wu","Hao Fei","Mong-Li Lee","Wynne Hsu"],"authorCount":6,"org":"National University of Singapore","affiliations":[],"orgNote":null,"date":"2026-09-25","hfDaily":"2026-09-30","primaryCategory":"cs.CL","categories":["cs.CL"],"links":{"abs":"https://arxiv.org/abs/2609.31847","pdf":"https://arxiv.org/pdf/2609.31847","hf":"https://huggingface.co/papers/2609.31847","code":"https://github.com/any2any-mllm/Omni-IO-Skill","project":"https://github.com/any2any-mllm/Omni-IO-Skill"},"metrics":{"hfUpvotes":485,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":200},"hot":76.77,"tags":["agent","multimodal"],"summary":{"zh":"Omni-IO Skills 用分层技能和资产登记表，让现有 Agent 能读写图像、音频、视频、3D 等多种模态。","en":"Omni-IO Skills makes existing agents omni-modal through hierarchical skills, a common execution interface and an asset registry."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":9},{"id":"arxiv-2609.33325","arxiv":"2609.33325","title":"VisionHOPE: Visual Backbones as Self-Modifying Learning Systems","authors":["Siran Peng","Tianshuo Zhang","Tianyu Fu","Weisong Zhao","Haoyuan Zhang","Jiankuo Zhao","Minghui Wu","Ping Jiang"],"authorCount":11,"org":"Mininglamp Technology","affiliations":[],"orgNote":null,"date":"2026-09-27","hfDaily":"2026-09-29","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2609.33325","pdf":"https://arxiv.org/pdf/2609.33325","hf":"https://huggingface.co/papers/2609.33325","code":"https://github.com/PSRben/VisionHOPE","project":null},"metrics":{"hfUpvotes":296,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":1294},"hot":76.5,"tags":["vision"],"summary":{"zh":"VisionHOPE 把视觉骨干网络做成“自我修改”的学习系统，记什么、怎么学在单张图内共同演化。","en":"VisionHOPE turns a visual backbone into a self-modifying learner whose memory and update rules co-evolve within each image."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":10},{"id":"arxiv-2610.22471","arxiv":"2610.22471","title":"Has GPT-6-Astra Achieved Embodied Intelligence?","authors":[" SenseNova","Joint Research Team","Zhongang Cai","Ziang Cao","Tailai Chen","Han Chong Chua","Linjun Dai","Hanming Deng"],"authorCount":62,"org":"SenseNova","affiliations":[],"orgNote":null,"date":"2026-10-15","hfDaily":"2026-10-19","primaryCategory":"cs.RO","categories":["cs.RO","cs.CV","cs.LG","cs.MM"],"links":{"abs":"https://arxiv.org/abs/2610.22471","pdf":"https://arxiv.org/pdf/2610.22471","hf":"https://huggingface.co/papers/2610.22471","code":"https://github.com/OpenSenseNova/SenseNova-RoboRSI","project":"https://sensenova.ai/has-gpt-6-astra-achieved-embodied-intelligence"},"metrics":{"hfUpvotes":459,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":218},"hot":76.43,"tags":["robotics","benchmark"],"summary":{"zh":"在 12 个机器人基准上系统评测 GPT-6-Astra 的具身能力，并讨论通用大模型对机器人研究的影响。","en":"Evaluates GPT-6-Astra on 12 robotics benchmarks and discusses what a general model's embodied competence means for the field."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":11},{"id":"arxiv-2610.10528","arxiv":"2610.10528","title":"Long-WAM: Scaling the Context of World-Action Models","authors":["Wei Huang","Bohan Zhang","Chenzhi Liu","Isabella Liu","Shuai Yang","Weian Mao","Luozhou Wang","Yicheng Xiao"],"authorCount":16,"org":"NVIDIA","affiliations":[],"orgNote":null,"date":"2026-10-07","hfDaily":"2026-10-08","primaryCategory":"cs.RO","categories":["cs.RO","cs.AI","cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.10528","pdf":"https://arxiv.org/pdf/2610.10528","hf":"https://huggingface.co/papers/2610.10528","code":"https://github.com/NVlabs/LongLive","project":"https://nvlabs.github.io/LongLive/Long-WAM/"},"metrics":{"hfUpvotes":113,"hfComments":4,"citations":1,"citationsSource":"semantic_scholar","githubStars":2813},"hot":76.4,"tags":["robotics","efficiency"],"summary":{"zh":"Long-WAM 研究在实时控制约束下给世界-动作模型加长上下文，自回归预训练的视频底座收益最大。","en":"Long-WAM scales the context of world-action models under real-time limits and finds autoregressive video pretraining makes long history pay off."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":12},{"id":"arxiv-2609.38426","arxiv":"2609.38426","title":"LoopVL: Recurrent Visual Intelligence","authors":["Zhe Qian","Ziyang Gong","Zhongxing Xu","Hehan Li","Zhonghua Wang","Hanmeng Liu","Liwei Qian","Xin Pei"],"authorCount":16,"org":"Renmin University of China","affiliations":[],"orgNote":null,"date":"2026-09-29","hfDaily":"2026-10-01","primaryCategory":"cs.CV","categories":["cs.CV","cs.CL"],"links":{"abs":"https://arxiv.org/abs/2609.38426","pdf":"https://arxiv.org/pdf/2609.38426","hf":"https://huggingface.co/papers/2609.38426","code":"https://github.com/Tier-Flow/LoopVL","project":"https://huggingface.co/TierFlow/LoopVL"},"metrics":{"hfUpvotes":417,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":261},"hot":75.92,"tags":["multimodal","reasoning"],"summary":{"zh":"LoopVL 把循环 Transformer 扩展到视觉语言模型，在多次循环中反复更新视觉-语言状态。","en":"LoopVL brings looped (recurrent) Transformers to vision-language models, iteratively refining a shared vision-language state."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":13},{"id":"arxiv-2609.39601","arxiv":"2609.39601","title":"GroundingPI: A Grounding Foundation Model towards Physical Intelligence with Visual Primitives","authors":["Qize Yu","Lianrui Fan","Boyu Chen","Jiaqi Liang","Xini Ding","Yue Chen","Zetian Song","Yuran Wang"],"authorCount":26,"org":"XPENG AI","affiliations":[],"orgNote":null,"date":"2026-09-30","hfDaily":"2026-10-13","primaryCategory":"cs.CV","categories":["cs.CV","cs.AI","cs.LG","cs.RO"],"links":{"abs":"https://arxiv.org/abs/2609.39601","pdf":"https://arxiv.org/pdf/2609.39601","hf":"https://huggingface.co/papers/2609.39601","code":"https://github.com/groundingpi/GroundingPI","project":"https://groundingpi.github.io/"},"metrics":{"hfUpvotes":310,"hfComments":2,"citations":2,"citationsSource":"semantic_scholar","githubStars":140},"hot":75.81,"tags":["robotics","vision"],"summary":{"zh":"GroundingPI 是 4B 参数的定位基础模型，用坐标 token 输出点和框，为机器人控制提供快速精准的目标定位。","en":"GroundingPI is a 4B grounding foundation model that outputs points and boxes as tokens to give robots fast, precise targeting."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":14},{"id":"arxiv-2609.13814","arxiv":"2609.13814","title":"Realtime-Venus: A full-duplex interaction system with asynchronous delegation","authors":[" Ant Group"],"authorCount":1,"org":"inclusionAI","affiliations":[],"orgNote":null,"date":"2026-09-12","hfDaily":"2026-09-22","primaryCategory":"cs.CV","categories":["cs.CV","eess.AS"],"links":{"abs":"https://arxiv.org/abs/2609.13814","pdf":"https://arxiv.org/pdf/2609.13814","hf":"https://huggingface.co/papers/2609.13814","code":"https://github.com/inclusionAI/Realtime-Venus","project":"https://realtime-venus.github.io/"},"metrics":{"hfUpvotes":184,"hfComments":4,"citations":6,"citationsSource":"semantic_scholar","githubStars":200},"hot":75.66,"tags":["multimodal","agent"],"summary":{"zh":"Realtime-Venus 是全双工实时交互系统，可边听边看边说，并把复杂任务异步委派出去。","en":"Realtime-Venus is a full-duplex audio-visual interaction system that listens, watches and talks while delegating tasks asynchronously."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":15},{"id":"arxiv-2609.24984","arxiv":"2609.24984","title":"WorldCrafter: Consistent Video World Model with Implicit 3D-aware Memory","authors":["Wangbo Yu","Kunhao Liu","Wenbo Hu","Shenghai Yuan","Chaoran Feng","Haiyang Zhou","Yukun Huang","Yiran Wang"],"authorCount":11,"org":"ARC Lab, Tencent","affiliations":[],"orgNote":null,"date":"2026-09-21","hfDaily":"2026-09-22","primaryCategory":"cs.CV","categories":["cs.CV","cs.AI","cs.GR"],"links":{"abs":"https://arxiv.org/abs/2609.24984","pdf":"https://arxiv.org/pdf/2609.24984","hf":"https://huggingface.co/papers/2609.24984","code":"https://github.com/TencentARC/WorldCrafter","project":"https://drexubery.github.io/WorldCrafter"},"metrics":{"hfUpvotes":161,"hfComments":4,"citations":1,"citationsSource":"semantic_scholar","githubStars":512},"hot":75.66,"tags":["visual-gen"],"summary":{"zh":"WorldCrafter 给视频世界模型加入可按相机视角查询的隐式 3D 记忆，长时间探索也保持一致。","en":"WorldCrafter adds a camera-queryable implicit 3D memory to a video world model to keep long explorations consistent."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":16},{"id":"arxiv-2610.01762","arxiv":"2610.01762","title":"OneStreamer: Unifying Perception, Memory, and Proactive Response in Streaming Video Interaction","authors":["Xiangyu Zeng","Yuandong Yang","Zhiqiu Zhang","Yuhan Zhu","Xinhao Li","Qingyi Si","Dingyu Yao","Changlian Ma"],"authorCount":24,"org":"Nanjing University","affiliations":[],"orgNote":null,"date":"2026-10-01","hfDaily":"2026-10-02","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.01762","pdf":"https://arxiv.org/pdf/2610.01762","hf":"https://huggingface.co/papers/2610.01762","code":"https://github.com/MCG-NJU/OneStreamer","project":"https://mcg-nju.github.io/OneStreamer"},"metrics":{"hfUpvotes":238,"hfComments":2,"citations":3,"citationsSource":"semantic_scholar","githubStars":228},"hot":75.54,"tags":["multimodal"],"summary":{"zh":"OneStreamer 让流式视频模型边看边记分层字幕记忆，并在证据足够时主动回应。","en":"OneStreamer lets streaming video LLMs keep hierarchical caption memories and respond proactively once evidence is sufficient."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":17},{"id":"arxiv-2610.20571","arxiv":"2610.20571","title":"Any Time, Any Space: A World-State-Centric Framework for Agentic Video Generation","authors":["Weijie Wang","Haomin Zhang","Ye Li","Akide Liu","Zeyu Zhang","Haoyu Zhao","Yefei He","Donny Y. Chen"],"authorCount":10,"org":"Zhejiang University","affiliations":[],"orgNote":null,"date":"2026-10-14","hfDaily":"2026-10-15","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.20571","pdf":"https://arxiv.org/pdf/2610.20571","hf":"https://huggingface.co/papers/2610.20571","code":"https://github.com/ziplab/ATAS","project":"https://lhmd.top/atas/"},"metrics":{"hfUpvotes":364,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":312},"hot":75.44,"tags":["visual-gen","agent"],"summary":{"zh":"ATAS 用场景记忆和事件记忆把剧本编译成镜头，让长视频在时空上前后一致。","en":"ATAS compiles screenplays into shots backed by scene and event memories so long agentic videos stay spatially and temporally consistent."},"disciplines":["film-animation"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":18},{"id":"arxiv-2610.22200","arxiv":"2610.22200","title":"OpenGCUA: Toward a Fully Open Stack for Generalist Computer Use","authors":["Hanzhang Zhou","Long Li","Quyu Kong","Lei Wang","Panrong Tong","Long Chen","Tianyu Xia","Gongjie Zhang"],"authorCount":15,"org":"Tongyi-MAI","affiliations":[],"orgNote":null,"date":"2026-10-15","hfDaily":"2026-10-16","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.22200","pdf":"https://arxiv.org/pdf/2610.22200","hf":"https://huggingface.co/papers/2610.22200","code":"https://github.com/Tongyi-MAI/OpenGCUA","project":"https://tongyi-mai.github.io/OpenGCUA/"},"metrics":{"hfUpvotes":330,"hfComments":4,"citations":0,"citationsSource":"semantic_scholar","githubStars":396},"hot":75.4,"tags":["agent","code"],"summary":{"zh":"OpenGCUA 计划开源通用电脑操作 Agent 的完整技术栈，首版同时支持视觉界面和代码执行。","en":"OpenGCUA aims to open the full stack for generalist computer-use agents, with a first release spanning GUI and code execution."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":19},{"id":"arxiv-2610.12374","arxiv":"2610.12374","title":"AgentGarten: Code Worlds for Evolving Agents","authors":["Jiawei Chi","Shangchen Miao","Zhiyuan Shi","Kailu Wu","Hanyang Wang","Weiliang Chen","Qiyu Dai","Chengyang Li"],"authorCount":15,"org":"MirroS","affiliations":[],"orgNote":null,"date":"2026-10-08","hfDaily":"2026-10-09","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.12374","pdf":"https://arxiv.org/pdf/2610.12374","hf":"https://huggingface.co/papers/2610.12374","code":"https://github.com/MirroS-Lab/AgentGarten","project":"https://mirros-lab.github.io/agent-garten/"},"metrics":{"hfUpvotes":293,"hfComments":4,"citations":1,"citationsSource":"semantic_scholar","githubStars":201},"hot":75.29,"tags":["agent","visual-gen"],"summary":{"zh":"AgentGarten 把模拟器、游戏引擎和神经渲染器结合起来，为 Agent 搭建逼真且规则一致的代码世界。","en":"AgentGarten couples simulators and game engines with a neural renderer to build faithful, realistic worlds for training agents."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":20},{"id":"arxiv-2609.23088","arxiv":"2609.23088","title":"OmniEdu: Open Foundation Models for Learning and Teaching","authors":["Hao Liang","Qihan Lin","Meiyi Qiang","Linzhuang Sun","Hengyi Feng","Mingrui Chen","Sizhe Qiu","Wentao Zhang"],"authorCount":8,"org":"Peking University","affiliations":[],"orgNote":null,"date":"2026-09-19","hfDaily":"2026-09-22","primaryCategory":"cs.CL","categories":["cs.CL"],"links":{"abs":"https://arxiv.org/abs/2609.23088","pdf":"https://arxiv.org/pdf/2609.23088","hf":"https://huggingface.co/papers/2609.23088","code":"https://github.com/haolpku/Omni-Edu","project":"https://haolpku.github.io/Omni-Edu/"},"metrics":{"hfUpvotes":240,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":227},"hot":74.5,"tags":["llm"],"summary":{"zh":"OmniEdu 是面向中小学教与学的开源基础模型家族，覆盖解题、课程对齐、学情诊断和教学引导。","en":"OmniEdu is an open family of K-12 education foundation models covering problem solving, curriculum grounding, diagnosis and tutoring."},"disciplines":["education"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":21},{"id":"arxiv-2609.36484","arxiv":"2609.36484","title":"The Teacher Is a Direction, Not a Destination: Extrapolating RL-Induced Representation Residuals in On-Policy Distillation","authors":["Hao Li","MeiJia Chen","Weijie Ren","Donghan Li","Zijun Tian","Jingchun Huang","Naibo Wang"],"authorCount":7,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-09-29","hfDaily":"2026-10-01","primaryCategory":"cs.LG","categories":["cs.LG"],"links":{"abs":"https://arxiv.org/abs/2609.36484","pdf":"https://arxiv.org/pdf/2609.36484","hf":"https://huggingface.co/papers/2609.36484","code":"https://github.com/xixixixixxxx/RIDE","project":null},"metrics":{"hfUpvotes":511,"hfComments":4,"citations":2,"citationsSource":"semantic_scholar","githubStars":7},"hot":74.39,"tags":["efficiency","rl"],"summary":{"zh":"在线策略蒸馏中，沿 RL 引起的隐藏表征变化方向外推，让学生模型更稳定地超越教师。","en":"Extrapolates RL-induced hidden-representation residuals in on-policy distillation so students can surpass teachers more stably."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":22},{"id":"arxiv-2610.22350","arxiv":"2610.22350","title":"FDE-Gym: The Last Mile of AI Delivery","authors":["Dongxu Yang","Hongda Yan","Renxin Wang","Baoli Zhang","Haojia Zuo","Ying Zhao","Youmin Chen","Xiaojian Liao"],"authorCount":26,"org":"DeepLethe","affiliations":[],"orgNote":null,"date":"2026-10-15","hfDaily":"2026-10-16","primaryCategory":"cs.AI","categories":["cs.AI","cs.CL","cs.SE"],"links":{"abs":"https://arxiv.org/abs/2610.22350","pdf":"https://arxiv.org/pdf/2610.22350","hf":"https://huggingface.co/papers/2610.22350","code":"https://github.com/deeplethe/fde-gym","project":"https://fde-gym.com/paper"},"metrics":{"hfUpvotes":272,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":107},"hot":73.45,"tags":["agent","benchmark"],"summary":{"zh":"FDE-Gym 收录 95 个真实交付任务，考察 Agent 能否像驻场工程师一样澄清需求并交付可用方案。","en":"FDE-Gym offers 95 real delivery tasks testing whether agents can clarify vague requirements and ship working solutions like forward-deployed engineers."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":23},{"id":"arxiv-2610.12572","arxiv":"2610.12572","title":"DiffusionART: Efficient, Stable and Scalable Reinforcement of Diffusion Models","authors":["Zhenglin Wan","Jiayu Chen","Zhengfeiyang Zhang","Pengfei Zhou","Wangbo Zhao","Fangyuan Kong","Pengfei Wan","Kun Gai"],"authorCount":9,"org":"KlingAI Research","affiliations":[],"orgNote":null,"date":"2026-10-08","hfDaily":"2026-10-12","primaryCategory":"cs.LG","categories":["cs.LG"],"links":{"abs":"https://arxiv.org/abs/2610.12572","pdf":"https://arxiv.org/pdf/2610.12572","hf":"https://huggingface.co/papers/2610.12572","code":"https://github.com/vanzll/DiffusionART","project":"https://vanzll.github.io/diffusionart-project/"},"metrics":{"hfUpvotes":325,"hfComments":4,"citations":0,"citationsSource":"semantic_scholar","githubStars":154},"hot":72.89,"tags":["visual-gen","rl"],"summary":{"zh":"DiffusionART 先聚合奖励再微调扩散模型，在线强化学习更高效也更稳定。","en":"DiffusionART aggregates rewards before tilting the objective, making online RL fine-tuning of diffusion models efficient and stable."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":24},{"id":"arxiv-2610.20424","arxiv":"2610.20424","title":"BizAgentBench: Unveiling Hidden Challenges in Deploying Long-Horizon Agents for Real-World Business","authors":["Enhan Zhao","Yongqi Tong","Xuejun Yu","Hang Wang","Pan Wang","Yichao Wang","Di He","Jianshe Li"],"authorCount":9,"org":"ant-international","affiliations":[],"orgNote":null,"date":"2026-10-14","hfDaily":"2026-10-19","primaryCategory":"cs.CE","categories":["cs.CE"],"links":{"abs":"https://arxiv.org/abs/2610.20424","pdf":"https://arxiv.org/pdf/2610.20424","hf":"https://huggingface.co/papers/2610.20424","code":"https://github.com/ant-intl/BizAgentBench","project":null},"metrics":{"hfUpvotes":490,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":30},"hot":72.63,"tags":["agent","benchmark"],"summary":{"zh":"BizBench 基于真实支付业务流程，检验长任务 Agent 在执行中途是否违反政策与约束。","en":"BizBench, built on real payment procedures, checks whether long-horizon agents violate policies and constraints mid-execution."},"disciplines":["finance"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":25},{"id":"arxiv-2610.21046","arxiv":"2610.21046","title":"PremiseCheck: Efficient Fact-Checking of Human and AI Peer Reviews","authors":["Dingjie Song","Zhiling Yan","Ziqian Zhang","Yuxuan Zhang","Lichao Sun"],"authorCount":5,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-10-15","hfDaily":"2026-10-19","primaryCategory":"cs.CL","categories":["cs.CL","cs.AI","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2610.21046","pdf":"https://arxiv.org/pdf/2610.21046","hf":"https://huggingface.co/papers/2610.21046","code":"https://github.com/OpenLAIR/PremiseCheck","project":"https://openlair.github.io/premisecheck/"},"metrics":{"hfUpvotes":727,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":5},"hot":72.46,"tags":["benchmark","efficiency"],"summary":{"zh":"ReviewFact 标注审稿意见中的事实错误，PremiseCheck 结合大模型和轻量判别模型低成本核查。","en":"ReviewFact labels factual errors in peer reviews; PremiseCheck combines an LLM with cheap decision models to verify them efficiently."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":26},{"id":"arxiv-2610.22276","arxiv":"2610.22276","title":"MJev-Spatial: System One Spatial Intelligence via Native Attention Matching","authors":["Luozheng Qin","Yuhan Zhu","Changhai Zhou","Qian Qiao","Jun Gao","Yuhua Zhou","Harry Zhang","Xiangyu Zeng"],"authorCount":12,"org":"Mind Lab","affiliations":[],"orgNote":null,"date":"2026-10-15","hfDaily":"2026-10-16","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.22276","pdf":"https://arxiv.org/pdf/2610.22276","hf":"https://huggingface.co/papers/2610.22276","code":"https://github.com/Fr0zenCrane/MJev-Spatial","project":"https://fr0zencrane.github.io/mjev-spatial-project/"},"metrics":{"hfUpvotes":560,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":17},"hot":72.17,"tags":["multimodal","efficiency"],"summary":{"zh":"MJev-Spatial 把空间理解任务改成分类式打分，比自由生成文本更快地回答方位、距离和落点问题。","en":"MJev-Spatial casts spatial questions as classification over candidates, answering relation, distance and point queries faster than text generation."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":27},{"id":"arxiv-2609.34309","arxiv":"2609.34309","title":"MaLiang-Harness: A Programmable Path to Image and Video Generation","authors":["Haoyu Zhao","Zihao Zhang","Xudong Wang","Jiaxi Gu","Zuxuan Wu","Yu-Gang Jiang","Shuicheng Yan"],"authorCount":7,"org":"National University of Singapore","affiliations":[],"orgNote":null,"date":"2026-09-28","hfDaily":"2026-09-30","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2609.34309","pdf":"https://arxiv.org/pdf/2609.34309","hf":"https://huggingface.co/papers/2609.34309","code":"https://github.com/gulucaptain/MaLiang-Harness","project":"https://gulucaptain.github.io/MaLiang-Harness/"},"metrics":{"hfUpvotes":367,"hfComments":1,"citations":1,"citationsSource":"semantic_scholar","githubStars":36},"hot":71.92,"tags":["visual-gen","code"],"summary":{"zh":"MaLiang-Harness 让多模态模型写程序生成图像和视频，并在“构建-检查-修改”循环中缩小程序与画面的差距。","en":"MaLiang-Harness has MLLMs generate images and videos via programs in a build-inspect-revise loop that closes the program-to-visual gap."},"disciplines":["film-animation"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":28},{"id":"arxiv-2609.39102","arxiv":"2609.39102","title":"False Frontiers: Diagnosing and Mitigating Co-Cheating in Self-Evolving Search Agents","authors":["Meijia Chen","Hao Li","Zheng Lu","Hongshan Lin","Junbai Tian","Yichen Liu","Zijun Tian","Yufan Zou"],"authorCount":15,"org":"Rutgers University","affiliations":[],"orgNote":null,"date":"2026-09-30","hfDaily":"2026-10-01","primaryCategory":"cs.CL","categories":["cs.CL","cs.AI","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2609.39102","pdf":"https://arxiv.org/pdf/2609.39102","hf":"https://huggingface.co/papers/2609.39102","code":null,"project":null},"metrics":{"hfUpvotes":642,"hfComments":4,"citations":2,"citationsSource":"semantic_scholar","githubStars":null},"hot":71.47,"tags":["agent","rl"],"summary":{"zh":"揭示自进化搜索 Agent 的“共谋作弊”：出题者和解题者一起犯同样的错，并提出多样本验证来缓解。","en":"Identifies co-cheating in self-evolving search agents, where proposer and solver converge on shared errors, and mitigates it with multi-sample verification."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":29},{"id":"arxiv-2610.20951","arxiv":"2610.20951","title":"LEV: Thinking Fast and Deep with Looped Decision Models","authors":["Dingjie Song","Zhiling Yan","Yuxuan Zhang","Lichao Sun"],"authorCount":4,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-10-14","hfDaily":"2026-10-19","primaryCategory":"cs.AI","categories":["cs.AI","cs.CL","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2610.20951","pdf":"https://arxiv.org/pdf/2610.20951","hf":"https://huggingface.co/papers/2610.20951","code":"https://github.com/OpenLAIR/lev","project":"https://openlair.github.io/lev/"},"metrics":{"hfUpvotes":602,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":7},"hot":71.3,"tags":["reasoning","efficiency"],"summary":{"zh":"LEV 让决策模型通过循环重跑网络层来“想得更深”，不必生成推理文字。","en":"LEV lets decision models think deeper by looping their layers instead of generating reasoning tokens."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":30},{"id":"arxiv-2610.18813","arxiv":"2610.18813","title":"Self-Evolving Agent Harness","authors":["Chuyang Wei","Haoxiang Guan","Minghang Zhu","Kexin Zhou","Jiyan He","Xiren Zhou","Huanhuan Chen","Yifei Shen"],"authorCount":8,"org":"ZGCAGI","affiliations":[],"orgNote":null,"date":"2026-10-13","hfDaily":"2026-10-16","primaryCategory":"cs.AI","categories":["cs.AI","cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.18813","pdf":"https://arxiv.org/pdf/2610.18813","hf":"https://huggingface.co/papers/2610.18813","code":"https://github.com/zgcagi/Harness-RSI","project":"https://zgcagi.github.io/Harness-RSI/"},"metrics":{"hfUpvotes":217,"hfComments":7,"citations":0,"citationsSource":"semantic_scholar","githubStars":297},"hot":71.21,"tags":["agent"],"summary":{"zh":"给 Agent 框架的自我演化建立形式化理论，并给出泛化和保持性的理论界。","en":"Formalises self-evolving agent harnesses and derives generalisation and retention bounds for the process."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":31},{"id":"arxiv-2610.23479","arxiv":"2610.23479","title":"HexaShell: A Thin Harness for Exploring the Limits of VLM Agents as Robot Coding Agents","authors":["Hongcheng Gao","Alex Zhou","Jay Zhu","Zelin Zheng","Shijia Ge","Xiping Li","Di Wu","Yazhe Wang"],"authorCount":16,"org":"HexaFuture","affiliations":[],"orgNote":null,"date":"2026-10-16","hfDaily":"2026-10-19","primaryCategory":"cs.RO","categories":["cs.RO"],"links":{"abs":"https://arxiv.org/abs/2610.23479","pdf":"https://arxiv.org/pdf/2610.23479","hf":"https://huggingface.co/papers/2610.23479","code":"https://github.com/HexaFuture/HexaShell","project":"https://hexafuture.ai/home/"},"metrics":{"hfUpvotes":313,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":84},"hot":70.68,"tags":["robotics","agent","code"],"summary":{"zh":"HexaShell 只用一层很薄的框架，让编程 Agent 通过命令行直接控制双臂机器人，并在失败后自己写工具改进。","en":"HexaShell keeps the harness thin so a coding agent drives a two-arm robot via a command line and writes its own tools after failures."},"disciplines":["software-engineering","robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":32},{"id":"arxiv-2610.13527","arxiv":"2610.13527","title":"S-Space: Exploring Spatial Workspace in Multimodal Language Models","authors":["Zhiyuan Shi","Xiang Li","Xingzhuo Guo","Shangchen Miao","Weiliang Chen","Qiyu Dai","Jinshan Ren","Jiangran Lyu"],"authorCount":11,"org":"MirroS","affiliations":[],"orgNote":null,"date":"2026-10-09","hfDaily":"2026-10-12","primaryCategory":"cs.CV","categories":["cs.CV","cs.CL","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2610.13527","pdf":"https://arxiv.org/pdf/2610.13527","hf":"https://huggingface.co/papers/2610.13527","code":"https://github.com/mirros-lab/s-space","project":"https://mirros.ai/blog/s-space"},"metrics":{"hfUpvotes":270,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":175},"hot":70.6,"tags":["multimodal","safety"],"summary":{"zh":"发现多模态模型内部存在线性“空间工作区”S-Space，用连续坐标编码物体位置。","en":"Finds S-Space, a linear subspace inside MLLMs where object tokens encode continuous coordinates, acting as an internal spatial workspace."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":33},{"id":"arxiv-2609.25804","arxiv":"2609.25804","title":"The Tasteful Agent: Measuring and Improving Taste in Long-Horizon Tasks","authors":["Wenbo Pan","Zhichao Liu","Shujie Liu","Jingying Zeng","Chin-Yew Lin","Xianfeng Tang","Yan Lu","Qi He"],"authorCount":9,"org":"Microsoft","affiliations":[],"orgNote":null,"date":"2026-09-22","hfDaily":"2026-09-23","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2609.25804","pdf":"https://arxiv.org/pdf/2609.25804","hf":"https://huggingface.co/papers/2609.25804","code":"https://github.com/wbopan/tastebench","project":null},"metrics":{"hfUpvotes":163,"hfComments":6,"citations":1,"citationsSource":"semantic_scholar","githubStars":36},"hot":69.72,"tags":["agent","benchmark"],"summary":{"zh":"Taste-Bench 衡量 Agent 在长任务里选方向的“品味”，并探索如何提升这种决策能力。","en":"Taste-Bench measures an agent's 'taste' for long-horizon decisions and explores ways to improve it."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":34},{"id":"arxiv-2609.29233","arxiv":"2609.29233","title":"Post-Training Leaves Behavioral Shadows on Unrelated Decisions","authors":["Ziyang Zhang","Yubin Jing","Yuanhao Zeng","Yuyao Li","Haofan Wang","Yichen Gong"],"authorCount":6,"org":"Peking University","affiliations":[],"orgNote":null,"date":"2026-09-24","hfDaily":"2026-09-29","primaryCategory":"cs.CL","categories":["cs.CL","cs.AI","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2609.29233","pdf":"https://arxiv.org/pdf/2609.29233","hf":"https://huggingface.co/papers/2609.29233","code":"https://github.com/myboker/ATD","project":null},"metrics":{"hfUpvotes":273,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":18},"hot":69.57,"tags":["llm","safety"],"summary":{"zh":"发现后训练会在无关输出上留下“行为影子”，每个提示只需教师一个词就能把能力迁移给学生。","en":"Shows post-training leaves behavioural shadows in unrelated outputs; a single teacher word per prompt can transfer capabilities to a student."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":35},{"id":"arxiv-2609.34563","arxiv":"2609.34563","title":"Rethinking Latent Visual Reasoning: Grounding Latent Reasoning in Visual Evidence","authors":["Xi Xiao","Tianchen Zhao","Youngeun Kim","Zhuowei Li","Linghan Xu","Jiaye Wu","Zheng Zhang","Xiang Xu"],"authorCount":13,"org":"Amazon","affiliations":[],"orgNote":null,"date":"2026-09-28","hfDaily":"2026-10-01","primaryCategory":"cs.CV","categories":["cs.CV","cs.CL","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2609.34563","pdf":"https://arxiv.org/pdf/2609.34563","hf":"https://huggingface.co/papers/2609.34563","code":"https://github.com/xixiaouab/ReaLVR-code","project":"https://xixiaouab.github.io/projects/ReaLVR/"},"metrics":{"hfUpvotes":224,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":47},"hot":69.49,"tags":["reasoning","multimodal"],"summary":{"zh":"指出潜空间视觉推理的 token 对图像证据不敏感，并提出把潜推理锚定到视觉证据上的训练方法。","en":"Shows latent visual reasoning tokens barely respond to image evidence and proposes grounding them in visual evidence during training."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":36},{"id":"arxiv-2610.17953","arxiv":"2610.17953","title":"Agent-Reflex: Teaching Language Agents to Act Reflectively","authors":["Guanting Dong","Xinyu Yang","Jiajie Jin","Yuyang Hu","Xiaoshuai Song","Yifei Chen","Xiaoxi Li","Jiejun Tan"],"authorCount":12,"org":"NLPIR Lab @ RUC","affiliations":[],"orgNote":null,"date":"2026-10-13","hfDaily":"2026-10-14","primaryCategory":"cs.CL","categories":["cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.17953","pdf":"https://arxiv.org/pdf/2610.17953","hf":"https://huggingface.co/papers/2610.17953","code":"https://github.com/RUC-NLPIR/Agent-Reflex","project":null},"metrics":{"hfUpvotes":214,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":234},"hot":69.43,"tags":["agent","rl"],"summary":{"zh":"Agent-Reflex 分三阶段训练，让 Agent 在工具反馈有噪声或用户纠错时主动反思并恢复。","en":"Agent-Reflex trains agents in three stages to reflect proactively and recover when tool feedback is noisy or users intervene."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":37},{"id":"arxiv-2609.37959","arxiv":"2609.37959","title":"TabFM: A Zero-Shot Foundation Model for Tabular Data","authors":["Weihao Kong","Erez Louidor Ilan","Shuxin Nie","Taman Narayan","Rajat Sen","Yichen Zhou","Deqing Fu","Samet Oymak"],"authorCount":9,"org":"Google","affiliations":[],"orgNote":null,"date":"2026-09-29","hfDaily":"2026-09-30","primaryCategory":"cs.LG","categories":["cs.LG"],"links":{"abs":"https://arxiv.org/abs/2609.37959","pdf":"https://arxiv.org/pdf/2609.37959","hf":"https://huggingface.co/papers/2609.37959","code":"https://github.com/google-research/tabfm","project":"https://research.google/blog/introducing-tabfm-a-zero-shot-foundation-model-for-tabular-data/"},"metrics":{"hfUpvotes":21,"hfComments":2,"citations":22,"citationsSource":"semantic_scholar","githubStars":2789},"hot":68.93,"tags":["llm","benchmark"],"summary":{"zh":"TabFM 是 4 亿参数的表格基础模型，全靠合成数据训练，单次前向即可零样本预测。","en":"TabFM is a 400M-parameter tabular foundation model trained on synthetic tables that predicts zero-shot in a single forward pass."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":38},{"id":"arxiv-2610.05608","arxiv":"2610.05608","title":"Kandinsky 6.0 Video: Foundation Models for Synchronized Video and Audio Generation","authors":["Team Kandinsky","Julia Agafonova","Bulat Akhmatov","Mikhail Aksyutin","Grigorii Alekseenko","Anastasia Aliaskina","Olga Androsova","Vladimir Arkhipkin"],"authorCount":88,"org":"Kandinsky Lab","affiliations":[],"orgNote":null,"date":"2026-10-04","hfDaily":"2026-10-06","primaryCategory":"cs.CV","categories":["cs.CV","cs.AI","cs.LG","cs.MM"],"links":{"abs":"https://arxiv.org/abs/2610.05608","pdf":"https://arxiv.org/pdf/2610.05608","hf":"https://huggingface.co/papers/2610.05608","code":"https://github.com/kandinskylab/kandinsky-6","project":"https://kandinskylab.ai/"},"metrics":{"hfUpvotes":174,"hfComments":4,"citations":0,"citationsSource":"semantic_scholar","githubStars":364},"hot":68.81,"tags":["visual-gen","multimodal"],"summary":{"zh":"Kandinsky 6.0 Video 能同步生成 5 秒视频和 44kHz 音频（含口型），分 3B 和 29B 两个版本。","en":"Kandinsky 6.0 Video generates 5-second clips with synchronised 44 kHz audio and lip-sync, in 3B and 29B variants."},"disciplines":["film-animation"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":39},{"id":"arxiv-2609.28654","arxiv":"2609.28654","title":"Training Object Permanence in World Models","authors":["Haotian Zhang","Fengyuan Yu","Dezhi Luo","Haoran Sun","Zehong Zhao","Qingying Gao","Yihan Li","Siyuan An"],"authorCount":31,"org":"Carnegie Mellon University","affiliations":[],"orgNote":null,"date":"2026-09-23","hfDaily":"2026-09-25","primaryCategory":"cs.AI","categories":["cs.AI","cs.CV"],"links":{"abs":"https://arxiv.org/abs/2609.28654","pdf":"https://arxiv.org/pdf/2609.28654","hf":"https://huggingface.co/papers/2609.28654","code":"https://github.com/hokindeng/object-permanence","project":"https://www.object-permanence.world/"},"metrics":{"hfUpvotes":167,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":97},"hot":68.35,"tags":["visual-gen","reasoning"],"summary":{"zh":"构建 150 个认知科学启发任务，检验并训练视频世界模型的“物体恒存”常识。","en":"Builds 150 cognitive-science tasks to test and train object permanence in video world models."},"disciplines":["film-animation"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":40},{"id":"arxiv-2610.11287","arxiv":"2610.11287","title":"REMORY: Learning Residual Memory for Context Compaction","authors":["Hanchen Xia","Baoyou Chen","Yutang Ge","Naihao Deng","Senqiao Yang","Zilong Dong","Weihao Yuan","Siyu Zhu"],"authorCount":8,"org":"Fusion Lab: Generative Vision Lab of Fudan University and SII","affiliations":[],"orgNote":null,"date":"2026-10-08","hfDaily":"2026-10-09","primaryCategory":"cs.CL","categories":["cs.CL","cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.11287","pdf":"https://arxiv.org/pdf/2610.11287","hf":"https://huggingface.co/papers/2610.11287","code":"https://github.com/1ring2rta/Remory","project":"https://huggingface.co/mocoV3/Qwen3.8-27B-REMORY-1.9B"},"metrics":{"hfUpvotes":167,"hfComments":5,"citations":0,"citationsSource":"semantic_scholar","githubStars":303},"hot":68.22,"tags":["agent","efficiency"],"summary":{"zh":"REMORY 在文字摘要后补一串“软记忆 token”，帮长任务 Agent 压缩上下文时少丢信息。","en":"REMORY appends learned soft memory tokens after a summary so long-horizon agents lose less when compacting context."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":41},{"id":"arxiv-2610.08699","arxiv":"2610.08699","title":"nanoMuse: An Open-Source Personal Agent for Every Device You Own","authors":["Guangyi Liu","Yong Liu","Jiangning Zhang"],"authorCount":3,"org":"Zhejiang University","affiliations":[],"orgNote":null,"date":"2026-10-06","hfDaily":"2026-10-08","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.08699","pdf":"https://arxiv.org/pdf/2610.08699","hf":"https://huggingface.co/papers/2610.08699","code":"https://github.com/nano-muse/nanoMuse","project":"https://nanomuse.cn/"},"metrics":{"hfUpvotes":116,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":1551},"hot":68.07,"tags":["agent"],"summary":{"zh":"nanoMuse 是开源的个人 Agent，可跨设备记住用户、代办账户事务，定位为 Meta Muse 的开源对照。","en":"nanoMuse is an open-source personal agent that spans a user's devices and accounts, positioned as an open counterpart to Meta's Muse."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":42},{"id":"arxiv-2610.11765","arxiv":"2610.11765","title":"Thinking Inertia: LLMs Keep Thinking When Told Not To","authors":["Dianqiao Lei","Kevin Qinghong Lin","Pan Lu","Philip Torr","James Zou"],"authorCount":5,"org":"Stanford University","affiliations":[],"orgNote":null,"date":"2026-10-08","hfDaily":"2026-10-13","primaryCategory":"cs.CL","categories":["cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.11765","pdf":"https://arxiv.org/pdf/2610.11765","hf":"https://huggingface.co/papers/2610.11765","code":"https://github.com/thinking-inertia/code","project":"https://thinking-inertia.github.io/"},"metrics":{"hfUpvotes":241,"hfComments":6,"citations":0,"citationsSource":"semantic_scholar","githubStars":9},"hot":67.56,"tags":["reasoning","llm"],"summary":{"zh":"发现大模型即使被要求“不思考”仍会偷偷推理，并提出衡量“无思考”行为的三级指标。","en":"Shows LLMs keep reasoning even when told not to think, and proposes a three-level way to measure no-thinking behaviour."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":43},{"id":"arxiv-2610.09823","arxiv":"2610.09823","title":"UltraText Bench: A Comprehensive Bilingual Benchmark for Evaluating Visual Text Rendering in Image Generation","authors":["Deyuan Liu","Yihao Hu","Jingxuan Zhang","Xingying Li","Jun Xie","Jiacheng Liu","Jungang Li","Yu Huang"],"authorCount":16,"org":"Westlake University","affiliations":[],"orgNote":null,"date":"2026-10-07","hfDaily":"2026-10-08","primaryCategory":"cs.CV","categories":["cs.CV","cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.09823","pdf":"https://arxiv.org/pdf/2610.09823","hf":"https://huggingface.co/papers/2610.09823","code":"https://github.com/LINs-lab/UltraText_Bench","project":null},"metrics":{"hfUpvotes":197,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":123},"hot":67.01,"tags":["visual-gen","benchmark"],"summary":{"zh":"UltraText Bench 用 432 条中英双语提示，评测图像生成模型渲染密集文字的能力。","en":"UltraText Bench uses 432 bilingual prompts to evaluate dense visual text rendering in image generators."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":44},{"id":"arxiv-2609.34981","arxiv":"2609.34981","title":"What Makes World Action Models Generalize? An Empirical Study of Test-Time Future Modeling","authors":["Renping Zhou","Zanlin Ni","Zihao Fan","Guohao Fu","Zeyu Liu","Hao Shi","Jie Zhang","Chi Bene Chen"],"authorCount":11,"org":"Tsinghua-LeapLab","affiliations":[],"orgNote":null,"date":"2026-09-28","hfDaily":"2026-09-30","primaryCategory":"cs.CV","categories":["cs.CV","cs.RO"],"links":{"abs":"https://arxiv.org/abs/2609.34981","pdf":"https://arxiv.org/pdf/2609.34981","hf":"https://huggingface.co/papers/2609.34981","code":"https://github.com/LeapLabTHU/Simple-WAM","project":"https://zrporz.github.io/Simple-WAM-Web/"},"metrics":{"hfUpvotes":140,"hfComments":4,"citations":0,"citationsSource":"semantic_scholar","githubStars":82},"hot":66.95,"tags":["robotics"],"summary":{"zh":"实证发现：只在潜空间预测未来的世界-动作模型会丢掉泛化优势，显式生成未来画面仍有价值。","en":"Finds latent world-action models lose the generalisation benefits of explicitly generating future frames at test time."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":45},{"id":"arxiv-2610.18860","arxiv":"2610.18860","title":"AesCode: Aesthetic Code Generation with Decoupled Cross-Modal Rewards","authors":["Xin Gao","Yifan Yang","Xin Li","Pan Wang","Yan Li","Muzhao Tian","Mingxi Cheng","Bei Liu"],"authorCount":16,"org":"Microsoft","affiliations":[],"orgNote":null,"date":"2026-10-13","hfDaily":"2026-10-14","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.18860","pdf":"https://arxiv.org/pdf/2610.18860","hf":"https://huggingface.co/papers/2610.18860","code":"https://github.com/microsoft/AesCode","project":"https://microsoft.github.io/AesCode/"},"metrics":{"hfUpvotes":137,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":114},"hot":66.78,"tags":["code","visual-gen","rl"],"summary":{"zh":"AesCode 以同提示生成的图片做审美参考，用解耦的跨模态奖励训练模型写出更好看的幻灯片和海报代码。","en":"AesCode uses a generated image as aesthetic guidance and decoupled cross-modal rewards to produce better-looking slides and posters as code."},"disciplines":["software-engineering"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":46},{"id":"arxiv-2609.35347","arxiv":"2609.35347","title":"Beyond Teacher Assignment: Domain-Normalized Multi-Teacher On-Policy Distillation","authors":["Xin Li","Hao Jiang","Xin Gao","Annan Wang","Yuchen Xie","Jinghao Guo","Xingwei Qu","Yichi Zhang"],"authorCount":9,"org":"Nanyang Technological University","affiliations":[],"orgNote":null,"date":"2026-09-28","hfDaily":"2026-09-29","primaryCategory":"cs.LG","categories":["cs.LG"],"links":{"abs":"https://arxiv.org/abs/2609.35347","pdf":"https://arxiv.org/pdf/2609.35347","hf":"https://huggingface.co/papers/2609.35347","code":"https://github.com/LiXin97/DN-MOPD","project":"https://lixin.ai/DN-MOPD/"},"metrics":{"hfUpvotes":181,"hfComments":2,"citations":1,"citationsSource":"semantic_scholar","githubStars":57},"hot":66.61,"tags":["efficiency","rl"],"summary":{"zh":"多教师在线蒸馏中按领域归一化各教师的反馈强度，让学生同时学好数学、代码和指令遵循。","en":"Normalises each teacher's feedback strength by domain in multi-teacher on-policy distillation so one student learns maths, code and instruction following."},"disciplines":["mathematics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":47},{"id":"arxiv-2609.37725","arxiv":"2609.37725","title":"Context Language Models","authors":["Rulin Shao","Shannon Zejiang Shen","Junjie Oscar Yin","Yuetai Li","Minheng Wang","Hamish Ivison","Radha Poovendran","Nathan Lambert"],"authorCount":13,"org":"Meta","affiliations":[],"orgNote":null,"date":"2026-09-29","hfDaily":"2026-09-30","primaryCategory":"cs.AI","categories":["cs.AI","cs.CL","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2609.37725","pdf":"https://arxiv.org/pdf/2609.37725","hf":"https://huggingface.co/papers/2609.37725","code":"https://github.com/facebookresearch/context-language-models","project":"https://github.com/facebookresearch/context-language-models"},"metrics":{"hfUpvotes":47,"hfComments":6,"citations":2,"citationsSource":"semantic_scholar","githubStars":742},"hot":66.55,"tags":["llm","agent"],"summary":{"zh":"上下文语言模型把上下文当成可自由编辑的文件，自己决定保留什么，以更少算力取得更高准确率。","en":"Context Language Models treat their context as an editable file and decide what to keep, gaining accuracy with fewer FLOPs."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":48},{"id":"arxiv-2610.20341","arxiv":"2610.20341","title":"InfiniSplat v2: Analyzing and Improving Feed-Forward 3DGS from Any Views","authors":["Yongzhen Hu","Hao Yu","Jiawei Wang","Xianming Zeng","Jianyun Xu","Xingliang Liu","Ruizhen Hu","Xiaowei Zhou"],"authorCount":10,"org":"zju3dv","affiliations":[],"orgNote":null,"date":"2026-10-14","hfDaily":"2026-10-16","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.20341","pdf":"https://arxiv.org/pdf/2610.20341","hf":"https://huggingface.co/papers/2610.20341","code":"https://github.com/zju3dv/InfiniSplat","project":"https://zju3dv.github.io/InfiniSplat2/"},"metrics":{"hfUpvotes":131,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":523},"hot":66.13,"tags":["visual-gen","vision"],"summary":{"zh":"InfiniSplat v2 分析并改进前馈式 3D 高斯重建，在大视角变化下渲染质量更高。","en":"InfiniSplat v2 analyses and improves feed-forward 3D Gaussian reconstruction for better rendering under large viewpoint changes."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":49},{"id":"arxiv-2609.24981","arxiv":"2609.24981","title":"GAE: Learning a Geometry-Native Latent Space for 3D-Consistent World Generation","authors":["Jiahao Lu","Minghao Yin","Wenbo Hu","Hengyu Liu","Wang Zhao","Sai-Kit Yeung","Ying Shan","Yuan Liu"],"authorCount":8,"org":"ARC Lab, Tencent","affiliations":[],"orgNote":null,"date":"2026-09-21","hfDaily":"2026-09-23","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2609.24981","pdf":"https://arxiv.org/pdf/2609.24981","hf":"https://huggingface.co/papers/2609.24981","code":"https://github.com/TencentARC/GAE-GeometricAutoEncoder","project":"https://jiah-cloud.github.io/GAE.github.io/"},"metrics":{"hfUpvotes":78,"hfComments":4,"citations":0,"citationsSource":"semantic_scholar","githubStars":526},"hot":65.78,"tags":["visual-gen","vision"],"summary":{"zh":"GAE 把几何基础模型的特征重参数化成紧凑潜空间，让生成的世界在 3D 上保持一致。","en":"GAE reparameterises geometry foundation-model features into a compact latent space so generated worlds stay 3D-consistent."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":50},{"id":"arxiv-2610.12242","arxiv":"2610.12242","title":"TokenRouter: Efficient Serving System for Token-Level LLM Routing","authors":["Tianyu Fu","Tengxuan Liu","Ruoxi Wang","Yixin Dong","Yi Ge","Yichen You","Yu Wang"],"authorCount":7,"org":"Tsinghua-NICS-EFC","affiliations":[],"orgNote":null,"date":"2026-10-08","hfDaily":"2026-10-09","primaryCategory":"cs.CL","categories":["cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.12242","pdf":"https://arxiv.org/pdf/2610.12242","hf":"https://huggingface.co/papers/2610.12242","code":"https://github.com/thu-nics/TokenRouter","project":"https://fuvty.github.io/thinking_yard_project_page/projects/tokenrouter/"},"metrics":{"hfUpvotes":143,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":63},"hot":65.74,"tags":["efficiency"],"summary":{"zh":"TokenRouter 是支持 token 级多模型路由的推理服务系统，解决步调不同步和批处理延迟问题。","en":"TokenRouter is a serving system for token-level LLM routing that fixes step desynchronisation and batching delays."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":51},{"id":"arxiv-2610.18356","arxiv":"2610.18356","title":"Urbex: Agentic Spatial Grounding for City-Scale 3D Scenes","authors":["Yu Wang","Rui Dai","Wei Chen","Yi Wang","Bo Dang","Kaikui Liu","Xiangxiang Chu","Yansheng Li"],"authorCount":8,"org":"alibaba","affiliations":[],"orgNote":null,"date":"2026-10-13","hfDaily":"2026-10-14","primaryCategory":"cs.CV","categories":["cs.CV","cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.18356","pdf":"https://arxiv.org/pdf/2610.18356","hf":"https://huggingface.co/papers/2610.18356","code":"https://github.com/AMAP-ML/Urbex","project":"https://www.wangfaye.cn/paper/urbex"},"metrics":{"hfUpvotes":133,"hfComments":1,"citations":0,"citationsSource":"semantic_scholar","githubStars":105},"hot":65.68,"tags":["agent","vision","rl"],"summary":{"zh":"Urbex 让视觉语言 Agent 在城市级 3D 场景中主动搜索、缩放、换视角来定位目标，并用强化学习优化。","en":"Urbex treats city-scale 3D grounding as active search by a VLM agent that zooms and changes views, optimised with RL."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":52},{"id":"arxiv-2610.13361","arxiv":"2610.13361","title":"AgentSWE: Can Coding Agents Build the Agent You Actually Want?","authors":["Jiahao Wang","Hongjin Qian","Yuyang Hu","Jiajun Zhang","Zheng Liu"],"authorCount":5,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-10-09","hfDaily":"2026-10-13","primaryCategory":"cs.SE","categories":["cs.SE","cs.AI","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2610.13361","pdf":"https://arxiv.org/pdf/2610.13361","hf":"https://huggingface.co/papers/2610.13361","code":"https://github.com/VectorSpaceLab/AgentSWE","project":"https://vectorspacelab.github.io/AgentSWE/"},"metrics":{"hfUpvotes":233,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":42},"hot":65.6,"tags":["agent","code","benchmark"],"summary":{"zh":"AgentSWE 评测编程 Agent 能否按需求从零搭建、修改和优化真正好用的 Agent。","en":"AgentSWE benchmarks whether coding agents can create, edit and optimise agents that truly meet the stated requirement."},"disciplines":["software-engineering"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":53},{"id":"arxiv-2609.35432","arxiv":"2609.35432","title":"Self-Evolving Coding Agents: From Digital Programs to Physical-World Intelligence","authors":["Hongcheng Gao","Jingjing Zhou","Zelin Zheng","Shijia Ge","Jay Zhu","Yazhe Wang","Jianshu Zeng","Xuan Shangguan"],"authorCount":13,"org":"HexaFuture","affiliations":[],"orgNote":null,"date":"2026-09-28","hfDaily":"2026-09-29","primaryCategory":"cs.RO","categories":["cs.RO"],"links":{"abs":"https://arxiv.org/abs/2609.35432","pdf":"https://arxiv.org/pdf/2609.35432","hf":"https://huggingface.co/papers/2609.35432","code":"https://github.com/HexaFuture/PhysicalCoding","project":"https://hexafuture.ai/blog/physical-coding"},"metrics":{"hfUpvotes":112,"hfComments":2,"citations":4,"citationsSource":"semantic_scholar","githubStars":54},"hot":65.38,"tags":["robotics","code","agent"],"summary":{"zh":"主张把机器人任务写成可检查、可修改的代码，让自进化的编程 Agent 从数字世界走向物理世界。","en":"Argues for expressing robot tasks as inspectable, revisable code so self-evolving coding agents can move into the physical world."},"disciplines":["software-engineering","robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":54},{"id":"arxiv-2609.25001","arxiv":"2609.25001","title":"GameHorizon Suite: Multi-Horizon Data and Evaluation in Gameplay","authors":["Yiran Wang","Xingyilang Yin","Junfu Pu","Guangzhi Wang","Kaifeng Li","Mingyu Ouyang","Huiqiang Sun","Lingen Li"],"authorCount":15,"org":"Tencent","affiliations":[],"orgNote":null,"date":"2026-09-21","hfDaily":"2026-09-22","primaryCategory":"cs.CV","categories":["cs.CV","cs.AI"],"links":{"abs":"https://arxiv.org/abs/2609.25001","pdf":"https://arxiv.org/pdf/2609.25001","hf":"https://huggingface.co/papers/2609.25001","code":"https://github.com/TencentARC/GameHorizon","project":"https://gamehorizon-suite.github.io/"},"metrics":{"hfUpvotes":73,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":572},"hot":65,"tags":["benchmark","agent"],"summary":{"zh":"GameHorizon 提供多时间尺度的游戏数据标注和评测，衡量模型的规划和操作能力。","en":"GameHorizon provides multi-horizon gameplay data and evaluation to measure planning and control across model families."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":55},{"id":"arxiv-2610.23374","arxiv":"2610.23374","title":"MOSS-Image Technical Report","authors":["MOSI. AI"],"authorCount":1,"org":"OpenMOSS","affiliations":[],"orgNote":null,"date":"2026-10-16","hfDaily":"2026-10-19","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.23374","pdf":"https://arxiv.org/pdf/2610.23374","hf":"https://huggingface.co/papers/2610.23374","code":"https://github.com/OpenMOSS/MOSS-Image","project":"https://openmoss.ai/MOSS-Image/"},"metrics":{"hfUpvotes":89,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":1246},"hot":64.9,"tags":["visual-gen"],"summary":{"zh":"MOSS-Image 是开源 5B 文生图系统，擅长中英长文本渲染和复杂版式，并配有意图感知的提示增强。","en":"MOSS-Image is an open 5B text-to-image system strong at bilingual long-text rendering and layouts, with intent-aware prompt enhancement."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":56},{"id":"arxiv-2610.03543","arxiv":"2610.03543","title":"DuoMatching: Joint-Marginal Distribution Matching for Few-Step Video Generation","authors":["Jiahao Zhan","Yan Wang","Yongrui Ma","Qunliang Xing","Ruchang Yao","Runtao Liu","Shijie Zhao","Tianfan Xue"],"authorCount":8,"org":"ByteDance","affiliations":[],"orgNote":null,"date":"2026-10-02","hfDaily":"2026-10-07","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.03543","pdf":"https://arxiv.org/pdf/2610.03543","hf":"https://huggingface.co/papers/2610.03543","code":"https://github.com/JohnZhan2023/DuoMatching","project":"https://johnzhan2023.github.io/DuoMatching/"},"metrics":{"hfUpvotes":101,"hfComments":2,"citations":1,"citationsSource":"semantic_scholar","githubStars":56},"hot":64.78,"tags":["visual-gen","efficiency"],"summary":{"zh":"DuoMatching 同时匹配联合分布和逐帧边缘分布，提升少步视频生成的画质和语义对齐。","en":"DuoMatching matches both joint and per-frame marginal distributions to improve quality and alignment in few-step video generation."},"disciplines":["film-animation"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":57},{"id":"arxiv-2609.39600","arxiv":"2609.39600","title":"GroundAnything: Reconciling Parallel Decoding with Precise Visual Grounding at Flash Speed","authors":["Qize Yu","Lianrui Fan","Bowen Ping","Xini Ding","Zetian Song","Junbo Niu","Kaixuan Wang","Tianxing Chen"],"authorCount":22,"org":"XPENG AI","affiliations":[],"orgNote":null,"date":"2026-09-30","hfDaily":"2026-10-13","primaryCategory":"cs.CV","categories":["cs.CV","cs.AI","cs.LG","cs.RO"],"links":{"abs":"https://arxiv.org/abs/2609.39600","pdf":"https://arxiv.org/pdf/2609.39600","hf":"https://huggingface.co/papers/2609.39600","code":"https://github.com/groundingpi/GroundAnything","project":"https://groundingpi.github.io/groundanything/"},"metrics":{"hfUpvotes":94,"hfComments":2,"citations":4,"citationsSource":"semantic_scholar","githubStars":78},"hot":64.55,"tags":["vision","efficiency"],"summary":{"zh":"GroundAnything 用双向扩散并行解码目标位置，在保持精准的同时大幅加快视觉定位。","en":"GroundAnything decodes grounding outputs in parallel with bidirectional diffusion, keeping precision while being much faster."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":58},{"id":"arxiv-2610.02054","arxiv":"2610.02054","title":"UniWAM: Unified World-Action Model","authors":["Wenxuan Song","Jiayi Chen","Jingbo Wang","Shuai Zhou","Xicheng Gong","Zehua Fan","Ziyang Zhou","Junwu E"],"authorCount":16,"org":"HKUSTGZ","affiliations":[],"orgNote":null,"date":"2026-10-01","hfDaily":"2026-10-08","primaryCategory":"cs.RO","categories":["cs.RO"],"links":{"abs":"https://arxiv.org/abs/2610.02054","pdf":"https://arxiv.org/pdf/2610.02054","hf":"https://huggingface.co/papers/2610.02054","code":"https://github.com/UniWAM/UniWAM","project":"https://uniwam.github.io/"},"metrics":{"hfUpvotes":83,"hfComments":4,"citations":3,"citationsSource":"semantic_scholar","githubStars":131},"hot":64.47,"tags":["robotics","multimodal"],"summary":{"zh":"UniWAM 把物理推理、世界生成和动作预测统一到一个模型里，用于机器人控制。","en":"UniWAM unifies a physical reasoner, a world generator and an action predictor in one model for robot control."},"disciplines":["robotics","film-animation"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":59},{"id":"arxiv-2610.26825","arxiv":"2610.26825","title":"MemT-1.0: Towards Memory-Native Evolution","authors":["Zhiyu Li","Zeyu Zhang","Linyun Liu","Yiwei Bu","Chenyang Xi","Xichong Zhang","Yi Yu","Yu Yu"],"authorCount":74,"org":"MemTensor","affiliations":[],"orgNote":null,"date":"2026-10-19","hfDaily":"2026-10-20","primaryCategory":"cs.CL","categories":["cs.CL","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2610.26825","pdf":"https://arxiv.org/pdf/2610.26825","hf":"https://huggingface.co/papers/2610.26825","code":"https://github.com/MemTensor/MemT-1.0","project":"https://huggingface.co/collections/MemTensor/memt-1"},"metrics":{"hfUpvotes":119,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":272},"hot":63.98,"tags":["llm","efficiency"],"summary":{"zh":"MemT-1.0 把记忆做成模型计算的一部分，基于 Gated DeltaNet-2 和混合存储扩展原生记忆容量。","en":"MemT-1.0 makes memory intrinsic to the model using Gated DeltaNet-2 and Mixture-of-Storage to scale native memory."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":60},{"id":"arxiv-2610.26860","arxiv":"2610.26860","title":"OpenVLHarness: Real World Visual Agentic Tool-Use with Multimodal Memory","authors":["Yu Zhou","Cheng-Fu Yang","Rui Sun","Di Wu","Sibo Peng","Mingyang Zhang","Zhen Yang","Ruiyi Zhang"],"authorCount":11,"org":"Apple","affiliations":[],"orgNote":null,"date":"2026-10-19","hfDaily":"2026-10-20","primaryCategory":"cs.CL","categories":["cs.CL","cs.AI","cs.CV","cs.LG","cs.MM"],"links":{"abs":"https://arxiv.org/abs/2610.26860","pdf":"https://arxiv.org/pdf/2610.26860","hf":"https://huggingface.co/papers/2610.26860","code":"https://github.com/uclanlp/OpenVLHarness","project":"https://openvlharness.github.io/"},"metrics":{"hfUpvotes":102,"hfComments":1,"citations":null,"citationsSource":null,"githubStars":152},"hot":63.96,"tags":["multimodal","agent"],"summary":{"zh":"OpenVLHarness 为多模态模型提供工具能力层和持久多模态记忆，处理需要外部证据的视觉任务。","en":"OpenVLHarness gives MLLMs a tool capability layer and persistent multimodal memory for visual tasks needing external evidence."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":61},{"id":"arxiv-2610.14993","arxiv":"2610.14993","title":"Learning Where to Split: Dynamic Band Splitting for Audio Source Separation","authors":["Bingzhe Zhang","Kai Li","Xiaolin Hu"],"authorCount":3,"org":"Tsinghua Machine Learning Group","affiliations":[],"orgNote":null,"date":"2026-10-10","hfDaily":"2026-10-16","primaryCategory":"cs.SD","categories":["cs.SD"],"links":{"abs":"https://arxiv.org/abs/2610.14993","pdf":"https://arxiv.org/pdf/2610.14993","hf":"https://huggingface.co/papers/2610.14993","code":"https://github.com/JusperLee/Dynamic-Band-Splitting","project":"https://cslikai.cn/Dynamic-Band-Splitting/"},"metrics":{"hfUpvotes":143,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":30},"hot":63.93,"tags":["multimodal"],"summary":{"zh":"动态频带切分方法按每段混音自适应划分频带，提升音频源分离的效果和效率。","en":"Dynamic Band Splitting adapts frequency band boundaries per mixture to improve audio source separation within a fixed token budget."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":62},{"id":"arxiv-2610.19159","arxiv":"2610.19159","title":"Into the Unknown: Benchmarking AI agents on open scientific research problems","authors":["Seungone Kim","Jiayi Geng","Young-Jun Lee","Bo Liu","Juyoung Suk","Nicholas Lubbers","Mitchell Wood","Marcus Noack"],"authorCount":28,"org":"Carnegie Mellon University","affiliations":[],"orgNote":null,"date":"2026-10-13","hfDaily":"2026-10-19","primaryCategory":"cs.AI","categories":["cs.AI","cs.CL","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2610.19159","pdf":"https://arxiv.org/pdf/2610.19159","hf":"https://huggingface.co/papers/2610.19159","code":"https://github.com/prometheus-eval/into-the-unknown","project":"https://prometheus-eval.github.io/into-the-unknown/viewer/#/"},"metrics":{"hfUpvotes":155,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":19},"hot":63.64,"tags":["agent","ai4science","benchmark"],"summary":{"zh":"Into the Unknown 让 Agent 针对美国国家实验室科学家提供的未发表方向写续作论文，评测开放科研能力。","en":"Into the Unknown asks agents to write follow-up papers on unpublished directions from US national-lab scientists to test open research ability."},"disciplines":["chemistry","materials","physics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":63},{"id":"arxiv-2610.24087","arxiv":"2610.24087","title":"OR-License: Evaluating LLM Agents as Real-World Operations Research Practitioners","authors":["Peisong Wang","Feixue Liu","Yize Meng","Bowen Liu","Ruoyu Chen","Gaoyang Dai","Xing He","Aiqian Li"],"authorCount":19,"org":"ALLE AI","affiliations":[],"orgNote":null,"date":"2026-10-16","hfDaily":"2026-10-20","primaryCategory":"cs.CL","categories":["cs.CL","cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.24087","pdf":"https://arxiv.org/pdf/2610.24087","hf":"https://huggingface.co/papers/2610.24087","code":"https://github.com/NineAbyss/OR-License","project":"https://www.or-license.com/"},"metrics":{"hfUpvotes":291,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":5},"hot":63.33,"tags":["benchmark","agent"],"summary":{"zh":"OR-License 用 31 个真实付费运筹咨询项目的原始材料，测试大模型智能体能否直接交付专家级运筹方案。","en":"OR-License benchmarks whether LLM agents can turn original client materials from 31 paid operations-research engagements into expert-grade solutions."},"disciplines":["materials"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":64},{"id":"arxiv-2609.38537","arxiv":"2609.38537","title":"Systematically Exploring the Capabilities of GPT-6 Astra as Embodied Policies","authors":[" Galbot Team","Xuchuan Chen","Xiaoqian Cheng","Yu Deng","Lihe Ding","Shaocong Dong","Xiangjun Gao","Haozhe Jia"],"authorCount":34,"org":"Galbot","affiliations":[],"orgNote":null,"date":"2026-09-29","hfDaily":"2026-10-01","primaryCategory":"cs.RO","categories":["cs.RO"],"links":{"abs":"https://arxiv.org/abs/2609.38537","pdf":"https://arxiv.org/pdf/2609.38537","hf":"https://huggingface.co/papers/2609.38537","code":"https://github.com/anonymous-report-421/GPT-as-Policy","project":"https://galaxygeneralrobotics.github.io/astra-policy/"},"metrics":{"hfUpvotes":34,"hfComments":3,"citations":9,"citationsSource":"semantic_scholar","githubStars":637},"hot":62.99,"tags":["robotics","benchmark"],"summary":{"zh":"在六类具身任务上系统测试 GPT-6 Astra 直接输出机器人动作的能力，以及与专用策略的配合。","en":"Systematically tests GPT-6 Astra as an embodied policy across six domains, alone and combined with learned policies."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":65},{"id":"arxiv-2610.01415","arxiv":"2610.01415","title":"Beyond Memory: Harnessing Long-Horizon Agents with Explicit Belief States","authors":["Yu Luo","Jiamin Jiang","Yimin Zuo","Xidao Wen","Rongchen Gao","Yongqian Sun","Shenglin Zhang","Guiyang Liu"],"authorCount":12,"org":"alibaba","affiliations":[],"orgNote":null,"date":"2026-10-01","hfDaily":"2026-10-02","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.01415","pdf":"https://arxiv.org/pdf/2610.01415","hf":"https://huggingface.co/papers/2610.01415","code":"https://github.com/luoyu100/PoS","project":"https://luoyu100.github.io/projects/progression-of-states/project/"},"metrics":{"hfUpvotes":92,"hfComments":3,"citations":1,"citationsSource":"semantic_scholar","githubStars":32},"hot":62.92,"tags":["agent"],"summary":{"zh":"PoS 让 Agent 显式维护“信念状态”，记录对世界的估计和未完成的需求，避免长任务中陷入错误信念。","en":"PoS has agents maintain explicit belief states of the world and open requirements, avoiding belief traps in long tasks."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":66},{"id":"arxiv-2609.38155","arxiv":"2609.38155","title":"Beyond the Timeline: Augmenting Long-Video Memory with Grounded Entity Biographies","authors":["Hui Ren","Lei Fan","Henry Pao","Han Guo","Zeeshan Zia","Ying Chen","Alexander Schwing","Gang Hua"],"authorCount":8,"org":"Amazon Science","affiliations":[],"orgNote":null,"date":"2026-09-29","hfDaily":"2026-09-30","primaryCategory":"cs.CV","categories":["cs.CV","cs.AI","cs.CL","cs.IR","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2609.38155","pdf":"https://arxiv.org/pdf/2609.38155","hf":"https://huggingface.co/papers/2609.38155","code":"https://github.com/rhfeiyang/GEB","project":"https://geb-video.github.io/"},"metrics":{"hfUpvotes":116,"hfComments":1,"citations":0,"citationsSource":"semantic_scholar","githubStars":59},"hot":62.9,"tags":["multimodal"],"summary":{"zh":"GEB 把长视频中同一物体的观察串成“实体传记”，帮助回答跨时间的问题。","en":"GEB groups observations of the same physical object into entity biographies to answer questions over long videos."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":67},{"id":"arxiv-2610.08077","arxiv":"2610.08077","title":"Self-Retrospection Distillation: Turning Post-hoc Experiences into Prior Foresight","authors":["Haoxiang Zhang","Qinglin Chen","Hiroaki Hayashi","Zhuofeng Li","Siming Zhang","Jiaxin Zhang","Jixuan Chen","Fang Wu"],"authorCount":12,"org":"Salesforce AI Research","affiliations":[],"orgNote":null,"date":"2026-10-06","hfDaily":"2026-10-08","primaryCategory":"cs.AI","categories":["cs.AI","cs.CL","cs.IR","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2610.08077","pdf":"https://arxiv.org/pdf/2610.08077","hf":"https://huggingface.co/papers/2610.08077","code":"https://github.com/SalesforceAIResearch/SRD","project":null},"metrics":{"hfUpvotes":189,"hfComments":2,"citations":1,"citationsSource":"semantic_scholar","githubStars":10},"hot":62.89,"tags":["rl","agent"],"summary":{"zh":"自我回顾蒸馏把事后经验转成事前预判，弥补 RLVR 在奖励全相同时信号消失的问题。","en":"Self-Retrospection Distillation turns hindsight into foresight supervision, filling the gap when RLVR rewards are all identical."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":68},{"id":"arxiv-2610.13912","arxiv":"2610.13912","title":"RSI-Anything: Unlocking Recursive Self-Improvement with World Models","authors":["Taoyong Cui","Pinxian Zeng","Wanghan Xu","Hao Xiao","Shuhan Xue","Qiang Gao","Wanli Ouyang","Pheng Ann Heng"],"authorCount":11,"org":"PhAI Labs","affiliations":[],"orgNote":null,"date":"2026-10-09","hfDaily":"2026-10-12","primaryCategory":"cs.CL","categories":["cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.13912","pdf":"https://arxiv.org/pdf/2610.13912","hf":"https://huggingface.co/papers/2610.13912","code":"https://github.com/Gen-Verse/RSI-Anything","project":"https://huggingface.co/datasets/Gen-Verse/RSI-Anything"},"metrics":{"hfUpvotes":54,"hfComments":2,"citations":2,"citationsSource":"semantic_scholar","githubStars":809},"hot":62.86,"tags":["agent","ai4science"],"summary":{"zh":"RSI-Anything 借助预测式世界模型选择行动和训练数据，实现跨领域的递归自我改进。","en":"RSI-Anything uses predictive world models to guide actions and data selection for domain-general recursive self-improvement."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":69},{"id":"arxiv-2610.21601","arxiv":"2610.21601","title":"Waypoint-OPD: Transferring RL-Induced Policy Shifts via Gated On-Policy Distillation","authors":["Bingxi Hou","Guochao Jiang","Guofeng Quan","Weiqing Li","Wenfeng Feng","Guohua Liu","Yuewei Zhang"],"authorCount":7,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-10-15","hfDaily":"2026-10-16","primaryCategory":"cs.LG","categories":["cs.LG","cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.21601","pdf":"https://arxiv.org/pdf/2610.21601","hf":"https://huggingface.co/papers/2610.21601","code":"https://github.com/hhhbxloveai/WaypointOPD","project":null},"metrics":{"hfUpvotes":218,"hfComments":6,"citations":0,"citationsSource":"semantic_scholar","githubStars":10},"hot":62.8,"tags":["rl","efficiency"],"summary":{"zh":"Waypoint-OPD 借助教师 RL 训练中途的检查点，只迁移前后一致的策略变化。","en":"Waypoint-OPD uses an intermediate teacher checkpoint to transfer only consistent RL-induced policy shifts to the student."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":70},{"id":"arxiv-2610.26203","arxiv":"2610.26203","title":"Focused Omni-Modal GUI Agents","authors":["Fuyao Zhang","Wenwen He","Jiaming Zhang","Yurong Hao","Xuan Wang","Guowei Guan","Che Wang","Longtao Huang"],"authorCount":9,"org":"Nanyang Technological University","affiliations":[],"orgNote":null,"date":"2026-10-19","hfDaily":"2026-10-21","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.26203","pdf":"https://arxiv.org/pdf/2610.26203","hf":"https://huggingface.co/papers/2610.26203","code":"https://github.com/fyzhang1/FocusGUI","project":"https://fyzhang1.github.io/FocusGUI/"},"metrics":{"hfUpvotes":225,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":12},"hot":62.7,"tags":["agent","multimodal"],"summary":{"zh":"FocusGUI 用轻量适配器按步骤挑出最相关的屏幕、音频等信号注入冻结模型，缓解 GUI 智能体被无关上下文带偏。","en":"FocusGUI is a lightweight adapter that picks the step-relevant omni-modal evidence for a frozen GUI agent, curbing compounding distraction."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":71},{"id":"arxiv-2610.26855","arxiv":"2610.26855","title":"Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human Experience","authors":["Tianqi Liu","Hanfei Zhao","Qifeng Liang","Hui En Pang","Jiarui Xu","Sayang Mu","Penghao Qiu","YingSheng Zhang"],"authorCount":15,"org":"Ropedia","affiliations":[],"orgNote":null,"date":"2026-10-19","hfDaily":"2026-10-21","primaryCategory":"cs.CV","categories":["cs.CV","cs.RO"],"links":{"abs":"https://arxiv.org/abs/2610.26855","pdf":"https://arxiv.org/pdf/2610.26855","hf":"https://huggingface.co/papers/2610.26855","code":"https://github.com/Ropedia/ropedia-w0","project":"https://ropedia.github.io/ropedia-w0/"},"metrics":{"hfUpvotes":179,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":36},"hot":62.6,"tags":["robotics","visual-gen"],"summary":{"zh":"Ropedia W0 是跨人类与多种机器人的第一视角移动操作世界基础模型，用 1 万小时人类、3 千小时机器人数据训练。","en":"Ropedia W0 is an egocentric loco-manipulation world foundation model spanning humans and robots, trained on 10k hours of human and 3k hours of robot data."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":72},{"id":"arxiv-2609.38169","arxiv":"2609.38169","title":"STEPQuant: When and Where Errors Matter in Delta-Rule Recurrent State Quantization","authors":["Bingchen Yao","Haobo Xu","Haokun Lin","Yichen Wu","Ziyu Guo","Renrui Zhang","Zhichao Lu","Zhenan Sun"],"authorCount":9,"org":"Zhejiang University","affiliations":[],"orgNote":null,"date":"2026-09-29","hfDaily":"2026-10-08","primaryCategory":"cs.CL","categories":["cs.CL","cs.AI","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2609.38169","pdf":"https://arxiv.org/pdf/2609.38169","hf":"https://huggingface.co/papers/2609.38169","code":"https://github.com/Dreamer-Toby/STEPQuant","project":null},"metrics":{"hfUpvotes":131,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":101},"hot":62.47,"tags":["efficiency"],"summary":{"zh":"STEPQuant 分析线性注意力循环状态量化误差在时间和空间上的影响，实现低精度状态存储。","en":"STEPQuant studies where and when quantisation errors matter in linear-attention recurrent states, enabling low-precision state storage."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":73},{"id":"arxiv-2610.22341","arxiv":"2610.22341","title":"3DCodeVerse: Exploring 3D Worlds as Code","authors":["Yipeng Gao","Ziyao Zeng","Jinfa Huang","Yijiang Li","Haolin Xiong","Wang Qin","Xinlei Yu","Youheng Yao"],"authorCount":19,"org":"University of Southern California","affiliations":[],"orgNote":null,"date":"2026-10-15","hfDaily":"2026-10-16","primaryCategory":"cs.CV","categories":["cs.CV","cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.22341","pdf":"https://arxiv.org/pdf/2610.22341","hf":"https://huggingface.co/papers/2610.22341","code":"https://github.com/gaoypeng/3dcodeverse","project":"https://3dcodeverse.com/"},"metrics":{"hfUpvotes":117,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":185},"hot":62.42,"tags":["code","visual-gen"],"summary":{"zh":"3DCodeVerse 提供 200 万条可执行 3D 代码数据和编码框架，让 Agent 生成可编辑可验证的 3D 内容。","en":"3DCodeVerse offers 2M executable 3D programs and a coding harness so agents can create editable, verifiable 3D content."},"disciplines":["software-engineering"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":74},{"id":"arxiv-2609.24974","arxiv":"2609.24974","title":"Harness-Zero: Harness Distillation via Agent-as-Harness","authors":["Haoran Ye","Yuxing Lu","Haonan Dong","Zhaochen Su","Guojie Song"],"authorCount":5,"org":"Peking University","affiliations":[],"orgNote":null,"date":"2026-09-21","hfDaily":"2026-09-22","primaryCategory":"cs.AI","categories":["cs.AI","cs.CL","cs.NE"],"links":{"abs":"https://arxiv.org/abs/2609.24974","pdf":"https://arxiv.org/pdf/2609.24974","hf":"https://huggingface.co/papers/2609.24974","code":"https://github.com/metaevo-ai/harness-zero","project":null},"metrics":{"hfUpvotes":38,"hfComments":3,"citations":6,"citationsSource":"semantic_scholar","githubStars":115},"hot":62.38,"tags":["agent","efficiency"],"summary":{"zh":"Harness-Zero 把专用 Agent 框架带来的能力蒸馏进模型权重，换成通用框架也不掉点。","en":"Harness-Zero distils the gains of specialised agent harnesses into model weights so they persist under a single fixed harness."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":75},{"id":"arxiv-2610.00812","arxiv":"2610.00812","title":"Video Generation Models: A Survey of Post-Training and Alignment","authors":["Chaoyu Li","Xiaoyi Gu","Yogesh Kulkarni","Eun Woo Im","Mohammadmahdi Honarmand","Zeyu Wang","Juntong Song","Fei Du"],"authorCount":13,"org":"Arizona State University","affiliations":[],"orgNote":null,"date":"2026-09-30","hfDaily":"2026-10-02","primaryCategory":"cs.CV","categories":["cs.CV","cs.AI","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2610.00812","pdf":"https://arxiv.org/pdf/2610.00812","hf":"https://huggingface.co/papers/2610.00812","code":"https://github.com/people-robots/Awesome-Video-Generation-Post-Training","project":null},"metrics":{"hfUpvotes":63,"hfComments":2,"citations":3,"citationsSource":"semantic_scholar","githubStars":210},"hot":62.16,"tags":["visual-gen","alignment"],"summary":{"zh":"视频生成模型后训练与对齐的综述，梳理时序一致性、物理和安全约束等难点。","en":"A survey of post-training and alignment for video generation, covering temporal coherence, physics and safety challenges."},"disciplines":["film-animation"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":76},{"id":"arxiv-2609.35259","arxiv":"2609.35259","title":"On-Policy or Off-Policy Learning? A Systematic Study of Distillation Dynamics","authors":["Julianna Piskorz","Antonin Berthon","Mihaela van der Schaar"],"authorCount":3,"org":"University of Cambridge","affiliations":[],"orgNote":null,"date":"2026-09-28","hfDaily":"2026-10-02","primaryCategory":"cs.LG","categories":["cs.LG"],"links":{"abs":"https://arxiv.org/abs/2609.35259","pdf":"https://arxiv.org/pdf/2609.35259","hf":"https://huggingface.co/papers/2609.35259","code":null,"project":"https://antoninbrthn.github.io/on-off-policy-distillation/"},"metrics":{"hfUpvotes":205,"hfComments":3,"citations":1,"citationsSource":"semantic_scholar","githubStars":null},"hot":62.13,"tags":["efficiency","reasoning"],"summary":{"zh":"在受控实验中分离 on-policy 与 off-policy 蒸馏的作用，发现结论比“on-policy 总更好”复杂。","en":"A controlled study isolating on- vs off-policy rollouts in distillation, finding a more nuanced picture than 'on-policy is always better'."},"disciplines":["clinical-medicine"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":77},{"id":"arxiv-2609.38721","arxiv":"2609.38721","title":"UniEvo-VL: An On-policy Self-Distillation Training Recipe for Multimodal Model Self-improvement","authors":["Fang Wu","Da Xing","Yanjie Huang","Junxi Wang","Ji Wang","Hejia Geng","Guancheng Wan","Bowen Zuo"],"authorCount":19,"org":"Stanford NLP","affiliations":[],"orgNote":null,"date":"2026-09-30","hfDaily":"2026-10-01","primaryCategory":"cs.AI","categories":["cs.AI","cs.CV"],"links":{"abs":"https://arxiv.org/abs/2609.38721","pdf":"https://arxiv.org/pdf/2609.38721","hf":"https://huggingface.co/papers/2609.38721","code":null,"project":null},"metrics":{"hfUpvotes":266,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":61.95,"tags":["multimodal"],"summary":{"zh":"UniEvo-VL 让同一个多模态模型分别扮演教师和学生，用自我批评做在线自蒸馏实现自我提升。","en":"UniEvo-VL has one multimodal model act as both teacher and student, using its own critiques for on-policy self-distillation."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":78},{"id":"arxiv-2609.34759","arxiv":"2609.34759","title":"PanoVLN: Towards Effective Panoramic Vision-and-Language Navigation","authors":["Zhen Wang","Changpeng Wang","Zhe Liu","Zhangyang Qi","Yuxiang Lu","Zimo Zeng","Donglian Qi","Xi Chen"],"authorCount":8,"org":"Zhejiang University","affiliations":[],"orgNote":null,"date":"2026-09-28","hfDaily":"2026-09-30","primaryCategory":"cs.CV","categories":["cs.CV","cs.RO"],"links":{"abs":"https://arxiv.org/abs/2609.34759","pdf":"https://arxiv.org/pdf/2609.34759","hf":"https://huggingface.co/papers/2609.34759","code":"https://github.com/wangzhen-w/PanoVLN","project":"https://wangzhen-w.github.io/PanoVLN/"},"metrics":{"hfUpvotes":151,"hfComments":1,"citations":0,"citationsSource":"semantic_scholar","githubStars":70},"hot":61.94,"tags":["robotics","multimodal"],"summary":{"zh":"PanoVLN 研究如何让视觉语言导航真正用好全景图像，而不是简单替换透视图。","en":"PanoVLN studies how to make vision-and-language navigation actually benefit from panoramic observations."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":79},{"id":"arxiv-2610.08215","arxiv":"2610.08215","title":"Learn2Play Bench: How Well Do LLM Agents Learn from Experience in Unfamiliar Environments?","authors":["Yibo Li","Jinhang Qiu","Zhi Zheng","Qianyun Guo","Jiaying Wu","Shuo Ji","Bryan Hooi"],"authorCount":7,"org":"National University of Singapore","affiliations":[],"orgNote":null,"date":"2026-10-06","hfDaily":"2026-10-09","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.08215","pdf":"https://arxiv.org/pdf/2610.08215","hf":"https://huggingface.co/papers/2610.08215","code":"https://github.com/liushiliushi/Learn2Play-Bench","project":"https://liushiliushi.github.io/learn2play-bench-website/"},"metrics":{"hfUpvotes":153,"hfComments":3,"citations":1,"citationsSource":"semantic_scholar","githubStars":13},"hot":61.82,"tags":["agent","benchmark"],"summary":{"zh":"Learn2Play Bench 用规则新奇甚至反直觉的文字游戏，检验 Agent 能否从经验中学习。","en":"Learn2Play Bench uses text games with novel or counterintuitive rules to test whether agents learn from experience."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":80},{"id":"arxiv-2609.32712","arxiv":"2609.32712","title":"MassAlloc Attention: Let Attention Allocate Its Own Compute","authors":["Jingze Shi","Zhangyang Peng","Xianduo Li","Yanlin Qi","Xiaotian Lin","Haoxian Chen","Liangdong Wang","Guang Liu"],"authorCount":9,"org":"Beijing Academy of Artificial Intelligence","affiliations":[],"orgNote":null,"date":"2026-09-26","hfDaily":"2026-09-29","primaryCategory":"cs.AI","categories":["cs.AI","cs.CL"],"links":{"abs":"https://arxiv.org/abs/2609.32712","pdf":"https://arxiv.org/pdf/2609.32712","hf":"https://huggingface.co/papers/2609.32712","code":"https://github.com/HKUSTDial/flash-sparse-attention","project":null},"metrics":{"hfUpvotes":75,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":763},"hot":61.55,"tags":["efficiency"],"summary":{"zh":"MALA 注意力按每个位置的贡献自适应分配计算，跳过贡献极小的部分以提速。","en":"MALA attention allocates post-score computation by normalised contribution, skipping negligible work to save compute."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":81},{"id":"arxiv-2609.23986","arxiv":"2609.23986","title":"Jev-Mem: System-One-Controlled Agentic Memory for Efficient AI Agents","authors":["Dongming Jiang","Yi Li","Bingzhe Li"],"authorCount":3,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-09-21","hfDaily":"2026-09-22","primaryCategory":"cs.AI","categories":["cs.AI","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2609.23986","pdf":"https://arxiv.org/pdf/2609.23986","hf":"https://huggingface.co/papers/2609.23986","code":"https://github.com/libingzheren/Jev-Mem","project":null},"metrics":{"hfUpvotes":30,"hfComments":2,"citations":19,"citationsSource":"semantic_scholar","githubStars":224},"hot":61.51,"tags":["agent","efficiency"],"summary":{"zh":"Jev-Mem 用快速的 System 1 控制面管理 Agent 记忆，把昂贵的生成移出记忆操作的关键路径。","en":"Jev-Mem uses a fast System 1 control plane to manage agent memory, removing costly generation from the critical path."},"disciplines":["economics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":82},{"id":"arxiv-2610.08621","arxiv":"2610.08621","title":"Recursive Game Creator: An Agentic Product-Level Experience-Oriented Game Harness","authors":["Jiajun Chen","Haoyu Wu","Mingda Jia","Xihui Liu"],"authorCount":4,"org":"The University of Hong Kong","affiliations":[],"orgNote":null,"date":"2026-10-06","hfDaily":"2026-10-08","primaryCategory":"cs.AI","categories":["cs.AI","cs.MA","cs.SE"],"links":{"abs":"https://arxiv.org/abs/2610.08621","pdf":"https://arxiv.org/pdf/2610.08621","hf":"https://huggingface.co/papers/2610.08621","code":"https://github.com/IMBALDY/RecursiveGameCreator","project":"https://imbaldy.github.io/recursive-game-creator"},"metrics":{"hfUpvotes":97,"hfComments":3,"citations":1,"citationsSource":"semantic_scholar","githubStars":74},"hot":61.5,"tags":["agent","code"],"summary":{"zh":"Recursive Game Creator 用设计、构建、试玩、评审四个角色递归迭代，让 Agent 做出好玩的游戏。","en":"Recursive Game Creator iterates with designer, builder, player and reviewer roles so agents make games that are actually fun."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":83},{"id":"arxiv-2609.35748","arxiv":"2609.35748","title":"Improving Test-Time Scaling with Adaptive Looped Transformers","authors":["Yichen You","Tianyu Fu","Aosong Feng","Xingtai Lv","Xuefei Ning","Ning Ding","Yu Wang"],"authorCount":7,"org":"Tsinghua-NICS-EFC","affiliations":[],"orgNote":null,"date":"2026-09-28","hfDaily":"2026-09-29","primaryCategory":"cs.CL","categories":["cs.CL","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2609.35748","pdf":"https://arxiv.org/pdf/2609.35748","hf":"https://huggingface.co/papers/2609.35748","code":"https://github.com/thu-nics/TaH","project":"https://fuvty.github.io/thinking_yard_project_page/projects/tah2/"},"metrics":{"hfUpvotes":62,"hfComments":2,"citations":1,"citationsSource":"semantic_scholar","githubStars":102},"hot":61.44,"tags":["reasoning","efficiency"],"summary":{"zh":"研究循环 Transformer 在测试时扩展中的表现，并提出自适应循环以在同等算力下胜过基线。","en":"Studies looped transformers under test-time scaling and proposes adaptive looping to beat baselines at matched compute."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":84},{"id":"arxiv-2609.31620","arxiv":"2609.31620","title":"FuseReg: Regularizing Layer Fusion Mitigates the Reconstruction-Generation Gap in Representation Autoencoders","authors":["Hongyang Du","Yunfei Xie","Junjie Ye","Jiawei Yang","Xiaoyan Cong","Haodong Zhang","Yongchao Huang","Haiyu Wu"],"authorCount":16,"org":"USC Physical Superintelligence (PSI) Lab","affiliations":[],"orgNote":null,"date":"2026-09-25","hfDaily":"2026-09-28","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2609.31620","pdf":"https://arxiv.org/pdf/2609.31620","hf":"https://huggingface.co/papers/2609.31620","code":"https://github.com/Hongyang-Du/FuseReg","project":"https://hongyang-du.github.io/FuseReg/"},"metrics":{"hfUpvotes":138,"hfComments":3,"citations":1,"citationsSource":"semantic_scholar","githubStars":17},"hot":61.42,"tags":["visual-gen"],"summary":{"zh":"FuseReg 用可训练的层融合加正则，缓解表征自编码器在重建和生成之间的取舍。","en":"FuseReg learns regularised layer fusion to ease the reconstruction-generation trade-off in representation autoencoders."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":85},{"id":"arxiv-2610.01509","arxiv":"2610.01509","title":"Sharpening Tax in Post-Training","authors":["Changdae Oh","Qi Zeng","Qi Qi","Andrey Zhmoginov","Deren Lei","Yun He","Hoang Phan","Hangoo Kang"],"authorCount":10,"org":"Meta","affiliations":[],"orgNote":null,"date":"2026-10-01","hfDaily":"2026-10-02","primaryCategory":"cs.AI","categories":["cs.AI","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2610.01509","pdf":"https://arxiv.org/pdf/2610.01509","hf":"https://huggingface.co/papers/2610.01509","code":"https://github.com/changdaeoh/sharpening-tax","project":"https://changdaeoh.github.io/sharpening-tax/"},"metrics":{"hfUpvotes":111,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":30},"hot":61.42,"tags":["rl","agent"],"summary":{"zh":"发现 RL 后训练在 Agent 任务上有“锐化税”：预训练模型配轻量框架，多次尝试的覆盖率反而更高。","en":"Finds a 'sharpening tax': pre-trained LLMs with a light harness often beat post-trained ones on agentic coverage despite lower pass@1."},"disciplines":["mathematics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":86},{"id":"arxiv-2609.26780","arxiv":"2609.26780","title":"SpeakerMem-R1: Speaker-Centered Dual-Track Memory for Multi-Party Dialogue","authors":["Haobo Zheng","Tan Tang","Yan Chen","Weijie Wang","Yingcai Wu"],"authorCount":5,"org":"Zhejiang University","affiliations":[],"orgNote":null,"date":"2026-09-22","hfDaily":"2026-09-24","primaryCategory":"cs.CL","categories":["cs.CL","cs.AI","cs.IR","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2609.26780","pdf":"https://arxiv.org/pdf/2609.26780","hf":"https://huggingface.co/papers/2609.26780","code":"https://github.com/2022hpsk/SpeakerMemR1","project":"https://2022hpsk.github.io/SpeakerMemR1/"},"metrics":{"hfUpvotes":103,"hfComments":5,"citations":0,"citationsSource":"semantic_scholar","githubStars":114},"hot":60.99,"tags":["agent","rl"],"summary":{"zh":"SpeakerMem-R1 以说话人为中心构建双轨记忆，解决多人对话中“谁说了什么”的记忆难题。","en":"SpeakerMem-R1 builds speaker-centred dual-track memory for multi-party dialogue, tracking who said what about whom."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":87},{"id":"arxiv-2610.12772","arxiv":"2610.12772","title":"Reflex: Visually Grounded Reactive Humanoid Control","authors":["Taoyang Jia","Weikai Huang","Linxin Song","Jared Darlington","Jieyu Zhao","Yue Wang","Jiafei Duan","Zhongzheng Ren"],"authorCount":9,"org":"University of Washington","affiliations":[],"orgNote":null,"date":"2026-10-08","hfDaily":"2026-10-12","primaryCategory":"cs.RO","categories":["cs.RO"],"links":{"abs":"https://arxiv.org/abs/2610.12772","pdf":"https://arxiv.org/pdf/2610.12772","hf":"https://huggingface.co/papers/2610.12772","code":"https://github.com/ToberJ/reflex","project":"https://toberj.github.io/reflex/"},"metrics":{"hfUpvotes":154,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":34},"hot":60.97,"tags":["robotics","rl"],"summary":{"zh":"Reflex 让人形机器人凭第一视角视觉在一秒内全身反应，例如接住抛来的箱子。","en":"Reflex enables humanoids to react with whole-body control from egocentric vision, such as catching a thrown box."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":88},{"id":"arxiv-2610.20939","arxiv":"2610.20939","title":"MAI-Gym: An Open All-in-One Platform for GUI Agent Evaluation and RL Training across Mobile and Computer Environments","authors":["Panrong Tong","Quyu Kong","Pengxiang Li","Tianyu Xia","Hanzhang Zhou","Lei Wang","Chenglin Cai","Long Li"],"authorCount":15,"org":"Tongyi-MAI","affiliations":[],"orgNote":null,"date":"2026-10-14","hfDaily":"2026-10-16","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.20939","pdf":"https://arxiv.org/pdf/2610.20939","hf":"https://huggingface.co/papers/2610.20939","code":"https://github.com/Tongyi-MAI/MAI-Gym","project":"https://tongyi-mai.github.io/MAI-Gym/"},"metrics":{"hfUpvotes":149,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":38},"hot":60.91,"tags":["agent","rl","benchmark"],"summary":{"zh":"MAI-Gym 是开源 GUI Agent 一体化平台，统一评测、数据采集和在线 RL，最多可并行 2 万个环境。","en":"MAI-Gym is an open all-in-one GUI agent platform unifying evaluation, data collection and online RL at up to 20,000 parallel environments."},"disciplines":["mathematics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":89},{"id":"arxiv-2609.32722","arxiv":"2609.32722","title":"Scaling Properties of Same-Family On-Policy Distillation","authors":["Yuntai Bao","Qinfeng Li","Guoqing Jiang","Liwei Chen","Zhiheng Qin","Xuanping Li","Wenqi Zhang","Xuhong Zhang"],"authorCount":8,"org":"Zhejiang University","affiliations":[],"orgNote":null,"date":"2026-09-26","hfDaily":"2026-09-30","primaryCategory":"cs.LG","categories":["cs.LG"],"links":{"abs":"https://arxiv.org/abs/2609.32722","pdf":"https://arxiv.org/pdf/2609.32722","hf":"https://huggingface.co/papers/2609.32722","code":null,"project":"https://colored-dye.github.io/blog/2026/opd-scaling/"},"metrics":{"hfUpvotes":327,"hfComments":5,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":60.62,"tags":["efficiency","rl"],"summary":{"zh":"研究同族模型间在线策略蒸馏的扩展规律：训练早期准确率随 KL 距离近似线性上升。","en":"Studies scaling of same-family on-policy distillation, finding accuracy rises roughly linearly with KL distance early in training."},"disciplines":["mathematics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":90},{"id":"arxiv-2610.20344","arxiv":"2610.20344","title":"Doc2Harness: Benchmarking LLMs on Long-Horizon Harness Engineering of Coding Agents","authors":["Yuyao Ge","Shenghua Liu","Peizhong Gao","Chendi Ge","Chenxu Liu","Yiwei Wang","Lingrui Mei","Baolong Bi"],"authorCount":11,"org":"Institute of Computing Technology","affiliations":[],"orgNote":null,"date":"2026-10-14","hfDaily":"2026-10-15","primaryCategory":"cs.SE","categories":["cs.SE","cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.20344","pdf":"https://arxiv.org/pdf/2610.20344","hf":"https://huggingface.co/papers/2610.20344","code":"https://github.com/YuyaoGe/doc2harness","project":"https://geyuyao.com/doc2harness/"},"metrics":{"hfUpvotes":110,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":113},"hot":60.58,"tags":["agent","benchmark","code"],"summary":{"zh":"Doc2Harness 直接评测大模型能否按开发文档从零搭建长任务编程 Agent 的执行框架。","en":"Doc2Harness directly benchmarks whether LLMs can engineer a long-horizon coding-agent harness from a development document."},"disciplines":["software-engineering"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":91},{"id":"arxiv-2609.32704","arxiv":"2609.32704","title":"CoWindow Attention: Full Causal Coverage Is a Collective Property","authors":["Jingze Shi","Zhangyang Peng","Xianduo Li","Yanlin Qi","Xiaotian Lin","Haoxian Chen","Liangdong Wang","Guang Liu"],"authorCount":9,"org":"Beijing Academy of Artificial Intelligence","affiliations":[],"orgNote":null,"date":"2026-09-26","hfDaily":"2026-09-29","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2609.32704","pdf":"https://arxiv.org/pdf/2609.32704","hf":"https://huggingface.co/papers/2609.32704","code":"https://github.com/flash-algo/flash-sparse-attention","project":null},"metrics":{"hfUpvotes":67,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":763},"hot":60.44,"tags":["efficiency"],"summary":{"zh":"CoWA 注意力让不同 KV 头分摊远程历史，各头稀疏但合起来仍完整覆盖因果上下文。","en":"CoWA attention splits long-range history across KV heads so each is sparse yet together they cover the full causal context."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":92},{"id":"arxiv-2610.26625","arxiv":"2610.26625","title":"EgoXperience: Learning World-Grounded Egocentric Representations from Human Experience","authors":["Xinhao Ji","Hui En Pang","Hao Li","Kairui Hu","Jiner Wang","Mingze Sun","Ziang Cao","Yalun Dai"],"authorCount":15,"org":"Ropedia","affiliations":[],"orgNote":null,"date":"2026-10-19","hfDaily":"2026-10-20","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.26625","pdf":"https://arxiv.org/pdf/2610.26625","hf":"https://huggingface.co/papers/2610.26625","code":"https://github.com/Ropedia/EgoXperience","project":"https://ropedia.github.io/EgoXperience"},"metrics":{"hfUpvotes":170,"hfComments":2,"citations":null,"citationsSource":null,"githubStars":18},"hot":60.43,"tags":["vision","robotics","benchmark"],"summary":{"zh":"EgoXperience 用自动标注引擎联合重建第一视角视频中的人体运动、场景几何与手物交互，并配套基准。","en":"EgoXperience jointly reconstructs observer motion, scene geometry and hand-object interaction from egocentric video, with an auto-annotation engine and benchmark."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":93},{"id":"arxiv-2610.17201","arxiv":"2610.17201","title":"Multics: Generative Multi-Player World Model with Hierarchical Communication","authors":["Kaiyi Huang","Wangbo Yu","Yukun Huang","Jiwen Yu","Kunhao Liu","Donglin Yang","Wenbo Hu","Ying Shan"],"authorCount":9,"org":"ARC Lab, Tencent","affiliations":[],"orgNote":null,"date":"2026-10-12","hfDaily":"2026-10-14","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.17201","pdf":"https://arxiv.org/pdf/2610.17201","hf":"https://huggingface.co/papers/2610.17201","code":"https://github.com/TencentARC/Multics","project":"https://karine-huang.github.io/Multics/"},"metrics":{"hfUpvotes":74,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":100},"hot":60.36,"tags":["visual-gen"],"summary":{"zh":"Multics 用玩家-小组-全局的分层通信生成多人互动世界视频，玩家变多时开销可控。","en":"Multics generates multi-player world videos with hierarchical player-group-global communication that scales to more players."},"disciplines":["film-animation"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":94},{"id":"arxiv-2609.38288","arxiv":"2609.38288","title":"AREX-2: Advancing Self-Improving Agents through Long-Horizon Reflective Tasks","authors":["Hongjin Qian","Chaofan Li","Kun Luo","Wenqing Wei","Jianlyu Chen","Shuqi Lu","Yuyang Hu","Hongwang Xiao"],"authorCount":14,"org":"Beijing Academy of Artificial Intelligence","affiliations":[],"orgNote":null,"date":"2026-09-29","hfDaily":"2026-10-01","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2609.38288","pdf":"https://arxiv.org/pdf/2609.38288","hf":"https://huggingface.co/papers/2609.38288","code":"https://github.com/VectorSpaceLab/AREX-2","project":"https://github.com/VectorSpaceLab/AREX-2"},"metrics":{"hfUpvotes":134,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":39},"hot":60.35,"tags":["agent","rl"],"summary":{"zh":"AREX-2 用机器学习和算法编程中的长程改进轨迹训练，让 Agent 在测试时反复反思、不断改进答案。","en":"AREX-2 trains on long improvement trajectories from ML and algorithm tasks so agents iteratively refine solutions at test time."},"disciplines":["software-engineering"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":95},{"id":"arxiv-2610.06100","arxiv":"2610.06100","title":"From Traces to Agentic Worlds: Agentic Language World Models for Interactive Environment Simulation","authors":["Quanyu Long","Xiao Chen","Jianda Chen","Haozhen Zhang","Qisheng Hu","Jianzhu Bao","Wenya Wang"],"authorCount":7,"org":"Nanyang Technological University Singapore","affiliations":[],"orgNote":null,"date":"2026-10-05","hfDaily":"2026-10-09","primaryCategory":"cs.CL","categories":["cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.06100","pdf":"https://arxiv.org/pdf/2610.06100","hf":"https://huggingface.co/papers/2610.06100","code":"https://github.com/ruyue0001/trace2env","project":"https://quanyulong.net/trace2env/"},"metrics":{"hfUpvotes":119,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":75},"hot":60.35,"tags":["agent"],"summary":{"zh":"Trace2Env 不重建真实系统，而是用历史交互轨迹让一个“世界模型 Agent”扮演环境供其他 Agent 训练评测。","en":"Trace2Env lets a world-model agent role-play an environment reconstructed from historical traces for training and evaluating agents."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":96},{"id":"arxiv-2610.02205","arxiv":"2610.02205","title":"PROWBench: Do Video Models Render What the Program Specifies?","authors":["Zheng-Hui Huang","Guixu Lin","Yu-Ju Tsai","Jian-Kai Zhu","Fengbo Lan","Yu-Lun Liu","Yung-Yu Chuang","Kaipeng Zhang"],"authorCount":9,"org":"Alaya Lab","affiliations":[],"orgNote":null,"date":"2026-10-01","hfDaily":"2026-10-02","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.02205","pdf":"https://arxiv.org/pdf/2610.02205","hf":"https://huggingface.co/papers/2610.02205","code":"https://github.com/AlayaLab/PROWBench","project":"https://alaya-lab.github.io/PROWBench"},"metrics":{"hfUpvotes":74,"hfComments":2,"citations":3,"citationsSource":"semantic_scholar","githubStars":49},"hot":60.15,"tags":["visual-gen","benchmark"],"summary":{"zh":"PROWBench 检验视频模型是否按程序指定的世界事件准确渲染，面向下一代游戏引擎。","en":"PROWBench tests whether video models faithfully render program-specified world events, aimed at next-gen game engines."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":97},{"id":"arxiv-2609.39045","arxiv":"2609.39045","title":"RSIGame: Autonomous Agentic Game Development with Recursive Self-improvement","authors":["Wenyi Wu","Minghao Fu","Jieyu You","Kun Zhou","Siqi Liu","Aayush Salvi","Yiheng Lin","Ce Zhang"],"authorCount":13,"org":"RSIGame","affiliations":[],"orgNote":null,"date":"2026-09-30","hfDaily":"2026-10-01","primaryCategory":"cs.CL","categories":["cs.CL","cs.GT","cs.LG","cs.MA"],"links":{"abs":"https://arxiv.org/abs/2609.39045","pdf":"https://arxiv.org/pdf/2609.39045","hf":"https://huggingface.co/papers/2609.39045","code":"https://github.com/WenyiWU0111/RSIGame","project":"https://huggingface.co/spaces/RSIGame/rsigame-page"},"metrics":{"hfUpvotes":97,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":146},"hot":59.97,"tags":["agent","code"],"summary":{"zh":"RSIGame 用局部和全局两层循环递归改进 Agent 生成的游戏，减少只对少数测试过拟合的问题。","en":"RSIGame improves agent-built games with local and global recursive loops to avoid overfitting a few test cases."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":98},{"id":"arxiv-2610.17665","arxiv":"2610.17665","title":"RoTA-Bench: Evaluating 3D Spatial Reasoning through Interactive Projection Alignment","authors":["Qingchuan Ma","Yuexiao Ma","Xiang Meng","Yixuan Zou","Tianyu Xie","Wang Chen","Yuhui Zeng","Yuhang Wu"],"authorCount":10,"org":"MAC-AutoML","affiliations":[],"orgNote":null,"date":"2026-10-13","hfDaily":"2026-10-16","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.17665","pdf":"https://arxiv.org/pdf/2610.17665","hf":"https://huggingface.co/papers/2610.17665","code":"https://github.com/MAC-AutoML/RoTA-Bench","project":"https://www.modelscope.cn/studios/san2333333/rotabench-anon?mode=full&app_path=%2Fportal"},"metrics":{"hfUpvotes":206,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":6},"hot":59.84,"tags":["multimodal","benchmark","reasoning"],"summary":{"zh":"RoTA-Bench 通过交互式旋转物体对齐剪影，评测视觉语言模型的 3D 空间推理。","en":"RoTA-Bench evaluates VLMs' 3D spatial reasoning by having them rotate objects to match target silhouettes interactively."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":99},{"id":"arxiv-2610.28578","arxiv":"2610.28578","title":"Sage-A1.5: From Long-Horizon Execution to Scientific Exploration with Capability Composition","authors":["Wenjie Lou","Tianshuo Peng","Shiyang Feng","Shengji Tang","Runmin Ma","Ruizhe Chen","Jinxin Shi","Haoyang Peng"],"authorCount":99,"org":"Intern Science","affiliations":[],"orgNote":null,"date":"2026-10-20","hfDaily":"2026-10-21","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.28578","pdf":"https://arxiv.org/pdf/2610.28578","hf":"https://huggingface.co/papers/2610.28578","code":"https://github.com/InternScience/Agents-A1","project":"https://alpha-innovator.github.io/Sage-A1.5/"},"metrics":{"hfUpvotes":64,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":574},"hot":59.71,"tags":["agent","ai4science"],"summary":{"zh":"Sage-A1.5 是 35B MoE 智能体模型，新增子智能体并行协作、过程自检和开放式科学探索能力。","en":"Sage-A1.5 is a 35B MoE agentic model adding subagent orchestration, self-verification and open-ended scientific exploration."},"disciplines":["economics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":100},{"id":"arxiv-2610.14766","arxiv":"2610.14766","title":"Last Secure Code Benchmark: Is the Code You Vibe Secure?","authors":["Tianyi Wu","Chao Ge","Peng Chen","Fengji Zhang","Yongxing Wu","Yuchong Xie","Jian Li","Chaoyuan Peng"],"authorCount":15,"org":"Qwen","affiliations":[],"orgNote":null,"date":"2026-10-10","hfDaily":"2026-10-14","primaryCategory":"cs.CR","categories":["cs.CR","cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.14766","pdf":"https://arxiv.org/pdf/2610.14766","hf":"https://huggingface.co/papers/2610.14766","code":"https://github.com/alibaba/last-secure-code-benchmark","project":"https://last-secure-code-benchmark.github.io/"},"metrics":{"hfUpvotes":76,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":56},"hot":59.63,"tags":["code","safety","benchmark"],"summary":{"zh":"基于 150 个真实漏洞构建 450 个任务，检验编程 Agent“氛围编程”写出的代码是否安全。","en":"Builds 450 tasks from 150 real vulnerabilities to check whether code written by vibe-coding agents is secure."},"disciplines":["software-engineering","cybersecurity"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":101},{"id":"arxiv-2610.07987","arxiv":"2610.07987","title":"VisionWeave: Weaving Elastic Visual Representations as a Native Capability of MLLMs","authors":["Yuan Feng","Qize Yang","Ruizhe Chen","Sibo Song","Haolin He","Muzhi Zhu","Zihan Liu","Yunfei Chu"],"authorCount":12,"org":"Qwen","affiliations":[],"orgNote":null,"date":"2026-10-06","hfDaily":"2026-10-12","primaryCategory":"cs.CV","categories":["cs.CV","cs.AI","cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.07987","pdf":"https://arxiv.org/pdf/2610.07987","hf":"https://huggingface.co/papers/2610.07987","code":"https://github.com/FFY0/VisionWeave","project":null},"metrics":{"hfUpvotes":70,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":91},"hot":59.58,"tags":["multimodal","efficiency"],"summary":{"zh":"VisionWeave 让多模态模型端到端学会按内容分配视觉 token 粒度，细节处多给、简单处少给。","en":"VisionWeave lets MLLMs learn end-to-end where and how finely to spend visual tokens, adapting to image content."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":102},{"id":"arxiv-2610.12126","arxiv":"2610.12126","title":"SuperNav: An Agentic Navigation System for Any Task in Any Scene","authors":["Jinkai Zhang","Jingyi Xu","Yuanhong Yu","Jiarui Guo","Ruizhen Hu","Hujun Bao","Xiaowei Zhou","Sida Peng"],"authorCount":8,"org":"zju3dv","affiliations":[],"orgNote":null,"date":"2026-10-08","hfDaily":"2026-10-09","primaryCategory":"cs.RO","categories":["cs.RO","cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.12126","pdf":"https://arxiv.org/pdf/2610.12126","hf":"https://huggingface.co/papers/2610.12126","code":"https://github.com/zju3dv/SuperNav","project":"https://zju3dv.github.io/SuperNav/"},"metrics":{"hfUpvotes":86,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":198},"hot":59.54,"tags":["robotics","agent"],"summary":{"zh":"SuperNav 让多模态大模型负责理解请求和决策、把运动交给导航工具，实现任意场景任意任务的导航。","en":"SuperNav lets an MLLM interpret requests and decide while delegating motion to navigation tools, generalising across tasks and scenes."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":103},{"id":"arxiv-2610.04198","arxiv":"2610.04198","title":"ALoDLM: Adaptively Looped Diffusion Language Models","authors":["Liancheng Fang","Zhuowei Li","Youngeun Kim","Tianchen Zhao","Rajat Koner","Jiaye Wu","Linghan Xu","Xuanbai Chen"],"authorCount":13,"org":"Amazon","affiliations":[],"orgNote":null,"date":"2026-10-03","hfDaily":"2026-10-06","primaryCategory":"cs.AI","categories":["cs.AI","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2610.04198","pdf":"https://arxiv.org/pdf/2610.04198","hf":"https://huggingface.co/papers/2610.04198","code":"https://github.com/amazon-science/ALoDLM","project":"https://alo-dlm.github.io/"},"metrics":{"hfUpvotes":80,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":49},"hot":59.37,"tags":["llm","efficiency"],"summary":{"zh":"ALoDLM 在扩散语言模型中按 token 难度自适应循环计算，缩小与自回归模型的质量差距。","en":"ALoDLM adds token-adaptive latent recurrence to diffusion language models, narrowing the quality gap with autoregressive models."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":104},{"id":"arxiv-2609.26550","arxiv":"2609.26550","title":"JEV-as-a-Judge: Accept When Confident, Escalate When Unsure","authors":["Yubo Li","Yidi Miao","Ramayya Krishnan","Rema Padman"],"authorCount":4,"org":"Carnegie Mellon University","affiliations":[],"orgNote":null,"date":"2026-09-22","hfDaily":"2026-09-23","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2609.26550","pdf":"https://arxiv.org/pdf/2609.26550","hf":"https://huggingface.co/papers/2609.26550","code":null,"project":"https://yubol-bobo.github.io/jev-as-a-judge/"},"metrics":{"hfUpvotes":44,"hfComments":4,"citations":40,"citationsSource":"semantic_scholar","githubStars":null},"hot":59.34,"tags":["benchmark","efficiency"],"summary":{"zh":"用只做决策的轻量评审模型做第一轮打分，没把握时再升级给大模型，成本仅为后者的零头。","en":"Uses a decision-only judge as a cheap first pass and escalates to stronger judges only when unsure."},"disciplines":["economics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":105},{"id":"arxiv-2609.38839","arxiv":"2609.38839","title":"FrameMorrow: Future-guided Frame Selection with Prospective Tokens for Long-Horizon Video Generation","authors":["Bo Yin","Xiaobin Hu","Jiaqi Zhao","Shuicheng Yan"],"authorCount":4,"org":"National University of Singapore","affiliations":[],"orgNote":null,"date":"2026-09-30","hfDaily":"2026-10-05","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2609.38839","pdf":"https://arxiv.org/pdf/2609.38839","hf":"https://huggingface.co/papers/2609.38839","code":"https://github.com/YinBo0927/FrameMorrow","project":"https://yinbo0927.github.io/FrameMorrow/"},"metrics":{"hfUpvotes":124,"hfComments":4,"citations":0,"citationsSource":"semantic_scholar","githubStars":30},"hot":59.28,"tags":["visual-gen","efficiency"],"summary":{"zh":"FrameMorrow 用“前瞻 token”按未来需要挑选历史帧，提升长视频生成的效率和一致性。","en":"FrameMorrow selects history frames by future relevance via prospective tokens for efficient, consistent long video generation."},"disciplines":["film-animation"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":106},{"id":"arxiv-2610.26549","arxiv":"2610.26549","title":"Holi-World: Building Holistic World Intelligence from Human-Ego Experience","authors":["Hao Li","Xiyuan Zheng","Yuanyuan Gao","Yalun Dai","Xinhao Ji","Xuan-Li Chen","Jing Zhao","Fangzhou Hong"],"authorCount":10,"org":"Ropedia","affiliations":[],"orgNote":null,"date":"2026-10-19","hfDaily":"2026-10-20","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.26549","pdf":"https://arxiv.org/pdf/2610.26549","hf":"https://huggingface.co/papers/2610.26549","code":"https://github.com/Ropedia/holi-world","project":"https://ropedia.github.io/holi-world"},"metrics":{"hfUpvotes":156,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":13},"hot":59.24,"tags":["vision","benchmark"],"summary":{"zh":"Holi-World 是从人类第一视角经验构建世界智能的完整生态，覆盖采集、表征、数据构建、训练和评测。","en":"Holi-World is an ecosystem for learning world intelligence from human egocentric experience, from capture to evaluation."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":107},{"id":"arxiv-2610.12818","arxiv":"2610.12818","title":"LeWAM: End-to-End World and Action Modeling with JEPAs","authors":["Minghao Fu","Tavis Siebert","Eryk Halicki","Randall Balestriero"],"authorCount":4,"org":"LeWAM","affiliations":[],"orgNote":null,"date":"2026-10-08","hfDaily":"2026-10-14","primaryCategory":"cs.RO","categories":["cs.RO","cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.12818","pdf":"https://arxiv.org/pdf/2610.12818","hf":"https://huggingface.co/papers/2610.12818","code":"https://github.com/MinghaoFu/lewam","project":"https://le-wam.github.io/"},"metrics":{"hfUpvotes":75,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":287},"hot":59.11,"tags":["robotics","vision"],"summary":{"zh":"LeWAM 是端到端的 JEPA 世界-动作模型，同一模型既能直接执行策略也能高效规划。","en":"LeWAM is an end-to-end JEPA world-action model that supports both direct policy execution and efficient planning."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":108},{"id":"arxiv-2610.08430","arxiv":"2610.08430","title":"NeMo-DCR: Bit-Exact Delta-Compressed Refit for Scalable Agentic RL at Trillion-Parameter Scale","authors":["Songlin Jiang","Zhiyu Li","Terry Kong","Yu Yao","Youngeun Kwon","Bernard Nguyen","Ashwath Aithal","Mario Di Francesco"],"authorCount":8,"org":"NVIDIA","affiliations":[],"orgNote":null,"date":"2026-10-06","hfDaily":"2026-10-07","primaryCategory":"cs.DC","categories":["cs.DC","cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.08430","pdf":"https://arxiv.org/pdf/2610.08430","hf":"https://huggingface.co/papers/2610.08430","code":"https://github.com/NVIDIA-NeMo/RL","project":"https://docs.nvidia.com/nemo/rl/nightly/guides/refit.html"},"metrics":{"hfUpvotes":30,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":2076},"hot":59.08,"tags":["efficiency","rl"],"summary":{"zh":"NeMo-DCR 只传输每步变化的约 1% 权重并保证比特级一致，加速万亿参数 Agent RL 的权重同步。","en":"NeMo-DCR syncs only the ~1% of weights that change per step, bit-exactly, to speed trillion-parameter agentic RL refits."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":109},{"id":"arxiv-2609.38177","arxiv":"2609.38177","title":"Imagine3D-LLM: Teaching MLLMs to Imagine 3D Scenes Before Answering","authors":["Jaewoo Jung","Hyeonseo Yu","Honggyu An","Jisang Han","Mungyeom Kim","Minkyeong Jeon","Heeseong Shin","Wonjun Moon"],"authorCount":13,"org":"KAIST AI","affiliations":[],"orgNote":null,"date":"2026-09-29","hfDaily":"2026-10-01","primaryCategory":"cs.CV","categories":["cs.CV","cs.CL"],"links":{"abs":"https://arxiv.org/abs/2609.38177","pdf":"https://arxiv.org/pdf/2609.38177","hf":"https://huggingface.co/papers/2609.38177","code":"https://github.com/cvlab-kaist/Imagine3D-LLM","project":"https://cvlab-kaist.github.io/Imagine3D-LLM/"},"metrics":{"hfUpvotes":78,"hfComments":2,"citations":2,"citationsSource":"semantic_scholar","githubStars":38},"hot":58.9,"tags":["multimodal","reasoning"],"summary":{"zh":"Imagine3D-LLM 教多模态模型先在脑中“想象”3D 场景再回答多视角空间问题。","en":"Imagine3D-LLM teaches MLLMs to imagine a 3D scene before answering multi-view spatial questions."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":110},{"id":"arxiv-2610.11959","arxiv":"2610.11959","title":"MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement","authors":["Xiaomi LLM-Core Team"," :","Zongming Qiao","Ziyue Hua","Zirui Ou","Zihao Yue","Zihan Jiang","Zhuo Huang"],"authorCount":150,"org":"Xiaomi MiMo","affiliations":[],"orgNote":null,"date":"2026-10-08","hfDaily":"2026-10-09","primaryCategory":"cs.CL","categories":["cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.11959","pdf":"https://arxiv.org/pdf/2610.11959","hf":"https://huggingface.co/papers/2610.11959","code":null,"project":null},"metrics":{"hfUpvotes":92,"hfComments":2,"citations":6,"citationsSource":"semantic_scholar","githubStars":null},"hot":58.76,"tags":["llm","rl","multimodal"],"summary":{"zh":"小米 MiMo-V2.6 全模态模型技术报告，重点是从多个维度扩大强化学习算力以实现自我提升。","en":"Technical report of Xiaomi's omni-modal MiMo-V2.6, focused on scaling RL compute along several dimensions toward self-improvement."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":111},{"id":"arxiv-2610.00972","arxiv":"2610.00972","title":"VeriHarness: Scaling Agentic Verification for Long-Horizon Tasks","authors":["Caiqi Zhang","Rujun Han","Zifeng Wang","Zoey CuiZhu","Nigel Collier","Tomas Pfister","Chen-Yu Lee"],"authorCount":7,"org":"Google","affiliations":[],"orgNote":null,"date":"2026-10-01","hfDaily":"2026-10-05","primaryCategory":"cs.AI","categories":["cs.AI","cs.MA"],"links":{"abs":"https://arxiv.org/abs/2610.00972","pdf":"https://arxiv.org/pdf/2610.00972","hf":"https://huggingface.co/papers/2610.00972","code":"https://github.com/google-research/veriharness","project":"https://veriharness.com/"},"metrics":{"hfUpvotes":63,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":75},"hot":58.5,"tags":["agent","reasoning"],"summary":{"zh":"VeriHarness 把基础模型变成验证 Agent，利用多次采样间的分歧找出可信结论，无需参考答案。","en":"VeriHarness turns the base model into a verification agent that uses disagreement across samples to decide which claims to trust."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":112},{"id":"arxiv-2610.19789","arxiv":"2610.19789","title":"More Methods, Better Agents? A Reality Check on Self-Distillation in Agentic Reinforcement Learning","authors":["Linhao Zhong","Bozhen Fang","Xuehao Wang","Duochao Shi","Zekai Luo","Guangming Yao","Yi Yuan","Hao Chen"],"authorCount":10,"org":"Zhejiang University","affiliations":[],"orgNote":null,"date":"2026-10-14","hfDaily":"2026-10-15","primaryCategory":"cs.CL","categories":["cs.CL","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2610.19789","pdf":"https://arxiv.org/pdf/2610.19789","hf":"https://huggingface.co/papers/2610.19789","code":"https://github.com/aim-uofa/Self-Distillation-AgentRL","project":"https://aim-uofa.github.io/Self-Distillation-AgentRL/"},"metrics":{"hfUpvotes":92,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":83},"hot":58.48,"tags":["agent","rl","benchmark"],"summary":{"zh":"在统一条件下对比十种 Agent 强化学习自蒸馏方法与 GRPO，检查它们的真实收益和成本。","en":"Compares ten agentic-RL self-distillation methods against GRPO under identical conditions to check real gains and costs."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":113},{"id":"arxiv-2610.09326","arxiv":"2610.09326","title":"VIS-Ground: Video Interactive Storytelling with Contextual Grounding","authors":["Bingxuan Li","Yiwen Song","Xueqing Wu","Yanzhou Pan","Yang Li","Kuang Su","Jingyun Liu","Sebastian Ko"],"authorCount":13,"org":"Google","affiliations":[],"orgNote":null,"date":"2026-10-07","hfDaily":"2026-10-13","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.09326","pdf":"https://arxiv.org/pdf/2610.09326","hf":"https://huggingface.co/papers/2610.09326","code":"https://github.com/vis-ground/code","project":"https://vis-ground.github.io"},"metrics":{"hfUpvotes":102,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":12},"hot":58.41,"tags":["visual-gen"],"summary":{"zh":"VIS-Ground 让观众在交互式视频故事中随时干预，同时保持与素材和已生成内容的一致。","en":"VIS-Ground lets viewers steer interactive video stories while staying consistent with the source and rendered history."},"disciplines":["film-animation"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":114},{"id":"arxiv-2610.01939","arxiv":"2610.01939","title":"Fewer Tokens, Better Action: GPT-6 Astra Robot Agents with 14% Higher Success Rate but 65% Fewer Tokens","authors":["Ruiyang Si","Jianxin Bi","Shunyu Yang","Rui Ni","Wenbo Huang","Qiang Wang","Shulong Jiang","Yintianrun Zhang"],"authorCount":15,"org":"DAGroup-PKU","affiliations":[],"orgNote":null,"date":"2026-10-01","hfDaily":"2026-10-02","primaryCategory":"cs.CV","categories":["cs.CV","cs.RO"],"links":{"abs":"https://arxiv.org/abs/2610.01939","pdf":"https://arxiv.org/pdf/2610.01939","hf":"https://huggingface.co/papers/2610.01939","code":"https://github.com/DAGroup-PKU/PyRUA-Lean","project":"https://dagroup-pku.github.io/PyRUA-Lean/"},"metrics":{"hfUpvotes":52,"hfComments":2,"citations":5,"citationsSource":"semantic_scholar","githubStars":51},"hot":58.4,"tags":["robotics","agent","efficiency"],"summary":{"zh":"让机器人 Agent 把动作原语组合成可本地重试的 Python 代码块，成功率更高、token 用量更少。","en":"Has robot agents compose primitives into locally retried Python cells, raising success while cutting token use."},"disciplines":["mathematics","robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":115},{"id":"arxiv-2610.01780","arxiv":"2610.01780","title":"RealCompanion: Benchmarking Human Understanding from Reasoning over Longitudinal Real-World Conversations","authors":["Arman Behnam","Sunglyoung Kim","Jiayi Yu","Eric Huang","Liangwei Yang"],"authorCount":5,"org":"Quis Lab","affiliations":[],"orgNote":null,"date":"2026-10-01","hfDaily":"2026-10-05","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.01780","pdf":"https://arxiv.org/pdf/2610.01780","hf":"https://huggingface.co/papers/2610.01780","code":null,"project":null},"metrics":{"hfUpvotes":287,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":58.28,"tags":["benchmark","agent"],"summary":{"zh":"RealCompanion 公开 10 段长达 120 天、共 27218 条的真实 AI 陪伴对话，检验模型对人的长期理解。","en":"RealCompanion releases ten real AI-companion relationships (27,218 messages over up to 120 days) to test long-term understanding of a person."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":116},{"id":"arxiv-2610.02826","arxiv":"2610.02826","title":"Scaling Trajectories for Complex Tasks through Recursive Self-Rewrite","authors":["Zongxia Li","Yucheng Shi","Zhongzhi Li","Junyao Yang","Ruhan Wang","Chengsong Huang","Fuxiao Liu","Haitao Mi"],"authorCount":10,"org":"Tencent Hunyuan","affiliations":[],"orgNote":null,"date":"2026-10-02","hfDaily":"2026-10-05","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.02826","pdf":"https://arxiv.org/pdf/2610.02826","hf":"https://huggingface.co/papers/2610.02826","code":null,"project":null},"metrics":{"hfUpvotes":105,"hfComments":2,"citations":3,"citationsSource":"semantic_scholar","githubStars":null},"hot":57.83,"tags":["agent","rl"],"summary":{"zh":"递归自我重写：在各种专用框架下找到成功解，再改写成通用框架下可用的训练轨迹。","en":"Recursive Self-Rewrite finds solutions under specialised harnesses and rewrites them as training trajectories for a general harness."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":117},{"id":"arxiv-2610.07591","arxiv":"2610.07591","title":"Recurrent Looped Transformer","authors":["Yifan Zhang","Jichen Feng","Shihan Qin","Sun-Yuan Kung","Mengdi Wang"],"authorCount":5,"org":"Princeton University","affiliations":[],"orgNote":null,"date":"2026-10-06","hfDaily":"2026-10-08","primaryCategory":"cs.CL","categories":["cs.CL","cs.AI","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2610.07591","pdf":"https://arxiv.org/pdf/2610.07591","hf":"https://huggingface.co/papers/2610.07591","code":"https://github.com/yifanzhang-pro/recurrent-looped-tranformer","project":"https://github.com/yifanzhang-pro/recurrent-looped-tranformer"},"metrics":{"hfUpvotes":32,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":945},"hot":57.74,"tags":["llm","reasoning"],"summary":{"zh":"循环 Transformer 新结构 RLT 让计算深度随序列增长，奇偶校验等任务可泛化到远超训练长度。","en":"The Recurrent Looped Transformer lets compute depth grow with sequence length, generalising tasks like parity far beyond training length."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":118},{"id":"arxiv-2609.39982","arxiv":"2609.39982","title":"Mid-Harness: Scaling Actions Between Model and Harness for Terminal Agents","authors":["Minki Kang","Ryo Hachiuma","Shaokun Zhang","Subhashree Radhakrishnan","Yonggan Fu","Jindong Jiang","Mingjie Liu","Ehsan Hosseini-Asl"],"authorCount":11,"org":"NVIDIA","affiliations":[],"orgNote":null,"date":"2026-09-30","hfDaily":"2026-10-01","primaryCategory":"cs.CL","categories":["cs.CL","cs.AI","cs.MA"],"links":{"abs":"https://arxiv.org/abs/2609.39982","pdf":"https://arxiv.org/pdf/2609.39982","hf":"https://huggingface.co/papers/2609.39982","code":null,"project":"https://byungkwanlee.github.io/MidHarness-page/"},"metrics":{"hfUpvotes":120,"hfComments":5,"citations":1,"citationsSource":"semantic_scholar","githubStars":null},"hot":57.42,"tags":["agent","efficiency"],"summary":{"zh":"Mid-Harness 在模型与执行框架之间先采样并验证候选命令再执行，提高终端 Agent 的可靠性。","en":"Mid-Harness samples and verifies candidate actions before execution, improving terminal-agent reliability."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":119},{"id":"arxiv-2610.13521","arxiv":"2610.13521","title":"Beyond Domain-Routed Supervision: Cross-Teacher Guarding for Multi-Teacher On-Policy Distillation","authors":["Shi-Yu Tian","Yong Liu","Ximan Liu","Tianrui Zhu","Sule Bai","Xinwei Yang","Ruiqi Liu","Kun-Yang Yu"],"authorCount":13,"org":"jingdong","affiliations":[],"orgNote":null,"date":"2026-10-09","hfDaily":"2026-10-15","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.13521","pdf":"https://arxiv.org/pdf/2610.13521","hf":"https://huggingface.co/papers/2610.13521","code":null,"project":null},"metrics":{"hfUpvotes":251,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":57.37,"tags":["efficiency"],"summary":{"zh":"在多教师在线蒸馏中加入跨教师“护栏”，减少不同领域教师之间的梯度冲突。","en":"Adds cross-teacher guarding to multi-teacher on-policy distillation to reduce gradient conflicts between domain teachers."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":120},{"id":"arxiv-2609.22966","arxiv":"2609.22966","title":"Transferring the Intelligence of VLMs to Robotic Control","authors":["Meng-Hao Guo","Zhe-Han Mo","Jia-Jun Wang","Yi Zhang","Kejin Wang","Yi-Xuan Deng","Jia-Peng Zhang","Yongming Rao"],"authorCount":9,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-09-19","hfDaily":"2026-09-22","primaryCategory":"cs.RO","categories":["cs.RO"],"links":{"abs":"https://arxiv.org/abs/2609.22966","pdf":"https://arxiv.org/pdf/2609.22966","hf":"https://huggingface.co/papers/2609.22966","code":null,"project":null},"metrics":{"hfUpvotes":89,"hfComments":4,"citations":11,"citationsSource":"semantic_scholar","githubStars":null},"hot":57.35,"tags":["robotics","multimodal"],"summary":{"zh":"RoboDawn 用一组简单的离散平移、旋转和夹爪命令，让视觉语言 Agent 直接控制机器人，检验数字智能能否迁移到物理世界。","en":"RoboDawn exposes robot control as simple discrete commands to an agentic VLM, testing whether digital intelligence transfers to the physical world."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":121},{"id":"arxiv-2610.02193","arxiv":"2610.02193","title":"Hierarchical Continuous Diffusion Language Models","authors":["Hui Ren","Zihan Li","Chang Liu","Huidong Liu","Alexander Schwing"],"authorCount":5,"org":"University of Illinois at Urbana-Champaign","affiliations":[],"orgNote":null,"date":"2026-10-01","hfDaily":"2026-10-02","primaryCategory":"cs.CL","categories":["cs.CL","cs.AI","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2610.02193","pdf":"https://arxiv.org/pdf/2610.02193","hf":"https://huggingface.co/papers/2610.02193","code":"https://github.com/rhfeiyang/HC-DLM","project":"https://hc-dlm.github.io/"},"metrics":{"hfUpvotes":92,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":62},"hot":57.33,"tags":["llm"],"summary":{"zh":"分层连续扩散语言模型在连续状态去噪的同时约束到合法 token，改善并行解码的一致性。","en":"Hierarchical continuous diffusion LMs tie the denoised continuous state to valid tokens, improving coherence of parallel decoding."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":122},{"id":"arxiv-2610.18861","arxiv":"2610.18861","title":"Hybrid Reasoning Agents","authors":["Jiaru Zou","Yunzhe Qi","Sirui Chen","Federico Bianchi","Shang Zhu","Pan Lu","Jingrui He","James Zou"],"authorCount":8,"org":"Stanford University","affiliations":[],"orgNote":null,"date":"2026-10-13","hfDaily":"2026-10-15","primaryCategory":"cs.CL","categories":["cs.CL","cs.AI","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2610.18861","pdf":"https://arxiv.org/pdf/2610.18861","hf":"https://huggingface.co/papers/2610.18861","code":"https://github.com/togethercomputer/HybridAgent","project":null},"metrics":{"hfUpvotes":38,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":386},"hot":57.18,"tags":["reasoning","agent"],"summary":{"zh":"混合推理 Agent 在潜空间推理和文本交互之间动态切换，兼顾效率与工具调用。","en":"Hybrid Reasoning Agents switch dynamically between latent reasoning and text interaction, combining efficiency with tool access."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":123},{"id":"arxiv-2609.37372","arxiv":"2609.37372","title":"Think Before You Score: Thinking Reward Model for Visual Generation","authors":["Xuehai Bai","Zhenchen Tang","Yang Shi","Dianyi Wang","Tengfei Liu","Wanshun Su","Xuanyu Zhu","Ruohui Wang"],"authorCount":12,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-09-29","hfDaily":"2026-09-30","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2609.37372","pdf":"https://arxiv.org/pdf/2609.37372","hf":"https://huggingface.co/papers/2609.37372","code":"https://github.com/bxhsort/Thinking_Reward_Model","project":"https://bxhsort.github.io/Thinking-Reward-Model/"},"metrics":{"hfUpvotes":104,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":35},"hot":57.15,"tags":["visual-gen","alignment"],"summary":{"zh":"思考型奖励模型先为每个示例制定评分要点再打分，给视觉生成提供更细的奖励信号。","en":"The Thinking Reward Model first sets case-specific rubrics and then scores, giving finer rewards for visual generation."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":124},{"id":"arxiv-2609.24983","arxiv":"2609.24983","title":"onPanda: Efficient Annotation of On-Policy Alignment Data for LLMs and Agents via Token-Level Correction","authors":["Lei Yang","Mengyin Liu","Jia Wang","Hangyu Guo","Liang Zhao","Zheng Ge","Kang An","Binxing Jiao"],"authorCount":12,"org":"StepFun","affiliations":[],"orgNote":null,"date":"2026-09-21","hfDaily":"2026-09-22","primaryCategory":"cs.CL","categories":["cs.CL","cs.HC","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2609.24983","pdf":"https://arxiv.org/pdf/2609.24983","hf":"https://huggingface.co/papers/2609.24983","code":"https://github.com/on-panda/on-panda","project":"https://on-panda.github.io/research/"},"metrics":{"hfUpvotes":55,"hfComments":4,"citations":0,"citationsSource":"semantic_scholar","githubStars":63},"hot":57.06,"tags":["alignment","agent"],"summary":{"zh":"onPanda 是对齐数据标注工具：找到第一个错 token 改掉后让模型续写，低成本标注 Agent 轨迹。","en":"onPanda is an annotation tool where annotators fix the first bad token and let the model continue, cheaply producing alignment data."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":125},{"id":"arxiv-2610.18072","arxiv":"2610.18072","title":"LensErase: A Foundation Model for Erase Anything","authors":["Yunuo Chen","Xiuyu Wu","Qinhong Yang","Jinjing Zhao","Yanchen Dong","Sirui Zhang","Dong Chen","Fangyun Wei"],"authorCount":14,"org":"Microsoft","affiliations":[],"orgNote":null,"date":"2026-10-13","hfDaily":"2026-10-15","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.18072","pdf":"https://arxiv.org/pdf/2610.18072","hf":"https://huggingface.co/papers/2610.18072","code":"https://github.com/LensFamily/LensErase","project":"https://lensfamily.github.io/LensErase/"},"metrics":{"hfUpvotes":52,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":103},"hot":56.96,"tags":["visual-gen"],"summary":{"zh":"LensErase 是通用“擦除一切”模型，基于 1 亿对数据训练，可移除物体、文字、水印及其影子等效果。","en":"LensErase is a general erase-anything model trained on 100M pairs, removing objects, text, watermarks and their visual effects."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":126},{"id":"arxiv-2609.38597","arxiv":"2609.38597","title":"PixelUMM: Encoder-Free Unified Image and Video Understanding and Generation","authors":["Cong Wei","Xuanchi Ren","Bryan Chu","Weiming Ren","Huan Ling","Jiahui Huang","Laura Leal-Taixé","Sanja Fidler"],"authorCount":11,"org":"NVIDIA","affiliations":[],"orgNote":null,"date":"2026-09-29","hfDaily":"2026-10-01","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2609.38597","pdf":"https://arxiv.org/pdf/2609.38597","hf":"https://huggingface.co/papers/2609.38597","code":"https://github.com/nv-tlabs/PixelUMM","project":"https://nv-tlabs.github.io/PixelUMM/"},"metrics":{"hfUpvotes":34,"hfComments":1,"citations":1,"citationsSource":"semantic_scholar","githubStars":227},"hot":56.94,"tags":["multimodal","visual-gen"],"summary":{"zh":"PixelUMM 不用视觉编码器，直接在像素空间统一图像和视频的理解与生成。","en":"PixelUMM unifies image and video understanding and generation directly in pixel space without a visual encoder."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":127},{"id":"arxiv-2609.23863","arxiv":"2609.23863","title":"Grounded Action Model: 3D Grounding as a Foundation for Robotics","authors":["Gehao Zhang","Weikai Huang","Shailesh Shailesh","Yiyan Peng","Jiafei Duan","Ranjay Krishna"],"authorCount":6,"org":"National University of Singapore","affiliations":[],"orgNote":null,"date":"2026-09-20","hfDaily":"2026-09-22","primaryCategory":"cs.RO","categories":["cs.RO"],"links":{"abs":"https://arxiv.org/abs/2609.23863","pdf":"https://arxiv.org/pdf/2609.23863","hf":"https://huggingface.co/papers/2609.23863","code":"https://github.com/GehaoZhang6/Grounded-Action-Model","project":"https://grounded-action-model.github.io/"},"metrics":{"hfUpvotes":88,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":51},"hot":56.84,"tags":["robotics"],"summary":{"zh":"提出以 3D 定位为基础的机器人基础模型 GAM，可用语言、点或框指定要操作的物体。","en":"Proposes Grounded Action Models, robot foundation models built on 3D grounding and promptable with language, points or boxes."},"disciplines":["robotics","film-animation"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":128},{"id":"arxiv-2610.26802","arxiv":"2610.26802","title":"Scaling Parameter-Efficient Fine-Tuning for Brain Encoding","authors":["Zhuofan Josh Ying","Guanglong Sun","Nolan Ho","Kieran Liu","Rui Yang","Andrew Chen","Nikolaus Kriegeskorte","Pony Ma"],"authorCount":9,"org":"Mind Lab","affiliations":[],"orgNote":null,"date":"2026-10-19","hfDaily":"2026-10-20","primaryCategory":"cs.LG","categories":["cs.LG"],"links":{"abs":"https://arxiv.org/abs/2610.26802","pdf":"https://arxiv.org/pdf/2610.26802","hf":"https://huggingface.co/papers/2610.26802","code":null,"project":null},"metrics":{"hfUpvotes":237,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":56.8,"tags":["ai4science","efficiency"],"summary":{"zh":"用参数高效微调把最大 104B 的多模态模型适配到脑编码任务，更准确地预测大脑对刺激的反应。","en":"Uses parameter-efficient fine-tuning to adapt multimodal backbones of up to 104B for brain encoding, better predicting neural responses."},"disciplines":["neuroscience"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":129},{"id":"arxiv-2610.13166","arxiv":"2610.13166","title":"Scaling Agentic Data for Long-Horizon Terminal Intelligence","authors":["Shun Zou","Ziyu Ma","Yi Zou","Yong Wang","Lin Chen","Zehui Chen","Guanhua Chen","Xiangxiang Chu"],"authorCount":9,"org":"alibaba","affiliations":[],"orgNote":null,"date":"2026-10-09","hfDaily":"2026-10-13","primaryCategory":"cs.CL","categories":["cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.13166","pdf":"https://arxiv.org/pdf/2610.13166","hf":"https://huggingface.co/papers/2610.13166","code":null,"project":"https://zs1314.github.io/TerminalHorizon"},"metrics":{"hfUpvotes":165,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":56.77,"tags":["agent","benchmark"],"summary":{"zh":"研究如何规模化合成高质量长程终端任务轨迹，训练能完成专业任务的终端 Agent。","en":"Scales synthesis of high-quality long-horizon terminal trajectories to train agents for professional tasks."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":130},{"id":"arxiv-2610.05107","arxiv":"2610.05107","title":"SearchJev: A Fast and Calibrated System-1 Model for Search Agents","authors":["Congfeng Cao","Lipeng Zuo","Konstantinos Papakostas","Qiwei Xu","Songwei Xu","Lun Zhou","Zhaochun Ren","Yougang Lyu"],"authorCount":9,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-10-04","hfDaily":"2026-10-06","primaryCategory":"cs.IR","categories":["cs.IR","cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.05107","pdf":"https://arxiv.org/pdf/2610.05107","hf":"https://huggingface.co/papers/2610.05107","code":"https://github.com/EvoScientist/SearchJev","project":"https://huggingface.co/SearchJev/models"},"metrics":{"hfUpvotes":76,"hfComments":2,"citations":1,"citationsSource":"semantic_scholar","githubStars":34},"hot":56.75,"tags":["agent","efficiency"],"summary":{"zh":"SearchJev 是快速且校准良好的 System 1 模型，专门替搜索 Agent 做相关性和下一步动作的判断。","en":"SearchJev is a fast, calibrated System 1 model that makes relevance and action decisions for search agents."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":131},{"id":"arxiv-2610.09872","arxiv":"2610.09872","title":"LiveMACE: Process-Aware Evaluation of LLM Agent Capabilities in Evolving Markets","authors":["Jun Zhao","Leiming Fu","Yanbo Wen","Yiding Wang","Xuantong Liu","Yang Shu","Yuyang Lu","Xuanran Xing"],"authorCount":12,"org":"Fudan University","affiliations":[],"orgNote":null,"date":"2026-10-07","hfDaily":"2026-10-19","primaryCategory":"cs.AI","categories":["cs.AI","cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.09872","pdf":"https://arxiv.org/pdf/2610.09872","hf":"https://huggingface.co/papers/2610.09872","code":"https://github.com/RaymondVoldemortFu/LiveMACE","project":null},"metrics":{"hfUpvotes":132,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":9},"hot":56.74,"tags":["agent","benchmark"],"summary":{"zh":"LiveMACE 用实时金融市场做测试场，从结果和过程两方面评估持续运行的大模型 Agent。","en":"LiveMACE uses live financial markets to evaluate persistent LLM agents on both outcomes and process diagnostics."},"disciplines":["finance","economics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":132},{"id":"arxiv-2610.21261","arxiv":"2610.21261","title":"Behavioral Compositionality of Sparse Autoencoder Representations","authors":["Haotian Ma","Ziye Mei","Frederic Sala"],"authorCount":3,"org":"OpenSAE","affiliations":[],"orgNote":null,"date":"2026-10-15","hfDaily":"2026-10-16","primaryCategory":"cs.LG","categories":["cs.LG","cs.AI","cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.21261","pdf":"https://arxiv.org/pdf/2610.21261","hf":"https://huggingface.co/papers/2610.21261","code":"https://github.com/OpenSAE/SAE-Behavioral-Compositionality","project":"https://opensae.github.io/SAE-Behavioral-Compositionality"},"metrics":{"hfUpvotes":110,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":23},"hot":56.69,"tags":["safety"],"summary":{"zh":"提出 SAR 指标，衡量稀疏自编码器特征在下游行为上能否“相加组合”，服务可解释性研究。","en":"Introduces SAR to measure whether sparse-autoencoder features compose additively in downstream behaviour, aiding interpretability."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":133},{"id":"arxiv-2610.03715","arxiv":"2610.03715","title":"4DCodeBench: Benchmarking Agents on Inverse Graphics of Dynamic Scenes","authors":["Ruihong Shen","Žiga Kovačič","Peter Kulits","Xingrui Wang","Zizhang Li","Joshua B. Tenenbaum","Alan Yuille","Jieneng Chen"],"authorCount":9,"org":"4DCodeBench","affiliations":[],"orgNote":null,"date":"2026-10-02","hfDaily":"2026-10-05","primaryCategory":"cs.CV","categories":["cs.CV","cs.AI","cs.GR"],"links":{"abs":"https://arxiv.org/abs/2610.03715","pdf":"https://arxiv.org/pdf/2610.03715","hf":"https://huggingface.co/papers/2610.03715","code":"https://github.com/4DCodeBench/4DCodeBench","project":"https://4dcodebench.com/"},"metrics":{"hfUpvotes":34,"hfComments":2,"citations":4,"citationsSource":"semantic_scholar","githubStars":178},"hot":56.61,"tags":["benchmark","code","vision"],"summary":{"zh":"4DCodeBench 要求 Agent 把动态视频重建为可执行的图形程序，考察 4D 逆向图形能力。","en":"4DCodeBench asks agents to reconstruct dynamic scenes from video as executable graphics programs."},"disciplines":["software-engineering"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":134},{"id":"arxiv-2610.28630","arxiv":"2610.28630","title":"SolarRenderer: Few-Step Generative Reference Rendering","authors":["Guian Fang","Shengju Qian","Dongyue Lu","Junchao Huang","Lutao Jiang","Xin Wang","Yiren Song","Li Jiang"],"authorCount":9,"org":"Show Lab","affiliations":[],"orgNote":null,"date":"2026-10-20","hfDaily":"2026-10-21","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.28630","pdf":"https://arxiv.org/pdf/2610.28630","hf":"https://huggingface.co/papers/2610.28630","code":"https://github.com/Enderfga/SolarRenderer","project":"https://enderfga.cn/SolarRenderer"},"metrics":{"hfUpvotes":47,"hfComments":2,"citations":1,"citationsSource":"semantic_scholar","githubStars":215},"hot":56.57,"tags":["visual-gen","efficiency"],"summary":{"zh":"SolarRenderer 给冻结的视频大模型加两个 LoRA，把游戏引擎粗结构视频四步渲染成真实画面，并支持实时流式。","en":"SolarRenderer adds two LoRAs to a frozen video model to render coarse structural video into realistic footage in four steps, including real-time streaming."},"disciplines":["software-engineering"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":135},{"id":"arxiv-2610.10429","arxiv":"2610.10429","title":"SGF+: Decoupling Gradient Flows for Autoregressive Video Generation","authors":["Zihan Su","Junhao Zhuang","Yaowei Li","Siwen Lu","Haoran Li","Lingen Li","Haoyu Wu","Weiyang Jin"],"authorCount":13,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-10-07","hfDaily":"2026-10-08","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.10429","pdf":"https://arxiv.org/pdf/2610.10429","hf":"https://huggingface.co/papers/2610.10429","code":"https://github.com/Zihan-Su/Self_Gradient_Forcing_Plus","project":"https://zihan-su.github.io/self-gradient-forcing-plus/"},"metrics":{"hfUpvotes":60,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":207},"hot":56.53,"tags":["visual-gen"],"summary":{"zh":"SGF+ 把自回归视频生成中“写上下文”和“去噪”分到不同参数，缓解两者梯度冲突。","en":"SGF+ separates parameters for context writing and denoising in autoregressive video generation to resolve conflicting gradients."},"disciplines":["film-animation"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":136},{"id":"arxiv-2610.12611","arxiv":"2610.12611","title":"SoftwareGym: Scaling Computer-Use Agent Training with Diverse Synthesized Software Environments","authors":["Tian Jin","Wenhao Wang","Jing Kang","Haoting Shi","Rui Ye","Yijun Lu","Siheng Chen"],"authorCount":7,"org":"Shanghai Jiao Tong University","affiliations":[],"orgNote":null,"date":"2026-10-08","hfDaily":"2026-10-12","primaryCategory":"cs.MA","categories":["cs.MA","cs.SE"],"links":{"abs":"https://arxiv.org/abs/2610.12611","pdf":"https://arxiv.org/pdf/2610.12611","hf":"https://huggingface.co/papers/2610.12611","code":"https://github.com/MrKiing-Jin/SoftwareGym","project":null},"metrics":{"hfUpvotes":111,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":20},"hot":56.44,"tags":["agent","benchmark"],"summary":{"zh":"SoftwareGym 自动合成多样、可运行且可验证的 GUI 软件环境，用来扩大电脑操作 Agent 的训练。","en":"SoftwareGym automatically synthesises diverse, executable, verifiable GUI software environments to scale computer-use agent training."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":137},{"id":"arxiv-2610.14709","arxiv":"2610.14709","title":"Vela: User Behavior as Supervision Sparks LLM Intelligence","authors":["Yang Xiao","Yusong Sun","Haoyi Wu","Wenyang Hui","Weiqiang Hang","Luqing Wang","Zejian Wang","Zirui Tu"],"authorCount":20,"org":"AllSpark Research","affiliations":[],"orgNote":null,"date":"2026-10-10","hfDaily":"2026-10-14","primaryCategory":"cs.CL","categories":["cs.CL","cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.14709","pdf":"https://arxiv.org/pdf/2610.14709","hf":"https://huggingface.co/papers/2610.14709","code":"https://github.com/AllSpark-Research/Vela","project":"https://allspark-research.github.io/Vela/"},"metrics":{"hfUpvotes":88,"hfComments":6,"citations":0,"citationsSource":"semantic_scholar","githubStars":30},"hot":56.39,"tags":["llm"],"summary":{"zh":"Vela 把真实用户长期行为记录当作监督信号，提升大模型推断偏好、意图和未来行为的能力。","en":"Vela uses real long-term user behaviour records as supervision to improve LLMs at inferring preferences, intent and future actions."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":138},{"id":"arxiv-2610.06647","arxiv":"2610.06647","title":"LoGRA: Scaling LLM Reinforcement Learning with Low-Rank Gradient Sketches","authors":["Shaokun Zhang","Yifan Zhang","Jian Hu","Yueying Li","Hao Zhang","Binfeng Xu","Jan Kautz","Yi Dong"],"authorCount":8,"org":"NVIDIA","affiliations":[],"orgNote":null,"date":"2026-10-05","hfDaily":"2026-10-06","primaryCategory":"cs.CL","categories":["cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.06647","pdf":"https://arxiv.org/pdf/2610.06647","hf":"https://huggingface.co/papers/2610.06647","code":null,"project":null},"metrics":{"hfUpvotes":152,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":56.38,"tags":["rl","efficiency"],"summary":{"zh":"LoGRA 用低秩梯度草图压缩 RL 后训练的显存，并按预测 KL 控制步长保持稳定。","en":"LoGRA cuts RL post-training memory with low-rank gradient sketches and keeps training stable with predicted-KL step control."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":139},{"id":"arxiv-2610.02162","arxiv":"2610.02162","title":"World Observer: Joint Actor-Observer Generation for Persistent World Modeling","authors":["Hyunwook Choi","Dahyun Chung","Hyunsung Kim","Siyoon Jin","Jinhyeok Choi","Junyoung Seo","Seungryong Kim"],"authorCount":7,"org":"KAIST AI","affiliations":[],"orgNote":null,"date":"2026-10-01","hfDaily":"2026-10-02","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.02162","pdf":"https://arxiv.org/pdf/2610.02162","hf":"https://huggingface.co/papers/2610.02162","code":"https://github.com/cvlab-kaist/world-observer","project":"https://cvlab-kaist.github.io/world-observer/"},"metrics":{"hfUpvotes":89,"hfComments":4,"citations":0,"citationsSource":"semantic_scholar","githubStars":34},"hot":56.3,"tags":["visual-gen"],"summary":{"zh":"World Observer 在行动者视角之外同时生成全景“观察者”画面，让离开视野的物体也持续演化。","en":"World Observer jointly generates panoramic observer views alongside the actor view so off-screen objects keep evolving consistently."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":140},{"id":"arxiv-2609.40340","arxiv":"2609.40340","title":"EvoDuet: Bilevel Co-Evolution of Web Searching and Task Solving for Scientific Discovery","authors":["Young-Jun Lee","Jinheon Baek","Soyeong Jeong","Minki Kang","Seungyeon Jwa","Jonghyun Choi","Seungho Han","Dongyeop Kang"],"authorCount":8,"org":"Minnesota NLP","affiliations":[],"orgNote":null,"date":"2026-09-30","hfDaily":"2026-10-01","primaryCategory":"cs.CL","categories":["cs.CL"],"links":{"abs":"https://arxiv.org/abs/2609.40340","pdf":"https://arxiv.org/pdf/2609.40340","hf":"https://huggingface.co/papers/2609.40340","code":"https://github.com/Open-Galapagos/EvoDuet","project":"https://open-galapagos.github.io/evoduet_project_page/"},"metrics":{"hfUpvotes":112,"hfComments":1,"citations":1,"citationsSource":"semantic_scholar","githubStars":7},"hot":56.28,"tags":["ai4science","agent"],"summary":{"zh":"EvoDuet 让解法和搜索查询双层协同进化，帮助大模型进化搜索在需要外部知识时不再停滞。","en":"EvoDuet co-evolves solutions and web-search queries in a bilevel loop so LLM evolutionary search doesn't stall for lack of knowledge."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":141},{"id":"arxiv-2610.02368","arxiv":"2610.02368","title":"Rethinking World-Action Model for Compositional and In-Context Robotic Manipulation","authors":["Shukai Gong","Xuanran Zhai","Yintianrun Zhang","Ruopeng Cui","Ye Huang","Yiyang Fu","Dexuan Lyu","Chaojie Li"],"authorCount":19,"org":"DAGroup-PKU","affiliations":[],"orgNote":null,"date":"2026-10-01","hfDaily":"2026-10-08","primaryCategory":"cs.RO","categories":["cs.RO"],"links":{"abs":"https://arxiv.org/abs/2610.02368","pdf":"https://arxiv.org/pdf/2610.02368","hf":"https://huggingface.co/papers/2610.02368","code":"https://github.com/DAGroup-PKU/ViGAR","project":"https://dagroup-pku.github.io/ViGAR/#"},"metrics":{"hfUpvotes":72,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":104},"hot":56.19,"tags":["robotics","reasoning"],"summary":{"zh":"ViGAR 把长程操作拆成“视觉子目标规划 + 子目标执行”，提升组合式和上下文学习式机器人操作。","en":"ViGAR splits long-horizon manipulation into a visual subgoal planner and a subgoal executor for compositional, in-context robot tasks."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":142},{"id":"arxiv-2610.20818","arxiv":"2610.20818","title":"Intern Lumina U2: A Multi-Codebook Diffusion Large Language Model for Omni-Visual Understanding and Image Generation","authors":["Qian Yu","Zongyang Qiu","Juncheng Yan","Hongliang Li","Qi Qin","Keqi Wang","Siqi Luo","Junkui Ju"],"authorCount":43,"org":"Shanghai AI Laboratory","affiliations":[],"orgNote":null,"date":"2026-10-14","hfDaily":"2026-10-16","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.20818","pdf":"https://arxiv.org/pdf/2610.20818","hf":"https://huggingface.co/papers/2610.20818","code":"https://github.com/InternLM/InternLumina-U2","project":"https://internlm.github.io/InternLumina-U2/"},"metrics":{"hfUpvotes":51,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":81},"hot":56.18,"tags":["multimodal","visual-gen"],"summary":{"zh":"Intern Lumina U2 用多码本扩散大模型在同一空间里统一视觉理解和图像生成。","en":"Intern Lumina U2 is a multi-codebook diffusion LLM unifying visual understanding and image generation in one space."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":143},{"id":"arxiv-2609.25186","arxiv":"2609.25186","title":"From Pattern Recognizers to Personalized Companions: A Survey of Large Language Models in Mental Health","authors":["He Hu","Yucheng Zhou","Qianning Wang","Yingjian Zou","Chiyuan Ma","Juzheng Si","Jianzhuang Liu","Zitong Yu"],"authorCount":11,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-09-21","hfDaily":"2026-09-23","primaryCategory":"cs.CY","categories":["cs.CY","cs.AI","cs.CL"],"links":{"abs":"https://arxiv.org/abs/2609.25186","pdf":"https://arxiv.org/pdf/2609.25186","hf":"https://huggingface.co/papers/2609.25186","code":"https://github.com/Emo-gml/Awesome-Mental-Health-LLMs","project":null},"metrics":{"hfUpvotes":33,"hfComments":3,"citations":4,"citationsSource":"semantic_scholar","githubStars":141},"hot":56.17,"tags":["llm","ai4science"],"summary":{"zh":"大模型在心理健康领域应用的综述，梳理从模式识别到个性化陪伴的演进脉络。","en":"A survey of LLMs in mental health, tracing the evolution from pattern recognisers to personalised companions."},"disciplines":["clinical-medicine","psychology"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":144},{"id":"arxiv-2610.20693","arxiv":"2610.20693","title":"Orbia: Do Generated Videos Form a Consistent 3D World?","authors":["Dongyue Lu","Guian Fang","Hur Lim","Yonggang Wu","Yanrong Wang","Hongshan Chen","Shengju Qian","Yue Liao"],"authorCount":23,"org":"Tencent","affiliations":[],"orgNote":null,"date":"2026-10-14","hfDaily":"2026-10-16","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.20693","pdf":"https://arxiv.org/pdf/2610.20693","hf":"https://huggingface.co/papers/2610.20693","code":"https://github.com/worldbench/orbia","project":"https://orbia-bench.github.io"},"metrics":{"hfUpvotes":43,"hfComments":2,"citations":1,"citationsSource":"semantic_scholar","githubStars":48},"hot":55.99,"tags":["visual-gen","benchmark"],"summary":{"zh":"Orbia 是首个系统评测交互式世界模型 3D 一致性的基准，检查回访时场景是否保持不变。","en":"Orbia is a benchmark for 3D consistency of interactive world models, testing whether scenes persist on revisits."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":145},{"id":"arxiv-2609.32607","arxiv":"2609.32607","title":"VoxMem: Benchmarking Multimodal Memory in Large Audio Language Models","authors":["Yang Xiao","Vidhyasaharan Sethu","Eun-Jung Holden","Ting Dang"],"authorCount":4,"org":"The University of Melbourne","affiliations":[],"orgNote":null,"date":"2026-09-26","hfDaily":"2026-09-30","primaryCategory":"eess.AS","categories":["eess.AS","cs.SD"],"links":{"abs":"https://arxiv.org/abs/2609.32607","pdf":"https://arxiv.org/pdf/2609.32607","hf":"https://huggingface.co/papers/2609.32607","code":"https://github.com/swagshaw/voxmem","project":"https://swagshaw.github.io/voxmem/"},"metrics":{"hfUpvotes":134,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":7},"hot":55.9,"tags":["multimodal","benchmark"],"summary":{"zh":"VoxMem 评测音频大模型能否记住语音里“谁说的、怎么说的、听到了什么”等转写之外的信息。","en":"VoxMem benchmarks whether audio LLMs remember who said what, how, and what was audible—information beyond the transcript."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":146},{"id":"arxiv-2609.26333","arxiv":"2609.26333","title":"Disaggregated Quantization: Specializing LLM Prefill and Decode","authors":["Andrei Panferov","Maximilian Kleinegger","Sweta Priyadarshi","Tijmen Blankevoort","Dan Alistarh"],"authorCount":5,"org":" IST Austria Distributed Algorithms and Systems Lab","affiliations":[],"orgNote":null,"date":"2026-09-22","hfDaily":"2026-09-28","primaryCategory":"cs.LG","categories":["cs.LG"],"links":{"abs":"https://arxiv.org/abs/2609.26333","pdf":"https://arxiv.org/pdf/2609.26333","hf":"https://huggingface.co/papers/2609.26333","code":"https://github.com/IST-DASLab/disaggregated-quantization","project":null},"metrics":{"hfUpvotes":92,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":28},"hot":55.82,"tags":["efficiency"],"summary":{"zh":"分离式量化：为预填充和解码两个阶段分别定制计算格式与权重，兼顾速度和精度。","en":"Disaggregated quantisation specialises formats and weights for prefill and decode separately to balance speed and accuracy."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":147},{"id":"arxiv-2610.18829","arxiv":"2610.18829","title":"OmniCompass: A Holistic Benchmark for Omni-Modal Video Grounding and Understanding","authors":["Xuyang Liu","Shihao Wang","Xinyu Wei","Minghan Li","Zeqing Wang","Zuhao Yang","Zichen Wen","Lei Zhang"],"authorCount":8,"org":"VCLab","affiliations":[],"orgNote":null,"date":"2026-10-13","hfDaily":"2026-10-16","primaryCategory":"cs.CV","categories":["cs.CV","cs.CL","cs.MM"],"links":{"abs":"https://arxiv.org/abs/2610.18829","pdf":"https://arxiv.org/pdf/2610.18829","hf":"https://huggingface.co/papers/2610.18829","code":"https://github.com/PolyU-VCLab/OmniCompass","project":"https://omnicompass.netlify.app/"},"metrics":{"hfUpvotes":60,"hfComments":4,"citations":0,"citationsSource":"semantic_scholar","githubStars":135},"hot":55.8,"tags":["multimodal","benchmark"],"summary":{"zh":"OmniCompass 分视觉和音视频两条赛道，全面评测模型的视频定位与理解能力。","en":"OmniCompass evaluates video grounding and understanding on separate vision-language and audio-visual tracks."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":148},{"id":"arxiv-2610.12461","arxiv":"2610.12461","title":"OuroWorld: Bringing Any 3D World Alive as Diverse, Endlessly Looping 3D Cinemagraphs","authors":["You-Zhe Xie","Ting-Wei Chou","Yu-Hsuan Li","Kaipeng Zhang","Zhixiang Wang","Yu-Lun Liu"],"authorCount":6,"org":"Alaya Lab","affiliations":[],"orgNote":null,"date":"2026-10-08","hfDaily":"2026-10-09","primaryCategory":"cs.CV","categories":["cs.CV","cs.GR"],"links":{"abs":"https://arxiv.org/abs/2610.12461","pdf":"https://arxiv.org/pdf/2610.12461","hf":"https://huggingface.co/papers/2610.12461","code":"https://github.com/youzhe0305/OuroWorld","project":"https://ouroworld.userwei.com/"},"metrics":{"hfUpvotes":64,"hfComments":2,"citations":1,"citationsSource":"semantic_scholar","githubStars":44},"hot":55.68,"tags":["visual-gen"],"summary":{"zh":"OuroWorld 把任意静态 3D 高斯场景变成可从任意视角无缝循环的“3D 动态照片”。","en":"OuroWorld turns any static 3D Gaussian scene into a seamlessly looping 3D cinemagraph viewable from any angle."},"disciplines":["film-animation"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":149},{"id":"arxiv-2609.37200","arxiv":"2609.37200","title":"Adaptive Reward Routing: Dynamic Multi-Reward Optimization for Joint Audio-Video Diffusion via Forward-Process RL","authors":["Songlin Yang","Xiaotong Zhao","Jiacheng Zhang","Zhe Wang","Toyota Li","Eric Liu","Alan Zhao","Anyi Rao"],"authorCount":8,"org":"Tencent","affiliations":[],"orgNote":null,"date":"2026-09-29","hfDaily":"2026-10-02","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2609.37200","pdf":"https://arxiv.org/pdf/2609.37200","hf":"https://huggingface.co/papers/2609.37200","code":null,"project":null},"metrics":{"hfUpvotes":140,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":55.57,"tags":["visual-gen","rl"],"summary":{"zh":"自适应奖励路由在训练中动态调整更新位置和多个奖励的权重，改进音视频联合生成。","en":"Adaptive Reward Routing dynamically adjusts where updates act and how rewards combine to improve joint audio-video diffusion."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":150},{"id":"arxiv-2610.16794","arxiv":"2610.16794","title":"TPPO: Tracking-Value PPO For Language Model Reasoning","authors":["Futing Wang","Chenxi Li","Jianhao Yan","Yun Luo","Zhi Wang","Yizhuo Li","Runzhe Zhan","Yu Cheng"],"authorCount":9,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-10-12","hfDaily":"2026-10-14","primaryCategory":"cs.LG","categories":["cs.LG","cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.16794","pdf":"https://arxiv.org/pdf/2610.16794","hf":"https://huggingface.co/papers/2610.16794","code":"https://github.com/LINs-lab/TPPO","project":null},"metrics":{"hfUpvotes":96,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":21},"hot":55.55,"tags":["rl","reasoning"],"summary":{"zh":"TPPO 用累加式价值参数化追踪推理过程中的价值变化，改进 PPO 的逐 token 信用分配。","en":"TPPO uses a cumulative value parameterisation to track value along the trajectory, improving token-level credit in PPO for reasoning."},"disciplines":["mathematics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":151},{"id":"arxiv-2609.23038","arxiv":"2609.23038","title":"Spatial-Interactor: Learning Spatial Reasoning through Interaction with the Observable Physical World","authors":["Kaixiang Yao","Xu Wang","Miao Pan","Hu Xiyue","Jiaxin Li","Weishi Wang","Daniel Dahlmeier","Jintao Chen"],"authorCount":13,"org":"ZJU-OmniAI","affiliations":[],"orgNote":null,"date":"2026-09-19","hfDaily":"2026-09-24","primaryCategory":"cs.AI","categories":["cs.AI","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2609.23038","pdf":"https://arxiv.org/pdf/2609.23038","hf":"https://huggingface.co/papers/2609.23038","code":"https://github.com/ZJU-OmniAI/Spatial-Interactor","project":"https://zju-omniai.github.io/Spatial-Interactor/"},"metrics":{"hfUpvotes":64,"hfComments":4,"citations":0,"citationsSource":"semantic_scholar","githubStars":90},"hot":55.44,"tags":["multimodal","reasoning"],"summary":{"zh":"Spatial-Interactor 用“观察-动作-新观察”的交互轨迹训练视觉语言模型的动态空间推理。","en":"Spatial-Interactor trains VLM spatial reasoning on observation-action-observation interaction trajectories."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":152},{"id":"arxiv-2610.10524","arxiv":"2610.10524","title":"GRACE: Generation-aware latent compression for efficient video generation","authors":["Jiyoung Kim","Paul Hyunbin Cho","Jisu Nam","Donghoon Lee","Hyunsung Go","Yeonkyeong Lee","Hansaem Kim","Seungryong Kim"],"authorCount":8,"org":"KAIST AI","affiliations":[],"orgNote":null,"date":"2026-10-07","hfDaily":"2026-10-08","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.10524","pdf":"https://arxiv.org/pdf/2610.10524","hf":"https://huggingface.co/papers/2610.10524","code":"https://github.com/cvlab-kaist/GRACE","project":"https://cvlab-kaist.github.io/GRACE/"},"metrics":{"hfUpvotes":90,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":26},"hot":55.43,"tags":["visual-gen","efficiency"],"summary":{"zh":"GRACE 在保持与预训练 DiT 兼容的前提下进一步压缩视频潜空间，加速视频生成。","en":"GRACE further compresses the video latent while staying compatible with the pretrained DiT, speeding up video generation."},"disciplines":["film-animation"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":153},{"id":"arxiv-2610.19820","arxiv":"2610.19820","title":"ActionOPD: Distilling Decisions, Not Tokens","authors":["Huining Yuan","Xiangmin Yi","Zelai Xu","Zixu Wang","Xinlei Chen","Xiao-Ping Zhang","Yu Wang","Chao Yu"],"authorCount":8,"org":"Tsinghua-NICS-EFC","affiliations":[],"orgNote":null,"date":"2026-10-14","hfDaily":"2026-10-15","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.19820","pdf":"https://arxiv.org/pdf/2610.19820","hf":"https://huggingface.co/papers/2610.19820","code":"https://github.com/thu-nics/ActionOPD","project":"https://thu-nics.github.io/ActionOPD/"},"metrics":{"hfUpvotes":64,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":20},"hot":55.43,"tags":["agent","efficiency"],"summary":{"zh":"ActionOPD 以“动作”而不是 token 为单位做在线蒸馏，把大模型 Agent 的决策能力迁移给小模型。","en":"ActionOPD distils agents at the level of semantic actions rather than tokens to transfer decisions to smaller models."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":154},{"id":"arxiv-2609.25165","arxiv":"2609.25165","title":"Ovis-Embedding: Pushing the Frontiers of Universal Omni-Modal Embeddings","authors":["Ovis-Embedding Team"],"authorCount":1,"org":"ATH-MaaS","affiliations":[],"orgNote":null,"date":"2026-09-21","hfDaily":"2026-09-23","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2609.25165","pdf":"https://arxiv.org/pdf/2609.25165","hf":"https://huggingface.co/papers/2609.25165","code":"https://github.com/ATH-MaaS/Ovis-Omni-Embedding","project":null},"metrics":{"hfUpvotes":77,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":58},"hot":55.41,"tags":["multimodal"],"summary":{"zh":"Ovis-Embedding 基于全模态底座，把文本、图像、视频和音频编码到同一个向量空间。","en":"Ovis-Embedding encodes text, images, video and audio into one shared embedding space using an omni-modal backbone."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":155},{"id":"arxiv-2610.25690","arxiv":"2610.25690","title":"ScaleNav: Scalable Internet Video Pretraining for Language-Conditioned Local Navigation","authors":["Yihui Yang","Yuanhong Yu","Jiawei Wang","Zhongrui Yu","Jia Shi","Kun Zhan","Ruizhen Hu","Xiaowei Zhou"],"authorCount":10,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-10-19","hfDaily":"2026-10-20","primaryCategory":"cs.RO","categories":["cs.RO"],"links":{"abs":"https://arxiv.org/abs/2610.25690","pdf":"https://arxiv.org/pdf/2610.25690","hf":"https://huggingface.co/papers/2610.25690","code":"https://github.com/MachEmbodied/ScaleNav","project":"https://zju3dv.github.io/ScaleNav/"},"metrics":{"hfUpvotes":84,"hfComments":2,"citations":null,"citationsSource":null,"githubStars":40},"hot":55.36,"tags":["robotics"],"summary":{"zh":"ScaleNav 从互联网第一视角视频中学习目标导向的视觉动态，用于语言指令下的机器人局部导航。","en":"ScaleNav learns goal-directed visual dynamics from internet egocentric videos for language-conditioned robot navigation."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":156},{"id":"arxiv-2609.23377","arxiv":"2609.23377","title":"One to More, More to One: Category-Aware Iterative Expert Training for Software Engineering Agents","authors":["Jie Zhao","Ziyu Jiang","Suhang Zheng","Minghui Shan","Xiaoxiao Xu","Lin Qu"],"authorCount":6,"org":"alibaba","affiliations":[],"orgNote":null,"date":"2026-09-20","hfDaily":"2026-09-22","primaryCategory":"cs.SE","categories":["cs.SE","cs.CL","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2609.23377","pdf":"https://arxiv.org/pdf/2609.23377","hf":"https://huggingface.co/papers/2609.23377","code":"https://github.com/alibaba/AgenticBigBang","project":"https://github.com/alibaba/AgenticBigBang"},"metrics":{"hfUpvotes":50,"hfComments":4,"citations":0,"citationsSource":"semantic_scholar","githubStars":45},"hot":55.3,"tags":["code","rl","agent"],"summary":{"zh":"按任务类别分别训练软件工程专家再合并，缓解统一 RL 训练中各类任务此消彼长的问题。","en":"Trains category-specific SWE experts and then integrates them, easing the see-saw across task categories in pooled RL."},"disciplines":["software-engineering"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":157},{"id":"arxiv-2609.38143","arxiv":"2609.38143","title":"Learning Meta-Skills for Agent Harness Design in Test-Time AI4AI","authors":["Cheng Qian","Kunlun Zhu","Beibin Li","Zhenhailong Wang","Heng Ji"],"authorCount":5,"org":"Apodex","affiliations":[],"orgNote":null,"date":"2026-09-29","hfDaily":"2026-10-01","primaryCategory":"cs.AI","categories":["cs.AI","cs.CL","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2609.38143","pdf":"https://arxiv.org/pdf/2609.38143","hf":"https://huggingface.co/papers/2609.38143","code":"https://github.com/qiancheng-apodex/MetaSkill-AI4AI","project":null},"metrics":{"hfUpvotes":86,"hfComments":2,"citations":1,"citationsSource":"semantic_scholar","githubStars":11},"hot":55.29,"tags":["agent"],"summary":{"zh":"让“构建者”模型学习元技能，为固定的目标模型自动搭建更好的执行环境（测试时 AI4AI）。","en":"Has a Builder model learn meta-skills to construct better harnesses for a fixed Target model at test time (AI-for-AI)."},"disciplines":["finance"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":158},{"id":"arxiv-2610.22347","arxiv":"2610.22347","title":"Speculative Forcing for Efficient Autoregressive Video Generation","authors":["Hyunwook Choi","Seungryong Kim","Dongyoon Han","Sangdoo Yun"],"authorCount":4,"org":"NAVER AI Lab","affiliations":[],"orgNote":null,"date":"2026-10-15","hfDaily":"2026-10-16","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.22347","pdf":"https://arxiv.org/pdf/2610.22347","hf":"https://huggingface.co/papers/2610.22347","code":"https://github.com/naver-ai/speculative-forcing","project":"https://speculative-forcing.github.io/"},"metrics":{"hfUpvotes":93,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":21},"hot":55.24,"tags":["visual-gen","efficiency"],"summary":{"zh":"推测式强制把中间去噪结果当草稿，提前接受已收敛的 token，加速自回归视频生成。","en":"Speculative Forcing treats intermediate estimates as drafts and accepts converged tokens early to speed up autoregressive video generation."},"disciplines":["film-animation"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":159},{"id":"arxiv-2609.38847","arxiv":"2609.38847","title":"Scoring Higher, Answering Worse: Mitigating Reward Hacking in Rubric-Based RL via Protocol-Level Rubrics","authors":["Maoqi Liu","Junwei He","Bowen Zhang","Feiran Li","Wentao Ma","Rongyi Lin","Shuhan Zhong","Quan Fang"],"authorCount":8,"org":"ByteDance","affiliations":[],"orgNote":null,"date":"2026-09-30","hfDaily":"2026-10-14","primaryCategory":"cs.LG","categories":["cs.LG","cs.AI"],"links":{"abs":"https://arxiv.org/abs/2609.38847","pdf":"https://arxiv.org/pdf/2609.38847","hf":"https://huggingface.co/papers/2609.38847","code":"https://github.com/Estrellajer/ProRubric","project":null},"metrics":{"hfUpvotes":68,"hfComments":1,"citations":0,"citationsSource":"semantic_scholar","githubStars":22},"hot":55.22,"tags":["rl","alignment","ai4science"],"summary":{"zh":"指出基于评分细则的 RL 用加权求和会被“刷分”（以临床问诊为例），并提出协议级细则加以缓解。","en":"Shows weighted-sum rubric rewards can be hacked (e.g. in clinical consultation) and proposes protocol-level rubrics to fix it."},"disciplines":["clinical-medicine"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":160},{"id":"arxiv-2609.33295","arxiv":"2609.33295","title":"TraceDance: An Automated System for Building Agent Behavior Benchmarks from Real-World Agent Deployment Traces","authors":["Dehai Min","Daoan Zhang","Yiming Zeng","Huayi Zhang","Ziyi Chen","Yan Zhang","Qinbo Bai","Mengyuan Chao"],"authorCount":16,"org":"ByteDance","affiliations":[],"orgNote":null,"date":"2026-09-27","hfDaily":"2026-09-29","primaryCategory":"cs.AI","categories":["cs.AI","cs.CL"],"links":{"abs":"https://arxiv.org/abs/2609.33295","pdf":"https://arxiv.org/pdf/2609.33295","hf":"https://huggingface.co/papers/2609.33295","code":"https://github.com/ZhishanQ/TraceDance","project":"https://zhishanq.github.io/TraceDance/"},"metrics":{"hfUpvotes":72,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":13},"hot":55.15,"tags":["agent","benchmark","safety"],"summary":{"zh":"TraceDance 从真实部署轨迹中自动构建针对特定不良行为的 Agent 测试集。","en":"TraceDance automatically builds targeted benchmarks for specific undesirable agent behaviours from real deployment traces."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":161},{"id":"arxiv-2610.12299","arxiv":"2610.12299","title":"Multi-Agent Egocentric World Model with Fine-Grained Embodied Interaction","authors":["Dahyun Chung","Siyoon Jin","Hyunwook Choi","Honggyu An","Junyoung Seo","Hyunsung Kim","Seung Wook Kim","Seungryong Kim"],"authorCount":8,"org":"KAIST AI","affiliations":[],"orgNote":null,"date":"2026-10-08","hfDaily":"2026-10-09","primaryCategory":"cs.CV","categories":["cs.CV","cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.12299","pdf":"https://arxiv.org/pdf/2610.12299","hf":"https://huggingface.co/papers/2610.12299","code":"https://github.com/cvlab-kaist/ME-World","project":"https://cvlab-kaist.github.io/ME-World/"},"metrics":{"hfUpvotes":72,"hfComments":2,"citations":1,"citationsSource":"semantic_scholar","githubStars":21},"hot":55.05,"tags":["visual-gen","robotics"],"summary":{"zh":"多智能体第一视角世界模型，同步生成多个 Agent 在同一环境中的细粒度交互画面。","en":"A multi-agent egocentric world model that generates synchronised first-person streams of agents interacting in a shared world."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":162},{"id":"arxiv-2609.36601","arxiv":"2609.36601","title":"SAKI: Maximal-Coupling-Routed Teacher Supervision for On-Policy Distillation","authors":["Miteto Wei","Xiaohan Wang","Zehao Chen","Jiajun Chai","Sichao Liu","Li Wang","Haoyuan Xu","Zhaoyu Hu"],"authorCount":10,"org":"meituan","affiliations":[],"orgNote":null,"date":"2026-09-29","hfDaily":"2026-09-30","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2609.36601","pdf":"https://arxiv.org/pdf/2609.36601","hf":"https://huggingface.co/papers/2609.36601","code":"https://github.com/Miteto-sudo/SAKI","project":null},"metrics":{"hfUpvotes":93,"hfComments":1,"citations":0,"citationsSource":"semantic_scholar","githubStars":29},"hot":54.98,"tags":["efficiency"],"summary":{"zh":"SAKI 用 KL 约束的教师引导采样和最大耦合来分配蒸馏监督，帮助弱学生模型学习。","en":"SAKI routes distillation supervision using KL-constrained teacher-guided rollouts and maximal coupling to help weak students."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":163},{"id":"arxiv-2609.28466","arxiv":"2609.28466","title":"The Past Frames the Future: Memory for Autoregressive Video Generation","authors":["Harold Haodong Chen","Rongjin Guo","Disen Lan","Wen-Jie Shu","Hongfei Zhang","Hanzhe Hu","Shengtao Yao","Zixin Zhang"],"authorCount":25,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-09-23","hfDaily":"2026-09-24","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2609.28466","pdf":"https://arxiv.org/pdf/2609.28466","hf":"https://huggingface.co/papers/2609.28466","code":"https://github.com/HaroldChen19/Awesome-AR-Video-Memory","project":null},"metrics":{"hfUpvotes":66,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":89},"hot":54.94,"tags":["visual-gen"],"summary":{"zh":"系统讨论自回归视频生成的记忆问题：如何在有限上下文中保住身份、状态和因果变化。","en":"Examines memory for autoregressive video generation: keeping identities, states and causal changes within bounded context."},"disciplines":["film-animation"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":164},{"id":"arxiv-2610.02508","arxiv":"2610.02508","title":"World Action Modeling with Progressive Visual Planning","authors":["Fei Zhang","Zhaochong An","Duncan Frost","Yikai Wang","Pengfei Liu","Ya Zhang","Michal Drozdzal","Amir Bar"],"authorCount":8,"org":"Meta","affiliations":[],"orgNote":null,"date":"2026-10-01","hfDaily":"2026-10-05","primaryCategory":"cs.AI","categories":["cs.AI","cs.CV","cs.RO"],"links":{"abs":"https://arxiv.org/abs/2610.02508","pdf":"https://arxiv.org/pdf/2610.02508","hf":"https://huggingface.co/papers/2610.02508","code":null,"project":"https://sii-ferenas.github.io/ProWAM-page/"},"metrics":{"hfUpvotes":98,"hfComments":3,"citations":1,"citationsSource":"semantic_scholar","githubStars":null},"hot":54.8,"tags":["robotics","efficiency"],"summary":{"zh":"ProWAM 预测动作和一串稀疏的视觉子目标，比生成整段视频更高效地引导长程操作。","en":"ProWAM predicts actions plus an ordered sequence of sparse visual subgoals, guiding long-horizon control more efficiently than full video."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":165},{"id":"arxiv-2610.05416","arxiv":"2610.05416","title":"Prism: Dynamic Sparse Attention for Native 2K Joint Video-Audio Generation Model Training","authors":["Shuyuan Tu","Qi Tian","Yinming Huang","Yue Wu","Xintong Han","Kaihang Pan","Weijie Kong","Jiangfeng Xiong"],"authorCount":11,"org":"Tencent Hunyuan","affiliations":[],"orgNote":null,"date":"2026-10-04","hfDaily":"2026-10-06","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.05416","pdf":"https://arxiv.org/pdf/2610.05416","hf":"https://huggingface.co/papers/2610.05416","code":"https://github.com/Tencent-Hunyuan/Prism","project":"https://francis-rings.github.io/Prism/"},"metrics":{"hfUpvotes":21,"hfComments":1,"citations":1,"citationsSource":"semantic_scholar","githubStars":619},"hot":54.8,"tags":["visual-gen","efficiency"],"summary":{"zh":"Prism 动态稀疏注意力聚焦发声区域，用于原生 2K 音视频联合生成模型的训练。","en":"Prism is dynamic sparse attention that focuses on sound-producing regions for training native 2K joint video-audio models."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":166},{"id":"arxiv-2609.37686","arxiv":"2609.37686","title":"EngiWorld: What Can Frontier Agents Deliver in Professional Engineering Environments?","authors":["Hongcheng Gao","Hailong Qu","Yu Lei","Henghui Sun","Haoyang Li","Yipeng Wei","Naihao Xue","Xiaohan Yu"],"authorCount":22,"org":"Zhiman AI","affiliations":[],"orgNote":null,"date":"2026-09-29","hfDaily":"2026-09-30","primaryCategory":"cs.AI","categories":["cs.AI","cs.CL"],"links":{"abs":"https://arxiv.org/abs/2609.37686","pdf":"https://arxiv.org/pdf/2609.37686","hf":"https://huggingface.co/papers/2609.37686","code":"https://github.com/Hongcheng-Gao/EngiWorld","project":"https://engiworld.github.io/"},"metrics":{"hfUpvotes":75,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":46},"hot":54.58,"tags":["agent","benchmark"],"summary":{"zh":"EngiWorld 收录 1301 个专家任务，覆盖 CAD、CAE、EDA 等 6 个工程领域和 26 款专业软件。","en":"EngiWorld has 1,301 expert tasks across six engineering domains (CAD, CAE, EDA…) and 26 professional software platforms."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":167},{"id":"arxiv-2609.38123","arxiv":"2609.38123","title":"HelixWorld: A Real-time Interactive Audio-Visual World Model","authors":["Lei Ke","Jiahao Pan","Zeyue Tian","Jiaming Wang","Haoyuan Huang","Kam Man Wu","Pengjun Fang","Hongyu Liu"],"authorCount":16,"org":"NoizAI","affiliations":[],"orgNote":null,"date":"2026-09-29","hfDaily":"2026-10-05","primaryCategory":"cs.CV","categories":["cs.CV","cs.MM","cs.SD"],"links":{"abs":"https://arxiv.org/abs/2609.38123","pdf":"https://arxiv.org/pdf/2609.38123","hf":"https://huggingface.co/papers/2609.38123","code":"https://github.com/NoizAI/HelixWorld","project":"https://helixworld.org"},"metrics":{"hfUpvotes":38,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":625},"hot":54.38,"tags":["visual-gen","multimodal"],"summary":{"zh":"HelixWorld 是实时可交互的音视频世界模型，画面和立体声随用户操作同步变化。","en":"HelixWorld is a real-time interactive world model where visuals and spatial stereo sound co-evolve with user input."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":168},{"id":"arxiv-2610.18784","arxiv":"2610.18784","title":"Pack Light, CarryOn: Building Object-Oriented Context for Long-Horizon Agents","authors":["Yuchen Lu","Julian McAuley","Lei Zhang"],"authorCount":3,"org":"McAuley-Lab","affiliations":[],"orgNote":null,"date":"2026-10-13","hfDaily":"2026-10-19","primaryCategory":"cs.AI","categories":["cs.AI","cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.18784","pdf":"https://arxiv.org/pdf/2610.18784","hf":"https://huggingface.co/papers/2610.18784","code":"https://github.com/Ethannju/carryon","project":"https://ethannju.github.io/carryon/"},"metrics":{"hfUpvotes":118,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":5},"hot":54.35,"tags":["agent","efficiency"],"summary":{"zh":"CarryOn 以可复用的“任务对象”组织长程 Agent 的上下文，而不是反复读历史记录。","en":"CarryOn organises long-horizon agent context around reusable task objects instead of reprocessing history."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":169},{"id":"arxiv-2609.29429","arxiv":"2609.29429","title":"Just Ask Jev: Reinforcement Learning for Calibrated Decisions as a Zero-Shot Detector of AI Alignment Failures","authors":["Ruoqi Guo","Yi Liu","Gelei Deng","Yuekang Li","Lida Zhao","Yutao Wu","Simin Chen","Ying Zhang"],"authorCount":9,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-09-24","hfDaily":"2026-09-25","primaryCategory":"cs.AI","categories":["cs.AI","cs.CL","cs.CR"],"links":{"abs":"https://arxiv.org/abs/2609.29429","pdf":"https://arxiv.org/pdf/2609.29429","hf":"https://huggingface.co/papers/2609.29429","code":"https://github.com/sumleo/RLCDAlignBench","project":"https://sumleo.github.io/RLCDAlignBench/"},"metrics":{"hfUpvotes":29,"hfComments":2,"citations":12,"citationsSource":"semantic_scholar","githubStars":28},"hot":54.34,"tags":["safety","alignment"],"summary":{"zh":"测试经 RL 校准的决策模型 Jev 能否零样本检测谄媚、越狱、欺骗等十类对齐失败。","en":"Tests whether Jev, a calibrated decision model, can zero-shot detect ten alignment failures such as sycophancy, jailbreaks and deception."},"disciplines":["cybersecurity"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":170},{"id":"arxiv-2609.36380","arxiv":"2609.36380","title":"LEGO-Anything: Coding Agents for 3D Scene Reconstruction","authors":["Xirui Li","Peng Shi","Mingwen Dong","Sheng Zhang","Zhuoyan Xu","Dongkyu Lee","Shuaichen Chang","Yi Xiang"],"authorCount":10,"org":"Amazon Web Services","affiliations":[],"orgNote":null,"date":"2026-09-28","hfDaily":"2026-09-30","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2609.36380","pdf":"https://arxiv.org/pdf/2609.36380","hf":"https://huggingface.co/papers/2609.36380","code":null,"project":"https://lego-anything.com/"},"metrics":{"hfUpvotes":129,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":54.32,"tags":["code","vision","agent"],"summary":{"zh":"LEGO-Anything 让编程 Agent 反复编写和执行 Blender 代码，从单张图片重建可编辑的 3D 场景。","en":"LEGO-Anything has a coding agent iteratively write Blender code to reconstruct editable 3D scenes from a single image."},"disciplines":["software-engineering"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":171},{"id":"arxiv-2610.14157","arxiv":"2610.14157","title":"OpenMobile-2: Building Versatile Mobile Agents with Scalable Environments and App-Native Tools","authors":["Chenyang Yan","Yingying Zhang","Kanzhi Cheng","Qiushi Sun","Zhengyuan Pan","Zheng Ma","Hang Yan","Yian Wang"],"authorCount":24,"org":"Nanjing University","affiliations":[],"orgNote":null,"date":"2026-10-09","hfDaily":"2026-10-19","primaryCategory":"cs.SE","categories":["cs.SE","cs.CL","cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.14157","pdf":"https://arxiv.org/pdf/2610.14157","hf":"https://huggingface.co/papers/2610.14157","code":"https://github.com/OpenMobile2/OpenMobile2-Code","project":"https://os-copilot.github.io/OpenMobile2-Home/"},"metrics":{"hfUpvotes":55,"hfComments":4,"citations":1,"citationsSource":"semantic_scholar","githubStars":33},"hot":54.26,"tags":["agent","rl"],"summary":{"zh":"OpenMobile-2 开放 110 多个应用的训练环境和约 1.2 万条交互轨迹，训练接近前沿水平的手机智能体。","en":"OpenMobile-2 releases environments spanning 110+ apps and ~12K trajectories to train a near-frontier, fully open mobile agent."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":172},{"id":"arxiv-2609.18703","arxiv":"2609.18703","title":"RayOrch: Programming and Executing Lineage-Controlled Multi-Grain Dataflows for Foundation-Model Data Preparation","authors":["Xiaochen Ma","Zimo Meng","Junzhu Liang","Youhe Jiang","Yue Cheng","Hao Liang","Bohan Zeng","Dengchun Li"],"authorCount":16,"org":"Peking University","affiliations":[],"orgNote":null,"date":"2026-09-16","hfDaily":"2026-09-28","primaryCategory":"cs.DC","categories":["cs.DC"],"links":{"abs":"https://arxiv.org/abs/2609.18703","pdf":"https://arxiv.org/pdf/2609.18703","hf":"https://huggingface.co/papers/2609.18703","code":"https://github.com/OpenDCAI/RayOrch","project":"https://opendcai.github.io/RayOrch-doc/"},"metrics":{"hfUpvotes":57,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":19},"hot":54.17,"tags":["efficiency"],"summary":{"zh":"RayOrch 是为基础模型数据准备设计的分布式数据流引擎，保留父子数据关系并高效批处理。","en":"RayOrch is a distributed dataflow engine for foundation-model data prep that preserves lineage while batching efficiently."},"disciplines":["software-engineering"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":173},{"id":"arxiv-2609.32577","arxiv":"2609.32577","title":"Groupwise Agentic Grading and Advantage Redistribution for Code Agent RL","authors":["Jinhao Dong","Liang Zhao","Zihao Yue","Wenhan Ma","Linghao Zhang","Lei Li","Shicheng Li","Yifan Song"],"authorCount":10,"org":"Xiaomi MiMo","affiliations":[],"orgNote":null,"date":"2026-09-26","hfDaily":"2026-09-29","primaryCategory":"cs.CL","categories":["cs.CL","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2609.32577","pdf":"https://arxiv.org/pdf/2609.32577","hf":"https://huggingface.co/papers/2609.32577","code":null,"project":null},"metrics":{"hfUpvotes":95,"hfComments":2,"citations":1,"citationsSource":"semantic_scholar","githubStars":null},"hot":54.06,"tags":["code","rl"],"summary":{"zh":"GAGAR 在编程 Agent RL 中按实现质量重新分配优势值，奖励更干净、更聚焦的代码修改。","en":"GAGAR redistributes advantages by implementation quality in code-agent RL, rewarding cleaner, targeted changes."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":174},{"id":"arxiv-2610.18774","arxiv":"2610.18774","title":"GreenWAM: A World Foundation Model for Robotics","authors":["Gennady Sidorov","Valerii Kornilov","Denis Grankin","Polina Fedotova","Aleksey Postnikov"],"authorCount":5,"org":"Sber Robotics Center","affiliations":[],"orgNote":null,"date":"2026-10-13","hfDaily":"2026-10-14","primaryCategory":"cs.RO","categories":["cs.RO"],"links":{"abs":"https://arxiv.org/abs/2610.18774","pdf":"https://arxiv.org/pdf/2610.18774","hf":"https://huggingface.co/papers/2610.18774","code":"https://github.com/greenvla/GreenWAM","project":"https://greenvla.github.io/GreenWAM/"},"metrics":{"hfUpvotes":110,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":6},"hot":54.04,"tags":["robotics"],"summary":{"zh":"GreenWAM 是机器人世界基础模型，用统一关节动作空间覆盖 20 类机器人，同时充当世界模型和策略。","en":"GreenWAM is a robotics world foundation model using a shared joint action space across 20 robot families, serving as both world model and policy."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":175},{"id":"arxiv-2609.38827","arxiv":"2609.38827","title":"More Choices, Fewer Decisions: Ordinal-Scale Bias in JEV-like Direct-Decision Models","authors":["Tianxiang Gao","Jinzhe Li","Zhiyuan Li","Yi Chang","Yuan Wu"],"authorCount":5,"org":"Jilin University","affiliations":[],"orgNote":null,"date":"2026-09-30","hfDaily":"2026-10-01","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2609.38827","pdf":"https://arxiv.org/pdf/2609.38827","hf":"https://huggingface.co/papers/2609.38827","code":"https://github.com/Glax147/jev_ordinal_scale_bia","project":null},"metrics":{"hfUpvotes":65,"hfComments":3,"citations":3,"citationsSource":"semantic_scholar","githubStars":5},"hot":54,"tags":["benchmark"],"summary":{"zh":"发现 JEV 类直接决策模型在有序打分时偏向中间档，没有充分利用整个评分尺度。","en":"Finds JEV-like direct-decision models under-use ordinal scales, over-predicting middle labels."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":176},{"id":"arxiv-2609.38057","arxiv":"2609.38057","title":"EVO-WAM: Evolving World Action Models through Video-Action Verification","authors":["Shiyang Zhou","Xionghao Wu","Wenbo Li","Shenghe Zheng","Jiyao Zhang","Songsong Yu","Yijun Yang","Jianhui Liu"],"authorCount":14,"org":"Harbin Institute of Techonology, Shenzhen","affiliations":[],"orgNote":null,"date":"2026-09-29","hfDaily":"2026-09-30","primaryCategory":"cs.CV","categories":["cs.CV","cs.RO"],"links":{"abs":"https://arxiv.org/abs/2609.38057","pdf":"https://arxiv.org/pdf/2609.38057","hf":"https://huggingface.co/papers/2609.38057","code":"https://github.com/Clausy9/EVO-WAM","project":"https://evo-wam.github.io/"},"metrics":{"hfUpvotes":49,"hfComments":3,"citations":1,"citationsSource":"semantic_scholar","githubStars":54},"hot":53.99,"tags":["robotics"],"summary":{"zh":"EVO-WAM 让世界-动作模型从自己生成并通过校验的视频-动作轨迹中学习新任务，无需新示范。","en":"EVO-WAM adapts world-action models to new tasks by learning from their own verified generated video-action trajectories."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":177},{"id":"arxiv-2610.19849","arxiv":"2610.19849","title":"DeltaStreamer: Towards Human-Like Agentic Streaming Video Understanding","authors":["Boyuan Sun","Deshui Miao","Wenzhao Gao","Shaoyong Jia","Shaohui Jiao","Chi Liu","Yan Shu","Derek Li"],"authorCount":11,"org":"ByteDance","affiliations":[],"orgNote":null,"date":"2026-10-14","hfDaily":"2026-10-16","primaryCategory":"cs.CV","categories":["cs.CV","cs.AI","cs.HC"],"links":{"abs":"https://arxiv.org/abs/2610.19849","pdf":"https://arxiv.org/pdf/2610.19849","hf":"https://huggingface.co/papers/2610.19849","code":"https://github.com/BBBBchan/DeltaStreamer","project":"https://github.com/BBBBchan/DeltaStreamer"},"metrics":{"hfUpvotes":37,"hfComments":4,"citations":0,"citationsSource":"semantic_scholar","githubStars":86},"hot":53.95,"tags":["multimodal","agent"],"summary":{"zh":"DeltaStreamer 模仿人类观察方式，结合变化感知记忆和按需回放，理解流式视频。","en":"DeltaStreamer mimics human observation with change-aware memory and query-driven replay for streaming video understanding."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":178},{"id":"arxiv-2610.06801","arxiv":"2610.06801","title":"MC-Sparse: Deconstructing and Closing the Dense-Sparse Attention Gap in Diffusion Transformers","authors":["Jiarui Chen","Zeqiang Lai","Jiangshan Wang","Ziheng Ouyang","Ye Huang","Xiangyu Yue","Cewu Lu","Chunchao Guo"],"authorCount":8,"org":"Tencent Hunyuan","affiliations":[],"orgNote":null,"date":"2026-10-05","hfDaily":"2026-10-09","primaryCategory":"cs.CV","categories":["cs.CV","cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.06801","pdf":"https://arxiv.org/pdf/2610.06801","hf":"https://huggingface.co/papers/2610.06801","code":"https://github.com/dodododddo/mcsparse","project":"https://dodododddo.github.io/mcsparse-project-page/"},"metrics":{"hfUpvotes":46,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":46},"hot":53.77,"tags":["visual-gen","efficiency"],"summary":{"zh":"MC-Sparse 免训练地逐个挑选 KV token 并缓存补偿，缩小扩散 Transformer 稀疏与稠密注意力的质量差。","en":"MC-Sparse is training-free sparse attention that selects individual KV tokens with cached compensation, closing the dense-sparse gap in DiTs."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":179},{"id":"arxiv-2609.38879","arxiv":"2609.38879","title":"Does Learning Protein Folding Generalize to Broader Reasoning?","authors":["Yong Liu","Zhanpeng Shi","Yizhou Dang","Zhongyue Zhang","Xiaoliang Shi","Zhijian Wei","Shuangjia Zheng"],"authorCount":7,"org":"Shanghai JiaoTong University","affiliations":[],"orgNote":null,"date":"2026-09-30","hfDaily":"2026-10-05","primaryCategory":"cs.LG","categories":["cs.LG","cs.AI"],"links":{"abs":"https://arxiv.org/abs/2609.38879","pdf":"https://arxiv.org/pdf/2609.38879","hf":"https://huggingface.co/papers/2609.38879","code":"https://github.com/GENTEL-lab/Fold2Reason","project":null},"metrics":{"hfUpvotes":51,"hfComments":4,"citations":0,"citationsSource":"semantic_scholar","githubStars":112},"hot":53.75,"tags":["ai4science","reasoning"],"summary":{"zh":"研究让大模型学习蛋白质折叠能否迁移成通用的空间与结构推理能力。","en":"Asks whether learning protein folding teaches LLMs reusable spatial and structural reasoning."},"disciplines":["life-sciences"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":180},{"id":"arxiv-2610.27999","arxiv":"2610.27999","title":"Matrix-Game 4.0: Harnessing Video Models for Real-Time Playable Game Worlds","authors":["Jihai Zhang","Runjia Qian","Kaiqin Kong","Biao Jiang","Puyi Wang","Siru Chen","Tianle Li","Guanjie Chen"],"authorCount":18,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-10-20","hfDaily":"2026-10-21","primaryCategory":"cs.CV","categories":["cs.CV","cs.GR"],"links":{"abs":"https://arxiv.org/abs/2610.27999","pdf":"https://arxiv.org/pdf/2610.27999","hf":"https://huggingface.co/papers/2610.27999","code":null,"project":"https://matrix-game-v4-0.github.io/page/"},"metrics":{"hfUpvotes":182,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":53.74,"tags":["visual-gen","agent"],"summary":{"zh":"Matrix-Game 4.0 让轻量引擎和编程智能体管游戏逻辑、视频模型管画面，单张 RTX 5090 跑出 60 帧。","en":"Matrix-Game 4.0 separates game logic (engine plus coding agent) from visuals (autoregressive video renderer), reaching playable 60 FPS on one RTX 5090."},"disciplines":["software-engineering"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":181},{"id":"arxiv-2609.32259","arxiv":"2609.32259","title":"Prefill-Free Cross-Family KV Cache Transfer for Heterogeneous Multi-Agent LLMs","authors":["Vincent-Daniel Yun","Woosang Lim","Haneul Yoo","Sungjoo Yoo","Murali Annavaram","Sai Praneeth Karimireddy"],"authorCount":6,"org":"University of Southern California","affiliations":[],"orgNote":null,"date":"2026-09-26","hfDaily":"2026-10-02","primaryCategory":"cs.AI","categories":["cs.AI","cs.LG","cs.MA"],"links":{"abs":"https://arxiv.org/abs/2609.32259","pdf":"https://arxiv.org/pdf/2609.32259","hf":"https://huggingface.co/papers/2609.32259","code":"https://github.com/daniel-eai/Prefill-Free-Multi-Agent-LLMs","project":"https://github.com/daniel-eai/Prefill-Free-Multi-Agent-LLMs"},"metrics":{"hfUpvotes":103,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":7},"hot":53.72,"tags":["efficiency","agent"],"summary":{"zh":"HeteroFold 让不同家族的大模型之间直接传递 KV 缓存，多智能体协作时免去重复预填充。","en":"HeteroFold transfers KV caches across different model families, avoiding repeated prefill in heterogeneous multi-agent systems."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":182},{"id":"arxiv-2609.38334","arxiv":"2609.38334","title":"EVOKE: Eliciting World Knowledge in Agents for Transferable Decision-Making","authors":["Yuhan Guo","Jinming Liu","Liang Xu","Ziqiang Li","Jianguo Huang","Zhicheng Wang","Hu Zhu","Qiuyu Chen"],"authorCount":11,"org":"Shanghai Jiao Tong University","affiliations":[],"orgNote":null,"date":"2026-09-29","hfDaily":"2026-10-01","primaryCategory":"cs.CL","categories":["cs.CL"],"links":{"abs":"https://arxiv.org/abs/2609.38334","pdf":"https://arxiv.org/pdf/2609.38334","hf":"https://huggingface.co/papers/2609.38334","code":"https://github.com/Gnonymous/EVOKE","project":"https://gnonymous.github.io/EVOKE"},"metrics":{"hfUpvotes":75,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":27},"hot":53.72,"tags":["agent"],"summary":{"zh":"EVOKE 通过训练激发大模型预训练中已有的世界知识，让 Agent 更好地迁移到新环境。","en":"EVOKE elicits world knowledge already learned in pretraining so LLM agents transfer better to unseen environments."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":183},{"id":"arxiv-2610.13990","arxiv":"2610.13990","title":"DoSA: Dots Sparse Attention","authors":["Fanxu Meng","Taosong Fang"],"authorCount":2,"org":"dots studio","affiliations":[],"orgNote":null,"date":"2026-09-26","hfDaily":"2026-10-14","primaryCategory":"cs.LG","categories":["cs.LG"],"links":{"abs":"https://arxiv.org/abs/2610.13990","pdf":"https://arxiv.org/pdf/2610.13990","hf":"https://huggingface.co/papers/2610.13990","code":"https://github.com/llm-sys/flash-indexer","project":null},"metrics":{"hfUpvotes":81,"hfComments":1,"citations":0,"citationsSource":"semantic_scholar","githubStars":30},"hot":53.69,"tags":["efficiency"],"summary":{"zh":"DoSA 用上下界证明提前排除不可能入选的 key，在不改变 Top-K 结果的前提下加速稀疏注意力。","en":"DoSA uses bounds to reject keys early, speeding up sparse attention without changing exact Top-K selection."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":184},{"id":"arxiv-2609.35690","arxiv":"2609.35690","title":"Agent Priors-guided Policy Learning","authors":["Puming Jiang","Tianrun Hu","Haozhe Du","Yibo Li","Zhiwei Xue","Xinhu Li","Harold Soh"],"authorCount":7,"org":"National University of Singapore","affiliations":[],"orgNote":null,"date":"2026-09-28","hfDaily":"2026-10-02","primaryCategory":"cs.RO","categories":["cs.RO"],"links":{"abs":"https://arxiv.org/abs/2609.35690","pdf":"https://arxiv.org/pdf/2609.35690","hf":"https://huggingface.co/papers/2609.35690","code":"https://github.com/Agentics-robotics/Agent-Priors-guided-Policy-Learning","project":"https://agentics-robotics.github.io/APPL/"},"metrics":{"hfUpvotes":87,"hfComments":2,"citations":1,"citationsSource":"semantic_scholar","githubStars":5},"hot":53.67,"tags":["robotics","agent"],"summary":{"zh":"把每个技能策略的结构先验纳入 Agent 的技能接口，让少样本机器人同时做到组合泛化和技能泛化。","en":"Exposes each skill policy's structural prior to the agent so few-shot robots generalise both compositionally and per skill."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":185},{"id":"arxiv-2610.28354","arxiv":"2610.28354","title":"S-Agent 2.0: Spatial Reasoning with a General Coding Agent","authors":["Rui Gong","Ruoyu Wang","Yalun Dai","Hao Li","Sidi Yang","Qihao Zhao","Hongyu Li","Zeyang Guan"],"authorCount":12,"org":"Ropedia","affiliations":[],"orgNote":null,"date":"2026-10-20","hfDaily":"2026-10-21","primaryCategory":"cs.AI","categories":["cs.AI","cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.28354","pdf":"https://arxiv.org/pdf/2610.28354","hf":"https://huggingface.co/papers/2610.28354","code":"https://github.com/Ropedia/S-Agent-2.0","project":"https://ropedia.github.io/S-Agent-2.0"},"metrics":{"hfUpvotes":94,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":12},"hot":53.6,"tags":["agent","reasoning"],"summary":{"zh":"S-Agent 2.0 给通用编程智能体配上命令行 3D 感知工具和场景记忆，显著提升空间推理基准成绩。","en":"S-Agent 2.0 equips a general coding agent with command-line 3D perception tools and scene memory, lifting spatial-reasoning benchmarks."},"disciplines":["software-engineering"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":186},{"id":"arxiv-2609.31948","arxiv":"2609.31948","title":"Duplex-MPE: Benchmarking Multi-Party Interaction in Full-Duplex Dialogue","authors":["Chengqian Ma","Wenhao Feng","Weixuan Jin","Gaole Dai","Tianyu Xie","Yuexiao Ma","Zhaolu Kang","Xiangyu Zhao"],"authorCount":10,"org":"Peking University","affiliations":[],"orgNote":null,"date":"2026-09-25","hfDaily":"2026-09-29","primaryCategory":"cs.SD","categories":["cs.SD","eess.AS"],"links":{"abs":"https://arxiv.org/abs/2609.31948","pdf":"https://arxiv.org/pdf/2609.31948","hf":"https://huggingface.co/papers/2609.31948","code":null,"project":"https://step-out.github.io/Duplex-MPE-Page/"},"metrics":{"hfUpvotes":90,"hfComments":2,"citations":1,"citationsSource":"semantic_scholar","githubStars":null},"hot":53.53,"tags":["multimodal","benchmark"],"summary":{"zh":"Duplex-MPE 用 2000 个多人对话场景，评测全双工语音助手何时该接话、沉默或停下。","en":"Duplex-MPE uses 2,000 multi-party scenarios to test when a full-duplex voice assistant should answer, stay silent or stop."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":187},{"id":"arxiv-2610.01215","arxiv":"2610.01215","title":"AutoGUIWorld: Image Generators as Visual World Models for GUI Agent","authors":["Cheng Yang","Yifan Wu","Yutao Huang","Zhaohua Zhang","Beiduo Chen","Muxi Chen","Chenchen Zhao","Hexuan Deng"],"authorCount":21,"org":"Tencent Hunyuan","affiliations":[],"orgNote":null,"date":"2026-10-01","hfDaily":"2026-10-02","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.01215","pdf":"https://arxiv.org/pdf/2610.01215","hf":"https://huggingface.co/papers/2610.01215","code":"https://github.com/ImYangC7/AutoGUIWorld","project":null},"metrics":{"hfUpvotes":63,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":11},"hot":53.45,"tags":["agent","visual-gen"],"summary":{"zh":"AutoGUIWorld 用图像生成模型充当 GUI 的视觉世界模型，合成多样的界面交互轨迹。","en":"AutoGUIWorld uses image generators as visual world models of GUIs to synthesise diverse interaction trajectories."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":188},{"id":"arxiv-2610.06617","arxiv":"2610.06617","title":"RealtimeWAM: One-Step Asynchronous World Action Models","authors":["Chengtao Lv","Jinyang Du","Shuyi Feng","Yang Yong","Shiqiao Gu","Shunzi Yang","Ruihao Gong","Shen Ren"],"authorCount":10,"org":"Nanyang Technological University","affiliations":[],"orgNote":null,"date":"2026-10-05","hfDaily":"2026-10-06","primaryCategory":"cs.CV","categories":["cs.CV","cs.LG","cs.RO"],"links":{"abs":"https://arxiv.org/abs/2610.06617","pdf":"https://arxiv.org/pdf/2610.06617","hf":"https://huggingface.co/papers/2610.06617","code":"https://github.com/ModelTC/LightX2V","project":null},"metrics":{"hfUpvotes":23,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":2928},"hot":53.38,"tags":["robotics","efficiency"],"summary":{"zh":"RealtimeWAM 一步生成动作并异步推理，让世界-动作模型满足实时控制需求。","en":"RealtimeWAM generates actions in one step with asynchronous inference to make world-action models real-time."},"disciplines":["robotics","film-animation"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":189},{"id":"arxiv-2610.21783","arxiv":"2610.21783","title":"Learning to Think and Explore in 3D World","authors":["Yijiang Li","Bingyang Wang","Nuno Vasconcelos"],"authorCount":3,"org":"University of California at San Diego","affiliations":[],"orgNote":null,"date":"2026-10-15","hfDaily":"2026-10-19","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.21783","pdf":"https://arxiv.org/pdf/2610.21783","hf":"https://huggingface.co/papers/2610.21783","code":null,"project":"https://williamium3000.github.io/act3d/"},"metrics":{"hfUpvotes":175,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":53.35,"tags":["multimodal","rl","reasoning"],"summary":{"zh":"让视觉语言模型像人一样一次看一张图，在 3D 环境中边推理边移动视角再作答，用 RL 训练。","en":"Trains VLMs with RL to reason and actively move their viewpoint in 3D, seeing one image at a time like a human."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":190},{"id":"arxiv-2609.24058","arxiv":"2609.24058","title":"All-in-One Multilingual Scene Text Recognition with Script-aware Mixture-of-Experts","authors":["Xingsong Ye","Yongkun Du","Jiaxin Zhang","Zhixian Li","Chong Sun","Chen Li","Jing Lyu","Lianwen Jin"],"authorCount":9,"org":"Fudan University","affiliations":[],"orgNote":null,"date":"2026-09-21","hfDaily":"2026-09-23","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2609.24058","pdf":"https://arxiv.org/pdf/2609.24058","hf":"https://huggingface.co/papers/2609.24058","code":"https://github.com/YesianRohn/ScriptMoE","project":"https://huggingface.co/spaces/Yesianrohn/MultilingualOCR-Demo"},"metrics":{"hfUpvotes":56,"hfComments":3,"citations":2,"citationsSource":"semantic_scholar","githubStars":12},"hot":53.32,"tags":["vision","efficiency"],"summary":{"zh":"用按文字体系分路的混合专家，做一个比逐语言模型简单、比大视觉语言模型轻的多语种场景文字识别器。","en":"A script-aware mixture-of-experts gives one multilingual scene-text recogniser that is simpler than per-language models and lighter than VLMs."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":191},{"id":"arxiv-2609.35427","arxiv":"2609.35427","title":"LLMs are General Asynchronous Agents","authors":["George Yakushev","Denis Mazur","Vladimir Bartenev","Vyacheslav Zhdanovskiy","Timofey Byzov","Vladimir Kaurkin","Vadim Pastushenko"],"authorCount":7,"org":"Yandex Research","affiliations":[],"orgNote":null,"date":"2026-09-28","hfDaily":"2026-09-30","primaryCategory":"cs.LG","categories":["cs.LG","cs.CL"],"links":{"abs":"https://arxiv.org/abs/2609.35427","pdf":"https://arxiv.org/pdf/2609.35427","hf":"https://huggingface.co/papers/2609.35427","code":"https://github.com/dvmazur/async_llm","project":"https://dvmazur.github.io/async_llm"},"metrics":{"hfUpvotes":86,"hfComments":1,"citations":0,"citationsSource":"semantic_scholar","githubStars":20},"hot":53.31,"tags":["agent"],"summary":{"zh":"提出通用异步 Agent：大模型在思考或执行时也能接收新输入，适配语音、具身、监控等并发场景。","en":"Proposes general asynchronous agents that accept new inputs while thinking or acting, covering voice, embodied and monitoring settings."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":192},{"id":"arxiv-2610.01257","arxiv":"2610.01257","title":"Science Utopia? Closed-Loop LLM Simulation of Academic Research Ecosystems","authors":["Yiqiao Jin","Yiyang Wang","Lucheng Fu","Bing He","Siheng Xiong","Yijia Xiao","B. Aditya Prakash","Josiah Hester"],"authorCount":11,"org":"Georgia Institute of Technology","affiliations":[],"orgNote":null,"date":"2026-10-01","hfDaily":"2026-10-05","primaryCategory":"cs.CL","categories":["cs.CL","cs.AI","cs.CY"],"links":{"abs":"https://arxiv.org/abs/2610.01257","pdf":"https://arxiv.org/pdf/2610.01257","hf":"https://huggingface.co/papers/2610.01257","code":"https://github.com/Ahren09/ScienceUtopia","project":"https://ahren09.github.io/ScienceUtopia/"},"metrics":{"hfUpvotes":47,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":164},"hot":53.12,"tags":["agent","ai4science"],"summary":{"zh":"SciUtopia 用大模型 Agent 闭环模拟学术生态：选题、合作、投稿、评审、引用和资助。","en":"SciUtopia simulates academic ecosystems with LLM agents: topic choice, collaboration, review, citation and funding."},"disciplines":["economics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":193},{"id":"arxiv-2610.22791","arxiv":"2610.22791","title":"Self-Evolving Coding Agents via Harness-Policy Optimization","authors":["Zhepei Wei","Wenlin Yao","Baolin Peng","Danqing Wang","Qianhui Wu","Hao Cheng","Yu Meng","Jianfeng Gao"],"authorCount":8,"org":"Microsoft Research","affiliations":[],"orgNote":null,"date":"2026-10-16","hfDaily":"2026-10-20","primaryCategory":"cs.CL","categories":["cs.CL","cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.22791","pdf":"https://arxiv.org/pdf/2610.22791","hf":"https://huggingface.co/papers/2610.22791","code":"https://github.com/MSR-Orchard/Self-HPO","project":"https://weizhepei.com/blog/self-hpo/"},"metrics":{"hfUpvotes":45,"hfComments":4,"citations":0,"citationsSource":"semantic_scholar","githubStars":28},"hot":53.12,"tags":["code","agent","rl"],"summary":{"zh":"微软 Self-HPO 交替优化编程智能体的“外壳”（harness）和策略模型，实现递归自我改进。","en":"Microsoft's Self-HPO alternately evolves a coding agent's harness and policy model for recursive self-improvement."},"disciplines":["software-engineering"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":194},{"id":"arxiv-2610.08760","arxiv":"2610.08760","title":"WorldSonus: Bringing Sound to Worlds","authors":["Pengjun Fang","Jingyi Fa","Kam Man Wu","Jiaming Wang","Haoyuan Huang","Yaguang Wu","Xiangjun Huang","Ziyang Ma"],"authorCount":12,"org":"NoizAI","affiliations":[],"orgNote":null,"date":"2026-10-06","hfDaily":"2026-10-08","primaryCategory":"cs.SD","categories":["cs.SD","cs.AI","cs.CV","eess.AS"],"links":{"abs":"https://arxiv.org/abs/2610.08760","pdf":"https://arxiv.org/pdf/2610.08760","hf":"https://huggingface.co/papers/2610.08760","code":"https://github.com/NoizAI/WorldSonus","project":"https://noizai.github.io/WorldSonus/"},"metrics":{"hfUpvotes":41,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":233},"hot":53.09,"tags":["visual-gen","multimodal"],"summary":{"zh":"WorldSonus 为世界模型实时生成与画面和相机运动对齐的立体声，并支持中途指令控制。","en":"WorldSonus adds real-time, spatially aligned stereo sound to world models with mid-stream sound control."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":195},{"id":"arxiv-2609.36322","arxiv":"2609.36322","title":"Periodic Weak Spots: Phase Sensitivity from Chunked KV-Cache Compression","authors":["Xingyu Zhu"," Pu"," Yi","Ziheng Cheng","Ang Lv","Jing Liu","Lexing Ying","Yiyuan Ma"],"authorCount":9,"org":"ByteDance Seed","affiliations":["Luke"],"orgNote":null,"date":"2026-09-28","hfDaily":"2026-09-30","primaryCategory":"cs.LG","categories":["cs.LG","cs.AI"],"links":{"abs":"https://arxiv.org/abs/2609.36322","pdf":"https://arxiv.org/pdf/2609.36322","hf":"https://huggingface.co/papers/2609.36322","code":null,"project":"https://ultimatejupiter.github.io/blog/periodic-weak-spots/"},"metrics":{"hfUpvotes":108,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":52.99,"tags":["efficiency"],"summary":{"zh":"发现分块 KV 缓存压缩带来“相位敏感”：信息在压缩窗口中的位置不同，检索准确率差异明显。","en":"Finds chunked KV-cache compression causes phase sensitivity: retrieval accuracy varies with position relative to window boundaries."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":196},{"id":"arxiv-2610.19975","arxiv":"2610.19975","title":"LatentLoop: Scaling Looped MoE Models with Asymmetric Compute Allocation","authors":["Zefeng He","Yunkai Chai","Xiaoyang Chen","Tong Zhu","Zongliang Wu","Yun Luo","Yafu Li","Yu Cheng"],"authorCount":8,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-10-14","hfDaily":"2026-10-15","primaryCategory":"cs.LG","categories":["cs.LG"],"links":{"abs":"https://arxiv.org/abs/2610.19975","pdf":"https://arxiv.org/pdf/2610.19975","hf":"https://huggingface.co/papers/2610.19975","code":"https://github.com/lcqysl/LatentLoop","project":"https://latentloop-project.github.io/"},"metrics":{"hfUpvotes":82,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":13},"hot":52.87,"tags":["efficiency"],"summary":{"zh":"LatentLoop 发现循环 MoE 中两次循环作用不对称，据此非对称分配算力以提升效果。","en":"LatentLoop finds loop visits in looped MoE play asymmetric roles and allocates compute asymmetrically for better results."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":197},{"id":"arxiv-2609.38079","arxiv":"2609.38079","title":"OmniTaskonomy: When Does Visual Generation Improve Visual Understanding?","authors":["Jiaxin Ge","Yiming Qin","Ji Xie","Haozhe Jiang","Xiaochuang Han","Junyi Zhang","Andrew Dai","Yinfei Yang"],"authorCount":16,"org":"UC Berkeley","affiliations":[],"orgNote":null,"date":"2026-09-29","hfDaily":"2026-09-30","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2609.38079","pdf":"https://arxiv.org/pdf/2609.38079","hf":"https://huggingface.co/papers/2609.38079","code":"https://github.com/para-lost/OmniTaskonomy","project":"https://omni-taskonomy.github.io/"},"metrics":{"hfUpvotes":57,"hfComments":1,"citations":0,"citationsSource":"semantic_scholar","githubStars":17},"hot":52.87,"tags":["multimodal","vision"],"summary":{"zh":"研究视觉生成训练何时能提升视觉理解：配方对时，图生图训练能带来理解任务的增益。","en":"Studies when visual generation training improves understanding, finding image-to-image training helps under the right recipe."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":198},{"id":"arxiv-2610.20556","arxiv":"2610.20556","title":"Speculative Sampling in Interactive Video Generation","authors":["Yubo Huang","Yuxuan Zhang","Jiaming Liu","Sirui Zhao","Enhong Chen"],"authorCount":5,"org":"university of science and technology  of china","affiliations":[],"orgNote":null,"date":"2026-10-14","hfDaily":"2026-10-19","primaryCategory":"cs.CV","categories":["cs.CV","cs.HC"],"links":{"abs":"https://arxiv.org/abs/2610.20556","pdf":"https://arxiv.org/pdf/2610.20556","hf":"https://huggingface.co/papers/2610.20556","code":"https://github.com/Yubo-Shankui/SpeculativeSampling","project":"https://yubo-shankui.github.io/speculative_sampling/"},"metrics":{"hfUpvotes":84,"hfComments":4,"citations":0,"citationsSource":"semantic_scholar","githubStars":10},"hot":52.84,"tags":["visual-gen","efficiency"],"summary":{"zh":"交互式视频生成中的“推测采样”：用大窗口预测并缓存结果，兼顾画质与低延迟。","en":"Speculative Sampling for interactive video generation caches predictions from large windows to get large-chunk quality at small-chunk latency."},"disciplines":["film-animation"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":199},{"id":"arxiv-2609.40325","arxiv":"2609.40325","title":"WorldAuditBench: Interactive 3D World Auditing with Multimodal Agents","authors":["Ziyan Jiang","Jingbo Yang","Jiabao Ji","Yujian Liu","Qiucheng Wu","Tommi Jaakkola","Yang Zhang","Shiyu Chang"],"authorCount":8,"org":"University of California, Santa Barbara","affiliations":[],"orgNote":null,"date":"2026-09-30","hfDaily":"2026-10-01","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2609.40325","pdf":"https://arxiv.org/pdf/2609.40325","hf":"https://huggingface.co/papers/2609.40325","code":"https://github.com/UCSB-NLP-Chang/WorldAuditBench","project":"https://ucsb-nlp-chang.github.io/WorldAuditBench/"},"metrics":{"hfUpvotes":106,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":4},"hot":52.83,"tags":["multimodal","agent","benchmark"],"summary":{"zh":"WorldAuditBench 让多模态 Agent 在 3D 世界中移动搜寻漂浮物体、穿墙等异常。","en":"WorldAuditBench has multimodal agents navigate 3D worlds to find anomalies like floating objects or walk-through walls."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":200},{"id":"arxiv-2610.06833","arxiv":"2610.06833","title":"Towards Looped Models Done Right, Part II: Rethinking at Fixed Points","authors":["Benhao Huang","Chufan Shi","Junlin Chen","Shicheng Wen","Zhengzhong Liu","Eric Xing","Xuezhe Ma"],"authorCount":7,"org":"Institute of Foundation Models","affiliations":[],"orgNote":null,"date":"2026-10-05","hfDaily":"2026-10-06","primaryCategory":"cs.LG","categories":["cs.LG"],"links":{"abs":"https://arxiv.org/abs/2610.06833","pdf":"https://arxiv.org/pdf/2610.06833","hf":"https://huggingface.co/papers/2610.06833","code":"https://github.com/ifm-ai/xllm-loop","project":"https://x.com/huskydogewoof/status/2106130737007091939"},"metrics":{"hfUpvotes":35,"hfComments":2,"citations":3,"citationsSource":"semantic_scholar","githubStars":49},"hot":52.81,"tags":["efficiency","llm"],"summary":{"zh":"循环语言模型的状态越接近不动点，训练、解码和强化学习都能走捷径，从而大幅省算力。","en":"Driving looped language models towards fixed points enables truncated backprop, KV sharing and cheaper RL with little accuracy loss."},"disciplines":["cybersecurity"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":201},{"id":"arxiv-2610.00531","arxiv":"2610.00531","title":"Science or Slop?: Benchmarking and Mitigating Scientific Slop in AI-Generated Papers","authors":["Yerim Oh","Young-Jun Lee","Jaewoo Ahn","Gunhee Kim","Dongyeop Kang"],"authorCount":5,"org":"Seoul National University","affiliations":[],"orgNote":null,"date":"2026-09-30","hfDaily":"2026-10-05","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.00531","pdf":"https://arxiv.org/pdf/2610.00531","hf":"https://huggingface.co/papers/2610.00531","code":"https://github.com/yerimoh/ScientificSlop","project":"https://yerimoh.github.io/scientific-slop-demo/"},"metrics":{"hfUpvotes":63,"hfComments":2,"citations":1,"citationsSource":"semantic_scholar","githubStars":14},"hot":52.78,"tags":["benchmark","ai4science"],"summary":{"zh":"SciSlopBench 收录 390 篇 AI 生成论文，检测看似合理但推理断裂的“科研垃圾”，并探索缓解办法。","en":"SciSlopBench collects 390 AI-generated papers to detect plausible-looking but broken 'scientific slop' and explores mitigations."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":202},{"id":"arxiv-2610.14136","arxiv":"2610.14136","title":"Correction Training for Diffusion Models","authors":["Donghoon Ahn","Runqian Wang","Alexei A. Efros","Taesung Park","Yash Belhe"],"authorCount":5,"org":"University of California, Berkeley","affiliations":[],"orgNote":null,"date":"2026-10-09","hfDaily":"2026-10-21","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.14136","pdf":"https://arxiv.org/pdf/2610.14136","hf":"https://huggingface.co/papers/2610.14136","code":"https://github.com/sunovivid/correction-training","project":"https://sunovivid.github.io/correction-training/"},"metrics":{"hfUpvotes":37,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":67},"hot":52.57,"tags":["visual-gen"],"summary":{"zh":"为缓解扩散模型的曝光偏差，额外训练一个“校正器”在每个去噪步修正预测，在 DiT、SiT、JiT 上明显提升质量。","en":"Trains a corrector that refines the denoiser's prediction at each step to reduce exposure bias, improving DiT, SiT and JiT samples."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":203},{"id":"arxiv-2610.13021","arxiv":"2610.13021","title":"EmbodiedCLBench: Evaluating Continual Learning for Self-Evolving Embodied Agents","authors":["Jie Huang","Yanan Chen","Ruixun Liu","Kaichen He","Xinyi He","Rui Huang","Zilong Zheng","Zhenliang Zhang"],"authorCount":9,"org":"PKU-VaLuE-Lab","affiliations":[],"orgNote":null,"date":"2026-10-09","hfDaily":"2026-10-12","primaryCategory":"cs.AI","categories":["cs.AI","cs.CL","cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.13021","pdf":"https://arxiv.org/pdf/2610.13021","hf":"https://huggingface.co/papers/2610.13021","code":"https://github.com/PKU-VaLuE-Lab/EmbodiedCLBench","project":"https://pku-value-lab.github.io/EmbodiedCLBench-homepage/"},"metrics":{"hfUpvotes":74,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":17},"hot":52.48,"tags":["robotics","agent","benchmark"],"summary":{"zh":"EmbodiedCLBench 是首个具身仿真中的持续学习基准，考察 Agent 能否组合旧技能完成新任务。","en":"EmbodiedCLBench is the first embodied continual-learning benchmark, testing whether agents recombine learned skills on new tasks."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":204},{"id":"arxiv-2609.38173","arxiv":"2609.38173","title":"In-context Robot Learning Made Simple: A Democratized Recipe for Manipulation Tasks","authors":["Minxing Li","Minghao Han","Weizhi Zhao","Hanwen Wang","Xiangshuo Liu","Shuyao Shang","Jingxiang Zhou","Mingchao Sun"],"authorCount":13,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-09-29","hfDaily":"2026-10-09","primaryCategory":"cs.RO","categories":["cs.RO"],"links":{"abs":"https://arxiv.org/abs/2609.38173","pdf":"https://arxiv.org/pdf/2609.38173","hf":"https://huggingface.co/papers/2609.38173","code":"https://github.com/simpleicl/simple-icl","project":"https://simpleicl.github.io/simpleicl"},"metrics":{"hfUpvotes":64,"hfComments":7,"citations":0,"citationsSource":"semantic_scholar","githubStars":18},"hot":52.22,"tags":["robotics"],"summary":{"zh":"先明确定义机器人上下文学习要学什么，再给出一套简单易复现的操作任务方案。","en":"Clearly defines what robot in-context learning should follow and offers a minimalist, reproducible recipe for manipulation."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":205},{"id":"arxiv-2609.36071","arxiv":"2609.36071","title":"LongCat-DeepResearch Technical Report","authors":[" Meituan LongCat Team","He Zhu","Yue Xu","Wanli Wu","Haolin Ren","Yuxin Bian","Jiarui Zhao","Rongzhi Zhang"],"authorCount":31,"org":"LongCat","affiliations":[],"orgNote":null,"date":"2026-09-28","hfDaily":"2026-09-30","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2609.36071","pdf":"https://arxiv.org/pdf/2609.36071","hf":"https://huggingface.co/papers/2609.36071","code":"https://github.com/meituan-longcat/LongCat-DeepResearch","project":"https://meituan-longcat.github.io/LongCat-DeepResearch/"},"metrics":{"hfUpvotes":74,"hfComments":1,"citations":0,"citationsSource":"semantic_scholar","githubStars":23},"hot":52.16,"tags":["agent"],"summary":{"zh":"美团 LongCat 深度研究系统技术报告：多 Agent 分工规划、并行调研和分节修订，生成有证据的长报告。","en":"Technical report for LongCat-DeepResearch, a multi-agent workflow that plans, researches in parallel and revises evidence-grounded reports."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":206},{"id":"arxiv-2610.22174","arxiv":"2610.22174","title":"Mosaic: Multi-Session Agentic Reinforcement Learning","authors":["Yilong Xu","Minrui Xu","Xinyu Zhou","Xiang Long","Haochen Tan","Zhijiang Guo"],"authorCount":6,"org":"LARK Lab@HKUST (GZ)","affiliations":[],"orgNote":null,"date":"2026-10-15","hfDaily":"2026-10-16","primaryCategory":"cs.CL","categories":["cs.CL","cs.AI","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2610.22174","pdf":"https://arxiv.org/pdf/2610.22174","hf":"https://huggingface.co/papers/2610.22174","code":"https://github.com/LARK-AI-Lab/Mosaic","project":"https://lark-ai-lab.github.io/Mosaic/"},"metrics":{"hfUpvotes":62,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":36},"hot":52.11,"tags":["rl","agent"],"summary":{"zh":"Mosaic 把智能体强化学习扩展到多会话轨迹，匹配上下文压缩和子智能体分叉等真实推理方式。","en":"Mosaic generalises agentic RL to multi-session trajectories, matching inference-time context compaction and sub-agent forking."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":207},{"id":"arxiv-2610.00314","arxiv":"2610.00314","title":"Predictive Credit: Measuring What Scientific Explanations Add to Experimental Forecasts","authors":["Jingjie Ning","Xueqi Li","Yibo Kong","Dongting Li"],"authorCount":4,"org":"Carnegie Mellon University","affiliations":[],"orgNote":null,"date":"2026-09-29","hfDaily":"2026-10-02","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.00314","pdf":"https://arxiv.org/pdf/2610.00314","hf":"https://huggingface.co/papers/2610.00314","code":null,"project":null},"metrics":{"hfUpvotes":103,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":52.09,"tags":["ai4science","benchmark"],"summary":{"zh":"提出“预测贡献”指标，衡量科研 Agent 对实验的解释是否真能改善结果预测；预注册检验结论为不确定。","en":"Proposes 'predictive credit' to measure whether research agents' explanations improve experimental forecasts; preregistered tests were inconclusive."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":208},{"id":"arxiv-2610.06293","arxiv":"2610.06293","title":"VepAgent: Bridging Causal-Transition via Tool-Augmented Reinforcement Learning for Video Event Prediction","authors":["Qiutong Chen","Yuchan Guo","Zhenlong Yuan","Haobo Yang","Fangfang Lin","Xinyi Long","Yin Wang","Zijian Song"],"authorCount":13,"org":"UCSC-VLAA","affiliations":[],"orgNote":null,"date":"2026-10-05","hfDaily":"2026-10-08","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.06293","pdf":"https://arxiv.org/pdf/2610.06293","hf":"https://huggingface.co/papers/2610.06293","code":null,"project":null},"metrics":{"hfUpvotes":145,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":51.92,"tags":["multimodal","rl","reasoning"],"summary":{"zh":"VepAgent 把因果转移推理与工具增强强化学习结合，用于预测视频中接下来会发生的事件。","en":"VepAgent combines causal-transition reasoning with tool-augmented RL for video event prediction."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":209},{"id":"arxiv-2610.23121","arxiv":"2610.23121","title":"Scene-Designer: Learning Design Intelligence for Agentic 3D Scene Generation","authors":["Xuanyu Liu","Hongyan Feng","Miao Pan","Sunlai Chen","Lizhou Liang","Bing Sun","Wenqi Zhang","Xuhong Zhang"],"authorCount":8,"org":"ZJU-OmniAI","affiliations":[],"orgNote":null,"date":"2026-10-16","hfDaily":"2026-10-19","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.23121","pdf":"https://arxiv.org/pdf/2610.23121","hf":"https://huggingface.co/papers/2610.23121","code":"https://github.com/ZJU-OmniAI/Scene-Designer","project":"https://zju-omniai.github.io/Scene-Designer/"},"metrics":{"hfUpvotes":62,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":27},"hot":51.84,"tags":["agent","vision"],"summary":{"zh":"Scene-Designer 从设计与评审两类经验中学习“设计智能”，用文字生成完整可用的 3D 室内场景。","en":"Scene-Designer learns design intelligence from design and critique experience to generate coherent 3D indoor scenes from text."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":210},{"id":"arxiv-2610.17309","arxiv":"2610.17309","title":"StartLux-Decision: AutoThink Decision Models with Reinforcement Learning for Calibrated Deliberative Decisions","authors":["Xialie Zhuang","Zunhai Su","Jia Pan","Ngai Wong","Guanhua Chen","Chuan-Wei Kuo"],"authorCount":6,"org":"StartLux","affiliations":[],"orgNote":null,"date":"2026-10-12","hfDaily":"2026-10-15","primaryCategory":"cs.CL","categories":["cs.CL","cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.17309","pdf":"https://arxiv.org/pdf/2610.17309","hf":"https://huggingface.co/papers/2610.17309","code":"https://github.com/StartLuxLabs/StartLux-Decision","project":"https://decision.startlux.com/"},"metrics":{"hfUpvotes":24,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":1110},"hot":51.8,"tags":["reasoning","rl"],"summary":{"zh":"StartLux-Decision 决策模型（0.8B–27B）对有把握的情况直接作答、不确定时才推理，并用强化学习校准决策概率。","en":"StartLux-Decision models (0.8B–27B) answer confident cases directly and reason only when uncertain, with RL that rewards calibrated decision probabilities."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":211},{"id":"arxiv-2609.29892","arxiv":"2609.29892","title":"Qwen-Planner-Agent: A Closed-Loop AI-for-AI Framework for Real-World Mobile Planner Agents","authors":["Tingyu Qu","Weigao Sun","Yuecheng Liu","Yucheng Zhao","Yi Zhu","Yifeng Ding","Qiyi Wang","Sihan Cao"],"authorCount":26,"org":"Tongyi-MAI","affiliations":[],"orgNote":null,"date":"2026-09-24","hfDaily":"2026-09-25","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2609.29892","pdf":"https://arxiv.org/pdf/2609.29892","hf":"https://huggingface.co/papers/2609.29892","code":"https://github.com/Tongyi-MAI/Qwen-Planner-Agent","project":"https://tongyi-mai.github.io/Qwen-Planner-Agent/"},"metrics":{"hfUpvotes":34,"hfComments":2,"citations":1,"citationsSource":"semantic_scholar","githubStars":111},"hot":51.77,"tags":["agent","rl"],"summary":{"zh":"通义 Qwen-Planner-Agent 用“AI 开发 AI”的闭环框架迭代训练真实手机上的规划智能体。","en":"Qwen-Planner-Agent uses a closed-loop AI-for-AI framework to iteratively build real-world mobile planner agents."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":212},{"id":"arxiv-2609.25270","arxiv":"2609.25270","title":"RULER: Instance-aware Rubric Rewards for SVG Generation","authors":["Hangyu Ran","Yuhao Zheng","Yingying Zhang","Kevin Qinghong Lin","Han Peng"],"authorCount":5,"org":"inclusionAI","affiliations":[],"orgNote":null,"date":"2026-09-21","hfDaily":"2026-09-23","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2609.25270","pdf":"https://arxiv.org/pdf/2609.25270","hf":"https://huggingface.co/papers/2609.25270","code":null,"project":"https://hangyuran.github.io/RULER/"},"metrics":{"hfUpvotes":104,"hfComments":4,"citations":1,"citationsSource":"semantic_scholar","githubStars":null},"hot":51.73,"tags":["rl","visual-gen"],"summary":{"zh":"RULER 用针对每条指令的多维评分细则当奖励，训练模型生成 SVG 矢量图，缓解奖励作弊。","en":"RULER uses instance-aware multi-axis rubrics as rewards for SVG generation, reducing the reward hacking seen with scalar metrics."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":213},{"id":"arxiv-2610.27628","arxiv":"2610.27628","title":"PAR: Progress-Aligned Recurrence for Length Generalization in State Tracking","authors":["Pan Wang","Yihao Hu","Guangyuan Dong","Yitong Lin","Hang Wang","Zimi Liu","Zhihao Wen","Jianshe Li"],"authorCount":11,"org":"ant-international","affiliations":[],"orgNote":null,"date":"2026-10-20","hfDaily":"2026-10-21","primaryCategory":"cs.CL","categories":["cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.27628","pdf":"https://arxiv.org/pdf/2610.27628","hf":"https://huggingface.co/papers/2610.27628","code":null,"project":null},"metrics":{"hfUpvotes":148,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":51.69,"tags":["llm","reasoning"],"summary":{"zh":"PAR 让循环语言模型每轮迭代对齐“已完成的步数”并锚定状态范数，状态追踪任务可外推到训练长度的 16 倍。","en":"PAR aligns each loop iteration of a looped LM with task progress and anchors state norms, extrapolating state tracking to 16x training length."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":214},{"id":"arxiv-2610.07557","arxiv":"2610.07557","title":"CheckerBench: Can Long-Horizon Agents Synthesize Static-Analysis Checkers?","authors":["Hang He","Li Wang","Hao Chen","Yuchen Shao","Yuling Shi","Lisheng Wang","Peiyang Liu","Goose Lin"],"authorCount":12,"org":"Humanlaya Data Lab","affiliations":[],"orgNote":null,"date":"2026-10-06","hfDaily":"2026-10-07","primaryCategory":"cs.SE","categories":["cs.SE","cs.AI","cs.CR"],"links":{"abs":"https://arxiv.org/abs/2610.07557","pdf":"https://arxiv.org/pdf/2610.07557","hf":"https://huggingface.co/papers/2610.07557","code":"https://github.com/ahang0712/CheckerBench-Dataset","project":null},"metrics":{"hfUpvotes":72,"hfComments":2,"citations":2,"citationsSource":"semantic_scholar","githubStars":2},"hot":51.69,"tags":["code","agent","benchmark"],"summary":{"zh":"CheckerBench 用源自 297 个 CVE 的 300 个任务，测智能体能否在真实仓库写出静态分析检查器。","en":"CheckerBench tests whether long-horizon agents can build working static-analysis checkers, with 300 tasks from 297 CVEs."},"disciplines":["cybersecurity"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":215},{"id":"arxiv-2610.05162","arxiv":"2610.05162","title":"Memadapter: Counterfactual Adaptation Against Memory-induced Sycophancy","authors":["Ruqing Ning","Haibo Meng","Zhishang Xiang","Zerui Chen","Jinsong Su","Xin Wang","Qinggang Zhang"],"authorCount":7,"org":"DEEP Group at Jilin University","affiliations":[],"orgNote":null,"date":"2026-10-04","hfDaily":"2026-10-06","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.05162","pdf":"https://arxiv.org/pdf/2610.05162","hf":"https://huggingface.co/papers/2610.05162","code":"https://github.com/DEEP-JLU/MemAdapter","project":"https://github.com/DEEP-JLU/MemAdapter.git"},"metrics":{"hfUpvotes":62,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":30},"hot":51.66,"tags":["alignment","agent"],"summary":{"zh":"Memadapter 用反事实适配缓解长期记忆引发的谄媚，即使记忆本身正确也会让 Agent 迎合用户。","en":"Memadapter uses counterfactual adaptation to reduce memory-induced sycophancy, which arises even from correct memories."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":216},{"id":"arxiv-2610.17227","arxiv":"2610.17227","title":"AReA: Adaptive Receptive-Field Attention for Efficient High-Resolution Diffusion Transformers","authors":["Javad Rajabi","Koorosh Roohi","David B. Lindell","Babak Taati"],"authorCount":4,"org":"University of Toronto Computer Science","affiliations":[],"orgNote":null,"date":"2026-10-12","hfDaily":"2026-10-14","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.17227","pdf":"https://arxiv.org/pdf/2610.17227","hf":"https://huggingface.co/papers/2610.17227","code":"https://github.com/KooroshRH/AReA","project":"https://rajabi2001.github.io/area/"},"metrics":{"hfUpvotes":58,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":38},"hot":51.58,"tags":["visual-gen","efficiency"],"summary":{"zh":"AReA 让扩散 Transformer 的注意力感受野随去噪步骤自适应收窄，更省算力地生成超高分辨率图像。","en":"AReA adapts the attention receptive field across denoising steps so diffusion transformers generate beyond-training resolutions more cheaply."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":217},{"id":"arxiv-2610.19366","arxiv":"2610.19366","title":"Seeing, Maintaining, and Learning: The Evolution of Multimodal Agent Memory","authors":["Minghao Guo","Qingyue Jiao","Xiujin Liu","Jianhong Pang","Ruimeng Yang","Naihao Xu","Liwei Che","Haochen Xue"],"authorCount":80,"org":"Fudan University","affiliations":[],"orgNote":null,"date":"2026-10-14","hfDaily":"2026-10-15","primaryCategory":"cs.AI","categories":["cs.AI","cs.CL","cs.RO"],"links":{"abs":"https://arxiv.org/abs/2610.19366","pdf":"https://arxiv.org/pdf/2610.19366","hf":"https://huggingface.co/papers/2610.19366","code":"https://github.com/Seeing-Maintaining-Learning/multimodal-agent-memory-evolution","project":null},"metrics":{"hfUpvotes":55,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":37},"hot":51.43,"tags":["multimodal","agent"],"summary":{"zh":"综述多模态智能体记忆机制的演进：从感知、维护到可学习的记忆管理。","en":"A survey of multimodal agent memory, from perceiving and maintaining memories to learned memory-management policies."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":218},{"id":"arxiv-2609.38653","arxiv":"2609.38653","title":"TERRA: Terrain-Aware Reconstruction, Retargeting and Control for Musculoskeletal Locomotion","authors":["Merkourios Simos","Chengkun Li","Bianca Ziliotto","Alexander Mathis"],"authorCount":4,"org":"Mathis Group @ EPFL","affiliations":[],"orgNote":null,"date":"2026-09-29","hfDaily":"2026-10-01","primaryCategory":"cs.RO","categories":["cs.RO","cs.AI","cs.LG","q-bio.NC"],"links":{"abs":"https://arxiv.org/abs/2609.38653","pdf":"https://arxiv.org/pdf/2609.38653","hf":"https://huggingface.co/papers/2609.38653","code":"https://github.com/amathislab/terra","project":"https://cnai.epfl.ch/terra/"},"metrics":{"hfUpvotes":40,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":153},"hot":51.38,"tags":["robotics","rl"],"summary":{"zh":"TERRA 把运动数据连同地形一起重定向到肌肉骨骼模型，让肌肉驱动的智能体在复杂地形上行走。","en":"TERRA retargets motion with terrain to musculoskeletal models so muscle-driven agents can walk on complex ground."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":219},{"id":"arxiv-2610.20572","arxiv":"2610.20572","title":"Recovering Reliable Supervision via Critique-Guided On-Policy Distillation","authors":["Kaixiang Zhao","Tianrun Yu","Shangzhe Li","Yuxiao Yang","Amanda Hughes","Weitong Zhang","Taylor W. Killian"],"authorCount":7,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-10-14","hfDaily":"2026-10-15","primaryCategory":"cs.LG","categories":["cs.LG","cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.20572","pdf":"https://arxiv.org/pdf/2610.20572","hf":"https://huggingface.co/papers/2610.20572","code":"https://github.com/kzhao5/CriticOPD","project":"https://kzhao5.github.io/CriticOPD-website/"},"metrics":{"hfUpvotes":91,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":4},"hot":51.32,"tags":["efficiency","reasoning"],"summary":{"zh":"CriticOPD 在在线蒸馏中先让教师点评学生的错误再改写，解决教师跟着学生错下去的问题。","en":"CriticOPD has the teacher critique wrong student rollouts before rewriting, restoring reliable supervision in on-policy distillation."},"disciplines":["mathematics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":220},{"id":"arxiv-2609.33304","arxiv":"2609.33304","title":"Relevance Does Not Imply Applicability: Experience Activation for Personal GUI Agents","authors":["Fuyao Zhang","Xuan Wang","Zherui Li","Jiaming Zhang","Longtao Huang","Wei Yang Bryan Lim"],"authorCount":6,"org":"Nanyang Technological University","affiliations":[],"orgNote":null,"date":"2026-09-27","hfDaily":"2026-10-13","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2609.33304","pdf":"https://arxiv.org/pdf/2609.33304","hf":"https://huggingface.co/papers/2609.33304","code":"https://github.com/fyzhang1/ExpActivator","project":"https://fyzhang1.github.io/ExpActivator/"},"metrics":{"hfUpvotes":46,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":72},"hot":51.31,"tags":["agent"],"summary":{"zh":"研究个人 GUI 智能体：检索到的历史经验主要只帮第一步，提出按需“激活”经验的方法。","en":"Finds retrieved history mainly helps a personal GUI agent's first step, and proposes activating experience only when it applies."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":221},{"id":"arxiv-2609.37226","arxiv":"2609.37226","title":"Follow the Entities: A Corpus Map for Agentic Search","authors":["Soyeong Jeong","Sujay Kumar Jauhar","Sung Ju Hwang","Andrew Joohun Nam"],"authorCount":4,"org":"Microsoft","affiliations":[],"orgNote":null,"date":"2026-09-29","hfDaily":"2026-09-30","primaryCategory":"cs.CL","categories":["cs.CL","cs.AI","cs.IR","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2609.37226","pdf":"https://arxiv.org/pdf/2609.37226","hf":"https://huggingface.co/papers/2609.37226","code":null,"project":null},"metrics":{"hfUpvotes":101,"hfComments":1,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":51.29,"tags":["agent","llm"],"summary":{"zh":"微软为大型文档库构建以实体为线索的“语料地图”，帮助智能体搜索跨文档的证据。","en":"Microsoft builds an entity-based corpus map so search agents can follow evidence across documents instead of rediscovering links."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":222},{"id":"arxiv-2610.20442","arxiv":"2610.20442","title":"Towards Agentic Robots: Model, Data, Environment, and Harness","authors":["Chonghan Qin","Dongming Wu","Kaituo Feng","Jiaxu Wang","Xinlei Yu","Peiwen Sun","Yijia Fan","Xilin He"],"authorCount":16,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-10-14","hfDaily":"2026-10-15","primaryCategory":"cs.RO","categories":["cs.RO"],"links":{"abs":"https://arxiv.org/abs/2610.20442","pdf":"https://arxiv.org/pdf/2610.20442","hf":"https://huggingface.co/papers/2610.20442","code":"https://github.com/chonghanqin/Awesome-Agentic-Robots","project":null},"metrics":{"hfUpvotes":43,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":111},"hot":51.29,"tags":["robotics","agent"],"summary":{"zh":"“Agent 化机器人”综述，从模型、数据、环境和执行框架四方面梳理这一新范式。","en":"A survey of agentic robots organised around model, data, environment and harness."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":223},{"id":"arxiv-2610.23986","arxiv":"2610.23986","title":"BlenderWorld: Benchmarking Visual Agents for 3D Creation from Real-World References","authors":["Xinyu Fang","Yuchi Wang","Weihao Yu","Rongyao Fang","Zhaohai Li","Zhixun Li","Tiancheng Gu","Xilin Wei"],"authorCount":15,"org":"Qwen","affiliations":[],"orgNote":null,"date":"2026-10-14","hfDaily":"2026-10-20","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.23986","pdf":"https://arxiv.org/pdf/2610.23986","hf":"https://huggingface.co/papers/2610.23986","code":"https://github.com/QwenLM/BlenderWorld","project":"https://blenderworld.github.io/"},"metrics":{"hfUpvotes":25,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":185},"hot":51.29,"tags":["benchmark","agent","multimodal"],"summary":{"zh":"Qwen 的 BlenderWorld 基准测视觉智能体能否照真实参考图在 Blender 中创建和编辑 3D 内容。","en":"Qwen's BlenderWorld benchmarks visual agents creating and editing 3D content in Blender from real-world references."},"disciplines":["materials","robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":224},{"id":"arxiv-2610.01496","arxiv":"2610.01496","title":"SALD: Self-Referenced Advantage Learning for Diffusion Models","authors":["Aryan Das","Surjo Dey","Koushik Biswas","Swalpa Kumar Roy","Moloud Abdar","Arnab Bhattacharya","Vinay Kumar Verma"],"authorCount":7,"org":"Indian Institute of Technology, Kanpur","affiliations":[],"orgNote":null,"date":"2026-10-01","hfDaily":"2026-10-14","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.01496","pdf":"https://arxiv.org/pdf/2610.01496","hf":"https://huggingface.co/papers/2610.01496","code":null,"project":null},"metrics":{"hfUpvotes":107,"hfComments":2,"citations":1,"citationsSource":"semantic_scholar","githubStars":null},"hot":51.24,"tags":["visual-gen"],"summary":{"zh":"SALD 让扩散模型在两个噪声等级下自我对照打分，不需要教师模型就能获得训练信号。","en":"SALD lets a diffusion model score each image-caption pair at two noise levels, yielding self-referenced training signals without a teacher."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":225},{"id":"arxiv-2610.21306","arxiv":"2610.21306","title":"VLA-$Δ$: Learning Physical State Transition for Vision-Language-Action Models","authors":["Qi Lu","Jiarun Li","Ce Zhang","Miao Pan","Yi Pan","Xu Wang","Wenqi Zhang","Xuhong Zhang"],"authorCount":8,"org":"ZJU-OmniAI","affiliations":[],"orgNote":null,"date":"2026-10-15","hfDaily":"2026-10-16","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.21306","pdf":"https://arxiv.org/pdf/2610.21306","hf":"https://huggingface.co/papers/2610.21306","code":"https://github.com/ZJU-OmniAI/VLA-Delta","project":"https://zju-omniai.github.io/VLA-Delta/"},"metrics":{"hfUpvotes":64,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":22},"hot":51.23,"tags":["robotics"],"summary":{"zh":"VLA-Δ 显式监督物理状态的变化，减少视觉-语言-动作模型对“观察→动作”对应关系的过拟合。","en":"VLA-Δ explicitly supervises physical state transitions to reduce VLA overfitting to observation-action pairs."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":226},{"id":"arxiv-2609.38748","arxiv":"2609.38748","title":"Hear the World in Stereo: Learning Dynamic Spatial Correspondence for Immersive Joint Video-Audio Generation","authors":["Hanmo Chen","Chengcheng Liu","Tianxiao Chen","Zheyu Zhang","Siming Zheng","Jinwei Chen","Xu Yang","Cheng Deng"],"authorCount":10,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-09-30","hfDaily":"2026-10-12","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2609.38748","pdf":"https://arxiv.org/pdf/2609.38748","hf":"https://huggingface.co/papers/2609.38748","code":"https://github.com/vivoCameraResearch/StereoBind","project":"https://vivocameraresearch.github.io/Stereo-Bind-Project/"},"metrics":{"hfUpvotes":68,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":14},"hot":51.19,"tags":["multimodal","visual-gen"],"summary":{"zh":"StereoBind 让生成视频中的立体声声源位置随画面物体运动而变化，用于沉浸式音视频生成。","en":"StereoBind ties stereo sound position to on-screen source motion for immersive joint video-audio generation."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":227},{"id":"arxiv-2609.39378","arxiv":"2609.39378","title":"EgoTools: Towards Tool-Centric Reasoning in Real-World Egocentric Videos","authors":["Shulin Tian","Junsu Kim","Shuai Liu","Hao Li","Yujiao Shen","Sihan Li","Zhe Yang","Yeongon Kim"],"authorCount":20,"org":"Ropedia","affiliations":[],"orgNote":null,"date":"2026-09-30","hfDaily":"2026-10-02","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2609.39378","pdf":"https://arxiv.org/pdf/2609.39378","hf":"https://huggingface.co/papers/2609.39378","code":"https://github.com/Ropedia/EgoTools","project":"https://ropedia.github.io/egotools/"},"metrics":{"hfUpvotes":75,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":11},"hot":51.17,"tags":["multimodal","reasoning","benchmark"],"summary":{"zh":"EgoTools 关注真实第一视角视频中的工具使用推理，评测并提升模型对手-工具-物体关系的理解。","en":"EgoTools targets tool-centric reasoning in real egocentric videos, covering hand-tool-object relations and procedures."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":228},{"id":"arxiv-2609.32993","arxiv":"2609.32993","title":"X-Tree: Tokenizing Reusable Experience for Efficient Agent Generalization","authors":["Sitao Cheng","Xunjian Yin","Zhiyuan Sun","Yuxuan Li","Ruiwen Zhou","Xiangru Jian","Victor Zhong"],"authorCount":7,"org":"University of Waterloo","affiliations":[],"orgNote":null,"date":"2026-09-26","hfDaily":"2026-10-02","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2609.32993","pdf":"https://arxiv.org/pdf/2609.32993","hf":"https://huggingface.co/papers/2609.32993","code":"https://github.com/sitaocheng/X-Tree","project":"https://sitaocheng.github.io/xtree/"},"metrics":{"hfUpvotes":80,"hfComments":2,"citations":1,"citationsSource":"semantic_scholar","githubStars":2},"hot":51.11,"tags":["agent","efficiency"],"summary":{"zh":"X-Tree 像分词器一样从轨迹中提取可复用的子流程，把分层经验直接训练进智能体权重。","en":"X-Tree tokenises recurring sub-procedures in agent trajectories and trains on that hierarchy, improving generalisation without LLM calls."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":229},{"id":"arxiv-2610.13630","arxiv":"2610.13630","title":"Multi-Stream Attention Residuals","authors":["Xiaohan Qin","Xiangdong Zhang","Tuo Dai","Sunan Zou","Huaijin Wu","Yebin Yang","Zhuo Xia","Lin Yao"],"authorCount":10,"org":"Shanghai Jiao Tong University","affiliations":[],"orgNote":null,"date":"2026-10-09","hfDaily":"2026-10-13","primaryCategory":"cs.CL","categories":["cs.CL","cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.13630","pdf":"https://arxiv.org/pdf/2610.13630","hf":"https://huggingface.co/papers/2610.13630","code":"https://github.com/jianke0604/MSAR","project":null},"metrics":{"hfUpvotes":50,"hfComments":5,"citations":0,"citationsSource":"semantic_scholar","githubStars":37},"hot":51.1,"tags":["llm","efficiency"],"summary":{"zh":"多流注意力残差把“沿网络深度的注意力”扩展为多头形式，以很小开销带来明显提升。","en":"Multi-Stream Attention Residuals extend depth-wise attention residuals to multiple heads, with sizeable gains at little extra cost."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":230},{"id":"arxiv-2610.17552","arxiv":"2610.17552","title":"FlashVSR+: Efficient Video Super-Resolution via Hybrid Attention and Reward-Yield Distribution Matching","authors":["Ruchang Yao","Shi Guo","Junhao Zhuang","Mingde Yao","Xin Cai","Jiahao Zhan","Bin Fu","Xinyuan Chen"],"authorCount":10,"org":"MMLab-CUHK","affiliations":[],"orgNote":null,"date":"2026-10-13","hfDaily":"2026-10-14","primaryCategory":"cs.CV","categories":["cs.CV","cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.17552","pdf":"https://arxiv.org/pdf/2610.17552","hf":"https://huggingface.co/papers/2610.17552","code":"https://github.com/OpenImagingLab/FlashVSR_Plus","project":"https://www.flashvsr.tech"},"metrics":{"hfUpvotes":41,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":123},"hot":51.08,"tags":["visual-gen","efficiency"],"summary":{"zh":"FlashVSR+ 用混合注意力和偏好对齐的分布匹配，让一步式视频超分辨率更快、更符合人眼偏好。","en":"FlashVSR+ speeds up one-step video super-resolution with hybrid attention and aligns it to human preference via reward-weighted distribution matching."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":231},{"id":"arxiv-2610.22131","arxiv":"2610.22131","title":"Intern-Decision: Efficient Multimodal Decision Making with Autoregressive Masked Language Modeling","authors":["Tong Zhu","Zhe Xu","Zefeng He","Yunkai Chai","Xuyang Hu","Dehong Zheng","Yun Luo","Yafu Li"],"authorCount":10,"org":"Intern Large Models","affiliations":[],"orgNote":null,"date":"2026-10-15","hfDaily":"2026-10-16","primaryCategory":"cs.CL","categories":["cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.22131","pdf":"https://arxiv.org/pdf/2610.22131","hf":"https://huggingface.co/papers/2610.22131","code":"https://github.com/InternLM/Intern-Decision","project":null},"metrics":{"hfUpvotes":25,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":826},"hot":51.02,"tags":["multimodal","agent","efficiency"],"summary":{"zh":"Intern-Decision 系列小模型直接输出结构化选项及其概率，替代逐字生成，让智能体决策更快。","en":"Intern-Decision (0.8B–4B) returns structured choices with probabilities instead of free text, making agent decisions faster."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":232},{"id":"arxiv-2609.32019","arxiv":"2609.32019","title":"Decentralized Master-Mind: Joint Action Refinement through Iterative Intent Denoising in Multi-Agent Pathfinding","authors":["Valeriy Vyaltsev","Anton Andreychuk","Taisia Zlotnikova","Konstantin Yakovlev","Aleksandr Panov","Alexey Skrynnik"],"authorCount":6,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-09-25","hfDaily":"2026-10-02","primaryCategory":"cs.AI","categories":["cs.AI","cs.MA"],"links":{"abs":"https://arxiv.org/abs/2609.32019","pdf":"https://arxiv.org/pdf/2609.32019","hf":"https://huggingface.co/papers/2609.32019","code":"https://github.com/CognitiveAISystems/DMM","project":null},"metrics":{"hfUpvotes":66,"hfComments":2,"citations":1,"citationsSource":"semantic_scholar","githubStars":5},"hot":50.95,"tags":["robotics","rl"],"summary":{"zh":"DMM 用迭代去噪联合细化多智能体路径规划中的动作，避免各自采样拼出相互冲突的方案。","en":"DMM refines joint actions in decentralised multi-agent pathfinding by iterative intent denoising, avoiding incompatible combinations."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":233},{"id":"arxiv-2610.10409","arxiv":"2610.10409","title":"RobotWorld: Benchmarking Multimodal Agents for Robot Use Across Diverse Tasks and Embodiments","authors":["Zhiqin Yang","Chenxin Li","Xiaomeng Hu","Yibin Liu","Weidong Huang","Jiankai Sun","Haitao Li","Zijian Wu"],"authorCount":33,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-10-07","hfDaily":"2026-10-08","primaryCategory":"cs.RO","categories":["cs.RO","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2610.10409","pdf":"https://arxiv.org/pdf/2610.10409","hf":"https://huggingface.co/papers/2610.10409","code":"https://github.com/robotworldai/robotworld","project":"https://robotworldai.github.io/"},"metrics":{"hfUpvotes":40,"hfComments":2,"citations":1,"citationsSource":"semantic_scholar","githubStars":41},"hot":50.9,"tags":["robotics","agent","benchmark"],"summary":{"zh":"RobotWorld 用 84 个仿真任务（操作、移动、驾驶、飞行）检验通用多模态 Agent 能否操控机器人。","en":"RobotWorld uses 84 simulated tasks (manipulation, locomotion, driving, aerial) to test multimodal agents at robot use."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":234},{"id":"arxiv-2610.07767","arxiv":"2610.07767","title":"TRACE: Rollout-Guided Quantization-Aware Training for FP4 Reinforcement Learning of MoE Language Models","authors":["Xin Wang","Hao Yu","Zhengyang Zhuge","Bochao Mao","Zheng Li","Junda Feng","Yuyan Luo","Yi Zhang"],"authorCount":12,"org":"Qwen","affiliations":[],"orgNote":null,"date":"2026-10-06","hfDaily":"2026-10-07","primaryCategory":"cs.LG","categories":["cs.LG","cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.07767","pdf":"https://arxiv.org/pdf/2610.07767","hf":"https://huggingface.co/papers/2610.07767","code":null,"project":null},"metrics":{"hfUpvotes":91,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":50.87,"tags":["efficiency","rl","llm"],"summary":{"zh":"Qwen 团队的 TRACE 对齐训练与 rollout 两条 FP4 量化路径，用于 MoE 大模型的低精度强化学习。","en":"Qwen's TRACE aligns the training and rollout FP4 quantisation paths for low-precision RL of MoE language models."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":235},{"id":"arxiv-2610.08963","arxiv":"2610.08963","title":"On KL-Regularized Policy Optimization","authors":["Yifan Zhang"],"authorCount":1,"org":"math-ai","affiliations":[],"orgNote":null,"date":"2026-10-06","hfDaily":"2026-10-08","primaryCategory":"cs.LG","categories":["cs.LG","cs.AI","cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.08963","pdf":"https://arxiv.org/pdf/2610.08963","hf":"https://huggingface.co/papers/2610.08963","code":"https://github.com/yifanzhang-pro/KLPO","project":"https://github.com/yifanzhang-pro/KLPO"},"metrics":{"hfUpvotes":27,"hfComments":2,"citations":1,"citationsSource":"semantic_scholar","githubStars":187},"hot":50.83,"tags":["rl","llm"],"summary":{"zh":"KLPO 把 KL 正则锚定在采样策略上，处理异步强化学习中策略过期和推理/训练概率不一致的问题。","en":"KLPO anchors the KL regulariser at the sampling policy to handle stale, mismatched rollouts in asynchronous LLM-agent RL."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":236},{"id":"arxiv-2609.30199","arxiv":"2609.30199","title":"ExplorationBench: Measuring AI Systems' Exploration in Verifiable Alien Worlds","authors":["Ming Zhang","Zhenghao Xiang","Peizhong Gao","Yujiong Shen","Yuhui Wang","Zhonghan Yue","Shihan Dou","Zhangyue Yin"],"authorCount":20,"org":"Tencent Hunyuan","affiliations":[],"orgNote":null,"date":"2026-09-24","hfDaily":"2026-09-25","primaryCategory":"cs.AI","categories":["cs.AI","cs.CL"],"links":{"abs":"https://arxiv.org/abs/2609.30199","pdf":"https://arxiv.org/pdf/2609.30199","hf":"https://huggingface.co/papers/2609.30199","code":"https://github.com/Tencent-Hunyuan/ExplorationBench","project":"https://www.explorationbench.com/"},"metrics":{"hfUpvotes":31,"hfComments":1,"citations":1,"citationsSource":"semantic_scholar","githubStars":26},"hot":50.71,"tags":["ai4science","benchmark","agent"],"summary":{"zh":"腾讯混元 ExplorationBench 在可验证的虚构世界中测 AI 的科学探索能力，排除预训练知识的干扰。","en":"Tencent Hunyuan's ExplorationBench measures scientific exploration in verifiable alien worlds, ruling out recall from pre-training."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":237},{"id":"arxiv-2610.07967","arxiv":"2610.07967","title":"DecepEval: A Benchmark for Evaluating Deception in LLM Agents","authors":["Yiming Xu","Hongyue Yu","Beihua Yang","Zihan Chen","Yixin Liu","Zhen Peng","Bin Shi","Bo Dong"],"authorCount":11,"org":"Xi'an Jiaotong University","affiliations":[],"orgNote":null,"date":"2026-10-06","hfDaily":"2026-10-08","primaryCategory":"cs.LG","categories":["cs.LG"],"links":{"abs":"https://arxiv.org/abs/2610.07967","pdf":"https://arxiv.org/pdf/2610.07967","hf":"https://huggingface.co/papers/2610.07967","code":"https://github.com/functy/DECEPEVAL","project":null},"metrics":{"hfUpvotes":85,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":5},"hot":50.67,"tags":["safety","benchmark","agent"],"summary":{"zh":"DecepEval 用 1532 个实例和“欺骗钻石”框架，评测 LLM Agent 在什么条件下更容易撒谎。","en":"DecepEval uses 1,532 instances and a 'deception diamond' framework to test when LLM agents are likely to deceive."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":238},{"id":"arxiv-2610.18067","arxiv":"2610.18067","title":"WorkRover: Scaling Long-Horizon Cowork Agent Training Via Recursive Environment Evolution","authors":["Hang He","Linting Wang","Guanming Liu","Bo Mao","Jinxiu Liu","Yiran Suo","Maosen Zhou","Zhiheng Xi"],"authorCount":17,"org":"Nex AGI","affiliations":[],"orgNote":null,"date":"2026-10-13","hfDaily":"2026-10-14","primaryCategory":"cs.AI","categories":["cs.AI","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2610.18067","pdf":"https://arxiv.org/pdf/2610.18067","hf":"https://huggingface.co/papers/2610.18067","code":"https://github.com/workrover/WorkRover","project":"https://workrover.github.io/"},"metrics":{"hfUpvotes":85,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":4},"hot":50.65,"tags":["agent","rl"],"summary":{"zh":"WorkRover 通过递归演化训练环境与验证器，扩展长程“协作办公”智能体的强化学习训练。","en":"WorkRover scales RL for long-horizon cowork agents by recursively evolving tasks, work contexts and verifiers."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":239},{"id":"arxiv-2610.22194","arxiv":"2610.22194","title":"AutoSFT: An Agentic System for Automated Data Research in Supervised Fine-Tuning","authors":["Bingxiang He","Pengqi Sun","Shihao Peng","Shaoxuan Li","Huiyu Xiao","Hongjian Shi","Zixuan Fu","Ziqing Qiao"],"authorCount":21,"org":"Tsinghua NLP Group","affiliations":[],"orgNote":null,"date":"2026-10-15","hfDaily":"2026-10-16","primaryCategory":"cs.AI","categories":["cs.AI","cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.22194","pdf":"https://arxiv.org/pdf/2610.22194","hf":"https://huggingface.co/papers/2610.22194","code":"https://github.com/Thinking-Space/AutoSFT","project":null},"metrics":{"hfUpvotes":41,"hfComments":4,"citations":0,"citationsSource":"semantic_scholar","githubStars":14},"hot":50.56,"tags":["agent","llm"],"summary":{"zh":"清华 AutoSFT 给智能体一整套数据操作（配比、剪枝、修复、合成），自动做监督微调的数据研究。","en":"Tsinghua's AutoSFT gives an agent a full data action space—mix, prune, repair, synthesise—to automate SFT data research."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":240},{"id":"arxiv-2610.03391","arxiv":"2610.03391","title":"Native Action-Prior Learning from Videos for World Action Models","authors":["Zhaochong An","Fei Zhang","Menglin Jia","Duncan Frost","Zijian Zhou","Yikai Wang","Xudong Wang","Aditya Patel"],"authorCount":13,"org":"Meta","affiliations":[],"orgNote":null,"date":"2026-10-02","hfDaily":"2026-10-05","primaryCategory":"cs.CV","categories":["cs.CV","cs.RO"],"links":{"abs":"https://arxiv.org/abs/2610.03391","pdf":"https://arxiv.org/pdf/2610.03391","hf":"https://huggingface.co/papers/2610.03391","code":null,"project":"https://zhaochongan.github.io/projects/NAVA-WAM/"},"metrics":{"hfUpvotes":88,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":50.53,"tags":["robotics"],"summary":{"zh":"NAVA-WAM 直接用无动作标注的视频预训练动作策略，降低世界-动作模型对机器人标注数据的依赖。","en":"NAVA-WAM pretrains the action policy directly from observation-only videos, reducing reliance on action-labelled robot data."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":241},{"id":"arxiv-2610.26824","arxiv":"2610.26824","title":"SpecMRL: Redesign Matryoshka Representation Learning within Parametric Spectrum","authors":["Zhongxin Chen","Songhao Wu","Heng Cui","Cong Li","Xiao Zhou","Rui Yan"],"authorCount":6,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-10-19","hfDaily":"2026-10-20","primaryCategory":"cs.CL","categories":["cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.26824","pdf":"https://arxiv.org/pdf/2610.26824","hf":"https://huggingface.co/papers/2610.26824","code":"https://github.com/CentreChen/Spec-MRL","project":null},"metrics":{"hfUpvotes":80,"hfComments":2,"citations":null,"citationsSource":null,"githubStars":6},"hot":50.46,"tags":["efficiency"],"summary":{"zh":"SpecMRL 用参数导出的谱基重排特征，让俄罗斯套娃式表示的前缀维度更好地承载关键信息。","en":"SpecMRL reorganises features on a parameter-derived spectral basis so Matryoshka prefixes carry the most important information."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":242},{"id":"arxiv-2609.38445","arxiv":"2609.38445","title":"AIM: Agentic Idea Management for Automated Research","authors":["Hyeong Kyu Choi","Bhavana Dalvi Mishra","Jiefeng Chen","Mihir Parmar","Rui Meng","Chun-Liang Li","Xiangru Tang","Sharon Li"],"authorCount":10,"org":"Google","affiliations":[],"orgNote":null,"date":"2026-09-29","hfDaily":"2026-10-01","primaryCategory":"cs.AI","categories":["cs.AI","cs.CE"],"links":{"abs":"https://arxiv.org/abs/2609.38445","pdf":"https://arxiv.org/pdf/2609.38445","hf":"https://huggingface.co/papers/2609.38445","code":null,"project":"https://imhgchoi.github.io/agentic-idea-manager/"},"metrics":{"hfUpvotes":53,"hfComments":3,"citations":2,"citationsSource":"semantic_scholar","githubStars":null},"hot":50.36,"tags":["agent","ai4science"],"summary":{"zh":"谷歌 AIM 借鉴贝叶斯优化，自动管理、筛选和推进自动化科研中的研究想法。","en":"Google's AIM manages, selects and explores research ideas for automated research, inspired by Bayesian optimisation."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":243},{"id":"arxiv-2610.22264","arxiv":"2610.22264","title":"SCOPE: Semantic Bounded Credit Optimization for Prompt Engineering in Text-to-image Generation","authors":["Yufei Jiao","Jinfa Huang","Zeyue Xue","Tianyu Xie","Yuexiao Ma","Rongrong Ji","Xiawu Zheng","Jiebo Luo"],"authorCount":8,"org":"University of Rochester","affiliations":[],"orgNote":null,"date":"2026-10-15","hfDaily":"2026-10-16","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.22264","pdf":"https://arxiv.org/pdf/2610.22264","hf":"https://huggingface.co/papers/2610.22264","code":"https://github.com/MAC-AutoML/SCOPE","project":null},"metrics":{"hfUpvotes":86,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":4},"hot":50.33,"tags":["rl","visual-gen"],"summary":{"zh":"SCOPE 基于 GRPO，按语义片段而非单个子词分配奖励，训练文生图提示词改写模型。","en":"SCOPE, a GRPO-based objective, credits semantic spans rather than sub-word tokens when training prompt rewriters for text-to-image."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":244},{"id":"arxiv-2610.05030","arxiv":"2610.05030","title":"EVISKILL: Grounding Skill Evolution in Replayable Evidence","authors":["Yan Zhou","Yili Wang","Yiwei Dai","Qinggang Zhang","Xin Wang"],"authorCount":5,"org":"DEEP Group at Jilin University","affiliations":[],"orgNote":null,"date":"2026-10-04","hfDaily":"2026-10-07","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.05030","pdf":"https://arxiv.org/pdf/2610.05030","hf":"https://huggingface.co/papers/2610.05030","code":"https://github.com/Zhouyaner/Eviskill","project":"https://github.com/Zhouyaner/Eviskill"},"metrics":{"hfUpvotes":55,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":28},"hot":50.32,"tags":["agent"],"summary":{"zh":"EVISKILL 把智能体技能演化建立在可回放的行为证据上，判断每次修改是否该保留。","en":"EVISKILL grounds agent skill evolution in replayable evidence to judge which edits should persist."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":245},{"id":"arxiv-2610.00906","arxiv":"2610.00906","title":"ActiveSaddler: Automated Curriculum Learning for Agent Harness Optimization","authors":["Sungho Park","Wonjoong Kim","Jue Zhang","Wook-Shin Han","Pengfei Gao","Chanyoung Park","Yongqiang Yao","Rao Fu"],"authorCount":11,"org":"Microsoft","affiliations":[],"orgNote":null,"date":"2026-10-01","hfDaily":"2026-10-02","primaryCategory":"cs.AI","categories":["cs.AI","cs.CL","cs.LG","cs.MA","cs.SE"],"links":{"abs":"https://arxiv.org/abs/2610.00906","pdf":"https://arxiv.org/pdf/2610.00906","hf":"https://huggingface.co/papers/2610.00906","code":null,"project":"https://autosaddler-projectpage.github.io/activesaddler/"},"metrics":{"hfUpvotes":86,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":50.31,"tags":["agent"],"summary":{"zh":"微软 ActiveSaddler 让训练场景随智能体外壳一起自适应调整，实现自动课程式的外壳优化。","en":"Microsoft's ActiveSaddler adapts the training curriculum alongside the evolving agent harness for automated harness optimisation."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":246},{"id":"arxiv-2609.38923","arxiv":"2609.38923","title":"GraphForge: Training Working Agents with Graph-Anchored Workspace Synthesis","authors":["Qisheng Su","Hanchen Wang","Guanru Zhu","Huicheng Jiang","Qiuyinzhe Zhang","Kou Shi","Zhen Fang","Ziao Zhang"],"authorCount":13,"org":"University of Science and Technology of China","affiliations":[],"orgNote":null,"date":"2026-09-30","hfDaily":"2026-10-02","primaryCategory":"cs.CL","categories":["cs.CL"],"links":{"abs":"https://arxiv.org/abs/2609.38923","pdf":"https://arxiv.org/pdf/2609.38923","hf":"https://huggingface.co/papers/2609.38923","code":null,"project":null},"metrics":{"hfUpvotes":123,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":50.28,"tags":["agent","benchmark"],"summary":{"zh":"GraphForge 以真实文件构建证据图，合成带可验证结果的办公任务来训练“干活型”智能体。","en":"GraphForge synthesises verifiable workspace tasks grounded in real files via evidence graphs to train working agents."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":247},{"id":"arxiv-2610.23279","arxiv":"2610.23279","title":"Xiaomi-Lens-0 Technical Report","authors":["Pei Fu","Yike Liu","Yue Sun","Hanmo Wang","Xiuwen Xi","Ruoceng Zhang","Shaojie Zhang","Tingwei Huang"],"authorCount":16,"org":"SeerRay-Lab","affiliations":[],"orgNote":null,"date":"2026-10-16","hfDaily":"2026-10-19","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.23279","pdf":"https://arxiv.org/pdf/2610.23279","hf":"https://huggingface.co/papers/2610.23279","code":"https://github.com/SeerRay-Lab/Xiaomi-Lens-0","project":"https://seerray-lab.github.io/Xiaomi-Lens-0"},"metrics":{"hfUpvotes":45,"hfComments":4,"citations":0,"citationsSource":"semantic_scholar","githubStars":45},"hot":50.27,"tags":["vision","multimodal"],"summary":{"zh":"Xiaomi-Lens-0 是基于 Qwen3.5 的 4B 多模态模型，能直接从自身视觉表征预测分割掩码。","en":"Xiaomi-Lens-0 is a 4B MLLM on Qwen3.5 that predicts segmentation masks from its own visual representations."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":248},{"id":"arxiv-2610.02304","arxiv":"2610.02304","title":"SimuVerity: Benchmarking Agents for Engineering-Grade Simulink Model Generation","authors":["Ruiqi Zhang","Jiahao Wang","Mingxuan Li","Haichen Luo","Chaoting Wang","Guoyu Mou","Keyu Lai","Hanchao Lv"],"authorCount":12,"org":"Xi'an Jiaotong University","affiliations":[],"orgNote":null,"date":"2026-10-01","hfDaily":"2026-10-05","primaryCategory":"cs.SE","categories":["cs.SE","cs.AI","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2610.02304","pdf":"https://arxiv.org/pdf/2610.02304","hf":"https://huggingface.co/papers/2610.02304","code":"https://github.com/SimuVerity/SimuVerity","project":"https://simuverity.github.io/SimuVerity/"},"metrics":{"hfUpvotes":56,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":21},"hot":50.24,"tags":["agent","benchmark","code"],"summary":{"zh":"SimuVerity 用 101 个任务检验智能体生成的 Simulink 模型是否满足工程需求，而不只是能运行。","en":"SimuVerity's 101 tasks check whether agent-generated Simulink models meet engineering requirements, not just whether they run."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":249},{"id":"arxiv-2610.12984","arxiv":"2610.12984","title":"Work2Skill: Capturing Human GUI Experience for Reusable Agent Skill Induction","authors":["Yuqi Huang","Hui Li","Yifei Cao","Zikun Cai","Tianci Xie","Yangfan Gao","Pinxue Guo","Chaoyun Zhang"],"authorCount":12,"org":"Nex AGI","affiliations":[],"orgNote":null,"date":"2026-10-09","hfDaily":"2026-10-12","primaryCategory":"cs.AI","categories":["cs.AI","cs.HC"],"links":{"abs":"https://arxiv.org/abs/2610.12984","pdf":"https://arxiv.org/pdf/2610.12984","hf":"https://huggingface.co/papers/2610.12984","code":"https://github.com/china-qijizhifeng/Work2Skill","project":"https://china-qijizhifeng.github.io/Work2Skill/"},"metrics":{"hfUpvotes":52,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":34},"hot":50.24,"tags":["agent","multimodal"],"summary":{"zh":"Work2Skill 记录人类在图形界面上的操作经验，归纳成可复用的电脑使用智能体技能。","en":"Work2Skill captures human GUI demonstrations and induces reusable skills for computer-use agents."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":250},{"id":"arxiv-2610.02521","arxiv":"2610.02521","title":"Spatial Memory Intelligence: Endowing World Models with Understanding-Driven Long-Term Memory","authors":["Ying Yang","Guiyu Zhang","Lianghua Huang","Chang Nie","Chenyang Si","Haofan Wang","Shaoshuai Shi","Li Jiang"],"authorCount":8,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-10-01","hfDaily":"2026-10-05","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.02521","pdf":"https://arxiv.org/pdf/2610.02521","hf":"https://huggingface.co/papers/2610.02521","code":"https://github.com/xbyym/SMI","project":"https://spatial-memory-intelligence.github.io/"},"metrics":{"hfUpvotes":60,"hfComments":1,"citations":0,"citationsSource":"semantic_scholar","githubStars":24},"hot":50.2,"tags":["visual-gen","multimodal"],"summary":{"zh":"用多模态大模型的空间理解来管理长期记忆，让世界模型在长视频生成中保持空间一致。","en":"Uses MLLM spatial understanding to manage long-term memory so world models stay spatially consistent in long video generation."},"disciplines":["robotics","film-animation"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":251},{"id":"arxiv-2610.10539","arxiv":"2610.10539","title":"Tetris3D: 3D Scene Generation With Objects That Fit Together","authors":["Jaeyeong Kim","Jinhyuk Jang","Jongmin Lee","Kyehong Park","Seungryong Kim"],"authorCount":5,"org":"KAIST AI","affiliations":[],"orgNote":null,"date":"2026-10-07","hfDaily":"2026-10-08","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.10539","pdf":"https://arxiv.org/pdf/2610.10539","hf":"https://huggingface.co/papers/2610.10539","code":"https://github.com/cvlab-kaist/Tetris3D","project":"https://cvlab-kaist.github.io/Tetris3D/"},"metrics":{"hfUpvotes":49,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":41},"hot":50.11,"tags":["vision"],"summary":{"zh":"Tetris3D 在单图 3D 场景重建中按邻近物体的几何和物理关系生成每个物体，让物体彼此契合。","en":"Tetris3D conditions each object on its neighbours' geometry and physical relations so single-image 3D scenes fit together."},"disciplines":["physics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":252},{"id":"arxiv-2609.35767","arxiv":"2609.35767","title":"Learning Native Reflection in Unified Models with Interleaved Reinforcement Learning","authors":["Yijia Fan","Ziqi Huang","Zhongang Cai","Yan Li","Zimo Wen","Wanqi Yin","Haiwen Diao","Ziwei Liu"],"authorCount":8,"org":"MMLab@NTU","affiliations":[],"orgNote":null,"date":"2026-09-28","hfDaily":"2026-09-29","primaryCategory":"cs.CV","categories":["cs.CV","cs.AI"],"links":{"abs":"https://arxiv.org/abs/2609.35767","pdf":"https://arxiv.org/pdf/2609.35767","hf":"https://huggingface.co/papers/2609.35767","code":"https://github.com/waltstephen/UMM-Reflection","project":"https://waltstephen.github.io/UMM-Reflection/"},"metrics":{"hfUpvotes":50,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":31},"hot":50.06,"tags":["multimodal","rl","visual-gen"],"summary":{"zh":"用交错强化学习联合训练统一多模态模型的“反思文本”和图像生成，让它能自查并修图。","en":"Interleaved RL jointly trains reflection text and image generation so unified models can diagnose and repair their own images."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":253},{"id":"arxiv-2610.21767","arxiv":"2610.21767","title":"From Research Goals to Scientific Closure: Evaluating Agents Through Evidence-Supported Physics Research","authors":["Yijia Mao","Song Dai","Tao Chen","Xiangyu Lu","Linqian Zeng","Chenxi Hu","Sitong Zhou","Dongyang He"],"authorCount":17,"org":"Shanghai AI Laboratory","affiliations":[],"orgNote":null,"date":"2026-10-15","hfDaily":"2026-10-19","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.21767","pdf":"https://arxiv.org/pdf/2610.21767","hf":"https://huggingface.co/papers/2610.21767","code":"https://github.com/uncertainty2019/PhysClosureBench","project":"https://uncertainty2019.github.io/PhysClosureBench/"},"metrics":{"hfUpvotes":53,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":4},"hot":49.99,"tags":["ai4science","agent","benchmark"],"summary":{"zh":"PhysClosureBench 用 50 个专家验证的物理研究任务，评估 Agent 能否组织研究并得出有证据支撑的结论。","en":"PhysClosureBench uses 50 expert-verified physics tasks to test whether agents can run investigations to well-supported conclusions."},"disciplines":["physics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":254},{"id":"arxiv-2609.37533","arxiv":"2609.37533","title":"E-MoE: Enhanced Mixture-of-Experts for Non-Factorized Diffusion Language Models","authors":["Arseny Ivanov","Alexander Kolesov","Alexander Korotin","Ivan Oseledets","Mikhail Goncharov"],"authorCount":5,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-09-29","hfDaily":"2026-10-02","primaryCategory":"cs.CL","categories":["cs.CL"],"links":{"abs":"https://arxiv.org/abs/2609.37533","pdf":"https://arxiv.org/pdf/2609.37533","hf":"https://huggingface.co/papers/2609.37533","code":"https://github.com/Arseny5/e-moe","project":null},"metrics":{"hfUpvotes":68,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":10},"hot":49.98,"tags":["llm","efficiency"],"summary":{"zh":"E-MoE 用增强的混合专家改进非分解式扩散语言模型，提升少步解码质量并避免后验坍塌。","en":"E-MoE improves non-factorised diffusion language models with enhanced mixture-of-experts, raising few-step quality without posterior collapse."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":255},{"id":"arxiv-2609.33803","arxiv":"2609.33803","title":"Diffusion Reward Models","authors":["Xiangyang Wang","Bingxiang He","Zeyuan Liu","Jiaze Wang","Ziqing Qiao","Yuxin Zuo","Huan-ang Gao","Cheng Qian"],"authorCount":16,"org":"Tsinghua NLP Group","affiliations":[],"orgNote":null,"date":"2026-09-27","hfDaily":"2026-09-29","primaryCategory":"cs.LG","categories":["cs.LG","cs.AI"],"links":{"abs":"https://arxiv.org/abs/2609.33803","pdf":"https://arxiv.org/pdf/2609.33803","hf":"https://huggingface.co/papers/2609.33803","code":"https://github.com/thunlp/DRM","project":null},"metrics":{"hfUpvotes":35,"hfComments":1,"citations":0,"citationsSource":"semantic_scholar","githubStars":37},"hot":49.97,"tags":["alignment","rl"],"summary":{"zh":"清华 DRM 把奖励建模改成条件密度估计，用扩散模型刻画人类偏好的多峰分布。","en":"Tsinghua's Diffusion Reward Model recasts reward modelling as conditional density estimation to capture multimodal human preferences."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":256},{"id":"arxiv-2609.34715","arxiv":"2609.34715","title":"PDE-JEPA: Predictive Representation Learning of Latent Dynamics Modeling for Parametric PDEs","authors":["Zhentao Tan","Jianrong Zhang","Ruijie Quan","Yi Yang"],"authorCount":4,"org":"Zhejiang University","affiliations":[],"orgNote":null,"date":"2026-09-28","hfDaily":"2026-10-05","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2609.34715","pdf":"https://arxiv.org/pdf/2609.34715","hf":"https://huggingface.co/papers/2609.34715","code":"https://github.com/Tanpig-X/PDE-JEPA","project":"https://tanpig-x.github.io/PDE-JEPA/"},"metrics":{"hfUpvotes":65,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":12},"hot":49.96,"tags":["ai4science","efficiency"],"summary":{"zh":"PDE-JEPA 用预测式表示预训练学习参数化偏微分方程的潜在动力学，替代重建式目标。","en":"PDE-JEPA learns latent dynamics of parametric PDEs with predictive (JEPA-style) pretraining instead of reconstruction."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":257},{"id":"arxiv-2610.02160","arxiv":"2610.02160","title":"4Director: Controlling Video World Models with Rigid 3D Geometry","authors":["Wei Cao","Hao Zhang","Vikram Voleti","Yuqun Wu","Mallikarjun B R","Shimon Vainer","Mark Boss","Yaoyao Liu"],"authorCount":8,"org":"Stability AI","affiliations":[],"orgNote":null,"date":"2026-10-01","hfDaily":"2026-10-02","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.02160","pdf":"https://arxiv.org/pdf/2610.02160","hf":"https://huggingface.co/papers/2610.02160","code":"https://github.com/VVeiCao/4Director","project":"https://stability-ai.github.io/4director/"},"metrics":{"hfUpvotes":39,"hfComments":2,"citations":1,"citationsSource":"semantic_scholar","githubStars":28},"hot":49.73,"tags":["visual-gen"],"summary":{"zh":"Stability AI 的 4Director 用刚体网格加逐帧变换的 4D 场景，精确控制视频世界模型的镜头与物体运动。","en":"Stability AI's 4Director controls camera and object motion in a video world model via explicit rigid meshes and per-frame transforms."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":258},{"id":"arxiv-2610.17771","arxiv":"2610.17771","title":"Many Right Answers, One Training Example: Removing Responder Noise from GRPO for LLM Recommenders","authors":["Xin Li","Hao Jiang","Jinghao Guo","Weisi Lin","Chau Yuen"],"authorCount":5,"org":"Nanyang Technological University","affiliations":[],"orgNote":null,"date":"2026-10-13","hfDaily":"2026-10-15","primaryCategory":"cs.CL","categories":["cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.17771","pdf":"https://arxiv.org/pdf/2610.17771","hf":"https://huggingface.co/papers/2610.17771","code":"https://github.com/LiXin97/ReMaRk","project":"https://lixin.ai/ReMaRk/"},"metrics":{"hfUpvotes":79,"hfComments":4,"citations":0,"citationsSource":"semantic_scholar","githubStars":3},"hot":49.7,"tags":["rl","llm"],"summary":{"zh":"ReMaRk 让大模型推荐器的 GRPO 训练对所有回复的并集打分，消除随机抽一个回复带来的噪声。","en":"ReMaRk scores recommendation lists against the union of all replies, removing responder noise from GRPO training of LLM recommenders."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":259},{"id":"arxiv-2610.09455","arxiv":"2610.09455","title":"RLHND: Video Foundation Models as Physically Grounded Hand Trackers for Robot Learning","authors":["Seungjun Moon","Subin Jeon","Sangwoo Kim","Hanbyul Joo","Jinwoo Shin"],"authorCount":5,"org":"RLWRLD","affiliations":[],"orgNote":null,"date":"2026-10-07","hfDaily":"2026-10-07","primaryCategory":"cs.CV","categories":["cs.CV","cs.RO"],"links":{"abs":"https://arxiv.org/abs/2610.09455","pdf":"https://arxiv.org/pdf/2610.09455","hf":"https://huggingface.co/papers/2610.09455","code":"https://github.com/seungjun-moon/rlhnd","project":"https://seungjun-moon.github.io/rlhnd/"},"metrics":{"hfUpvotes":45,"hfComments":1,"citations":1,"citationsSource":"semantic_scholar","githubStars":19},"hot":49.59,"tags":["robotics","vision"],"summary":{"zh":"RLHND 基于视频基础模型从第一视角视频估计手部姿态和触觉信息，让人类视频更好地用于机器人学习。","en":"RLHND uses a video foundation model to estimate hand pose and touch from egocentric video, making human videos more useful for robots."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":260},{"id":"arxiv-2610.09450","arxiv":"2610.09450","title":"Iris-3B: Going Beyond the Latent with Pixel-Space Diffusion Training, Conversion and Fine-Tuning","authors":["Hanqiu Li Cai","Chema Garabito"],"authorCount":2,"org":"speridlabs","affiliations":["SperidLabs"],"orgNote":null,"date":"2026-10-07","hfDaily":"2026-10-08","primaryCategory":"cs.CV","categories":["cs.CV","cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.09450","pdf":"https://arxiv.org/pdf/2610.09450","hf":"https://huggingface.co/papers/2610.09450","code":"https://github.com/speridlabs/iris-3b","project":"https://github.com/speridlabs/iris-3b"},"metrics":{"hfUpvotes":31,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":194},"hot":49.49,"tags":["visual-gen"],"summary":{"zh":"Iris-3B 从头预训练并转换得到像素空间扩散模型，检验去掉 VAE 是否有利于细节敏感任务。","en":"Iris-3B pretrains and converts pixel-space diffusion models to test whether skipping the VAE helps detail-sensitive tasks."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":261},{"id":"arxiv-2610.16811","arxiv":"2610.16811","title":"X-Ragent: Tool-Augmented Reinforcement Learning for Evidence-Grounded Chest X-Ray Reasoning","authors":["Yi Wang","Zhenlong Yuan","Yuchan Guo","Maolei Huang","Zijian Song","Dapeng Zhang","Rongbin Tan","Jiyuan Wang"],"authorCount":12,"org":"Carnegie Mellon University","affiliations":[],"orgNote":null,"date":"2026-10-12","hfDaily":"2026-10-14","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.16811","pdf":"https://arxiv.org/pdf/2610.16811","hf":"https://huggingface.co/papers/2610.16811","code":null,"project":null},"metrics":{"hfUpvotes":79,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":49.47,"tags":["rl","reasoning","ai4science"],"summary":{"zh":"X-Ragent 用工具增强强化学习训练胸片推理，学会何时调用辅助证据、过滤不适用的量化证据。","en":"X-Ragent uses tool-augmented RL for chest X-ray reasoning, learning when to call auxiliary evidence and when to ignore it."},"disciplines":["clinical-medicine"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":262},{"id":"arxiv-2610.07625","arxiv":"2610.07625","title":"Stateless Language Agents: Scaling Long-Horizon Automated Research","authors":["Qizheng Zhang","Changxiu Ji","Isaac Sun","Yuetai Li","Shubhangi Upasani","Sherry Ruan","Boyuan Ma","Fenglu Hong"],"authorCount":18,"org":"Stanford University","affiliations":[],"orgNote":null,"date":"2026-10-06","hfDaily":"2026-10-12","primaryCategory":"cs.LG","categories":["cs.LG","cs.AI","cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.07625","pdf":"https://arxiv.org/pdf/2610.07625","hf":"https://huggingface.co/papers/2610.07625","code":"https://github.com/Alex-q-z/stateless-language-agents","project":null},"metrics":{"hfUpvotes":30,"hfComments":3,"citations":1,"citationsSource":"semantic_scholar","githubStars":11},"hot":49.4,"tags":["agent","ai4science"],"summary":{"zh":"斯坦福“无状态语言智能体”把研究状态放到外部，解决长程自动化科研中重复劳动和停滞的问题。","en":"Stanford's Stateless Language Agents keep research state outside the agent, tackling duplication and stalls in long-horizon automated research."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":263},{"id":"arxiv-2610.02298","arxiv":"2610.02298","title":"EditHero: A Benchmark for Long-Horizon Part-Level 3D Editing and Vibe Modeling","authors":["Ruihan Yu","Yu-Ju Tsai","Muyao Niu","Runyi Li","Lian Fu","Hanqing Liu","Zheng-Hui Huang","Yonghao Yu"],"authorCount":12,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-10-01","hfDaily":"2026-10-05","primaryCategory":"cs.CV","categories":["cs.CV","cs.AI","cs.GR"],"links":{"abs":"https://arxiv.org/abs/2610.02298","pdf":"https://arxiv.org/pdf/2610.02298","hf":"https://huggingface.co/papers/2610.02298","code":"https://github.com/AlayaLab/EditHero","project":"https://alaya-lab.github.io/EditHero/"},"metrics":{"hfUpvotes":60,"hfComments":0,"citations":0,"citationsSource":"semantic_scholar","githubStars":26},"hot":49.35,"tags":["vision","benchmark"],"summary":{"zh":"EditHero 是首个长程、部件级 3D 编辑基准，检验连续多次修改时能否只改指定部分。","en":"EditHero is a benchmark for long-horizon, part-level 3D editing, checking that each revision changes only what was asked."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":264},{"id":"arxiv-2610.12978","arxiv":"2610.12978","title":"Learning Representations to Guide Pixel Diffusion Transformers","authors":["Zihao Pan","Mengting Chen","Cheng Shi","Taihang Hu","Zhengrong Yue","Zikang Wang","Yuan Wang","Wei Li"],"authorCount":9,"org":"alibaba","affiliations":[],"orgNote":null,"date":"2026-10-09","hfDaily":"2026-10-12","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.12978","pdf":"https://arxiv.org/pdf/2610.12978","hf":"https://huggingface.co/papers/2610.12978","code":"https://github.com/Pan-Zihao/Representation-Guidance","project":"https://pan-zihao.github.io/Representation-Guidance"},"metrics":{"hfUpvotes":33,"hfComments":6,"citations":0,"citationsSource":"semantic_scholar","githubStars":16},"hot":49.27,"tags":["visual-gen"],"summary":{"zh":"研究什么样的视觉表示真正有助于像素扩散 Transformer 生成，发现理解能力强不等于生成有用。","en":"Studies which visual representations actually help pixel diffusion transformers, finding better understanding does not imply better generation."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":265},{"id":"arxiv-2610.09087","arxiv":"2610.09087","title":"U-Space: Uncovering When and Why Uncertainty Arises in Language Models","authors":["Tobias Braun","Nils Loose","Alexander Herzog","Virginia Ceccatelli","Marcus Rohrbach","Thomas Eisenbarth","Lorenzo Cavallaro"],"authorCount":7,"org":"Multimodal AI Lab - Technical University of Darmstadt","affiliations":[],"orgNote":null,"date":"2026-10-06","hfDaily":"2026-10-09","primaryCategory":"cs.CL","categories":["cs.CL","cs.AI","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2610.09087","pdf":"https://arxiv.org/pdf/2610.09087","hf":"https://huggingface.co/papers/2610.09087","code":"https://github.com/s2labres/U-Space","project":"https://s2lab.org/U-Space/"},"metrics":{"hfUpvotes":54,"hfComments":6,"citations":0,"citationsSource":"semantic_scholar","githubStars":11},"hot":49.2,"tags":["safety","llm"],"summary":{"zh":"U-Space 分析语言模型的不确定性在何时、为何产生，帮助判断单个回答是否可信。","en":"U-Space analyses when and why uncertainty arises in language models, to help judge whether an individual answer can be trusted."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":266},{"id":"arxiv-2609.38078","arxiv":"2609.38078","title":"MotorMind: Scaffolding General Vision Language Models for Zero-Shot Robot Manipulation","authors":["Bingxuan Li","Siqi Song","Yizhuo Wu","Jiarui Yao","Tong Zhang","Huan Zhang"],"authorCount":6,"org":"University of Illinois at Urbana-Champaign","affiliations":[],"orgNote":null,"date":"2026-09-29","hfDaily":"2026-10-05","primaryCategory":"cs.RO","categories":["cs.RO"],"links":{"abs":"https://arxiv.org/abs/2609.38078","pdf":"https://arxiv.org/pdf/2609.38078","hf":"https://huggingface.co/papers/2609.38078","code":null,"project":"https://motor-mind.github.io"},"metrics":{"hfUpvotes":115,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":49.18,"tags":["robotics","multimodal"],"summary":{"zh":"MotorMind 为通用视觉语言模型搭建脚手架，让它不经专门训练就能零样本操控机器人。","en":"MotorMind scaffolds general vision-language models for zero-shot robot manipulation without specialised VLA training."},"disciplines":["software-engineering","robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":267},{"id":"arxiv-2609.24997","arxiv":"2609.24997","title":"VideoGen-Agent: Reinforcing Video Generation Agents","authors":["Binxu Li","Haoyi Duan","Yuhui Zhang","Yaohui Zhang","Zihao Lin","Kaituo Feng","Suozhi Huang","Xiangyi Li"],"authorCount":12,"org":"Princeton University","affiliations":[],"orgNote":null,"date":"2026-09-21","hfDaily":"2026-09-22","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2609.24997","pdf":"https://arxiv.org/pdf/2609.24997","hf":"https://huggingface.co/papers/2609.24997","code":null,"project":"https://andyca111.github.io/VideoGen_Agent/"},"metrics":{"hfUpvotes":73,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":49.12,"tags":["visual-gen","agent","rl"],"summary":{"zh":"VideoGen-Agent 用多任务智能体强化学习训练视频生成智能体调用外部工具，满足需要专门知识的提示。","en":"VideoGen-Agent trains a multimodal agent with multitask agentic RL to use tools for knowledge- and identity-heavy video prompts."},"disciplines":["film-animation"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":268},{"id":"arxiv-2610.20399","arxiv":"2610.20399","title":"Cortex: Consolidating Context into Parametric Memory","authors":["Jintian Zhang","Feng He","Xinnian Liang","Ziwen Xu","Haoming Xu","Haojing Huang","Shuangzhi Wu","Ningyu Zhang"],"authorCount":8,"org":"ByteDance Seed","affiliations":[],"orgNote":null,"date":"2026-10-14","hfDaily":"2026-10-15","primaryCategory":"cs.CL","categories":["cs.CL","cs.AI","cs.CV","cs.LG","cs.MA"],"links":{"abs":"https://arxiv.org/abs/2610.20399","pdf":"https://arxiv.org/pdf/2610.20399","hf":"https://huggingface.co/papers/2610.20399","code":null,"project":null},"metrics":{"hfUpvotes":73,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":49.12,"tags":["llm","efficiency"],"summary":{"zh":"Cortex 借鉴生物记忆巩固，把上下文带来的能力提升内化进模型参数，移除上下文后仍保留。","en":"Cortex consolidates gains from context into parametric memory, keeping them after the context is removed."},"disciplines":["life-sciences"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":269},{"id":"arxiv-2610.25736","arxiv":"2610.25736","title":"Learning from Fleet Experience: Collective Experience Learning Across 96 Robots","authors":["Chao Yu","Shu'ang Yu","Hao Lin","Hongzhi Zang","Yuxiang Yan","Chenxi Jiang","Zhen Guo","Chuankang Li"],"authorCount":44,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-10-19","hfDaily":"2026-10-20","primaryCategory":"cs.RO","categories":["cs.RO"],"links":{"abs":"https://arxiv.org/abs/2610.25736","pdf":"https://arxiv.org/pdf/2610.25736","hf":"https://huggingface.co/papers/2610.25736","code":null,"project":"https://cel-robot-learning.github.io"},"metrics":{"hfUpvotes":112,"hfComments":2,"citations":null,"citationsSource":null,"githubStars":null},"hot":48.92,"tags":["robotics"],"summary":{"zh":"提出“集体经验学习”：96 台机器人共享执行经验，按失败情境检索成功片段来微调共同策略，无需额外人工示范。","en":"Collective experience learning lets 96 deployed robots share execution data, retrieving successful segments for failures to fine-tune a common policy without new demos."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":270},{"id":"arxiv-2610.25078","arxiv":"2610.25078","title":"CoEvo-RSI: When Harnesses Evolve, Which Models Should Follow?","authors":["Jiaqi Su","Guibin Zhang","Jie-Jing Shao","Cong Pang","Tiankuo Yao","Ivor Tsang","Haiyan Yin","Shuicheng Yan"],"authorCount":8,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-10-18","hfDaily":"2026-10-21","primaryCategory":"cs.LG","categories":["cs.LG","cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.25078","pdf":"https://arxiv.org/pdf/2610.25078","hf":"https://huggingface.co/papers/2610.25078","code":"https://github.com/SuJQ0722/CoEvo-RSI","project":null},"metrics":{"hfUpvotes":68,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":5},"hot":48.9,"tags":["agent","llm"],"summary":{"zh":"CoEvo-RSI 研究模型与执行框架（harness）共同进化时的“脑体不匹配”，由元智能体协调框架演化和模型选择。","en":"CoEvo-RSI tackles the brain–body mismatch when models and agent harnesses co-evolve, using a meta-agent to coordinate harness evolution and checkpoint choice."},"disciplines":["neuroscience"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":271},{"id":"arxiv-2610.01892","arxiv":"2610.01892","title":"Selection-Based Structured Reasoning: Toward Efficient Multimodal Search Agents","authors":["Feiyu Gavin Zhu","Xiaoyu Zhu","Jiqi Yang","Rui Yang","Arnab Kumar Mondal","Yancheng Wang","Xinke Deng","Jean Oh"],"authorCount":12,"org":"Carnegie Mellon University","affiliations":[],"orgNote":null,"date":"2026-10-01","hfDaily":"2026-10-07","primaryCategory":"cs.LG","categories":["cs.LG","cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.01892","pdf":"https://arxiv.org/pdf/2610.01892","hf":"https://huggingface.co/papers/2610.01892","code":"https://github.com/zfy0314/ssr-unofficial","project":"https://zfy0314.github.io/ssr-webpage/"},"metrics":{"hfUpvotes":38,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":10},"hot":48.71,"tags":["reasoning","efficiency","agent"],"summary":{"zh":"SSR 让小模型从预设推理候选中“选择”而不是自由生成，降低多模态搜索 Agent 的推理成本。","en":"SSR has small models select from predefined reasoning candidates instead of free generation, cutting cost for multimodal search agents."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":272},{"id":"arxiv-2610.25043","arxiv":"2610.25043","title":"VLA-Rolling: Fewer Denoising Steps, Better Control via Adaptive Action Reuse","authors":["Sicheng He","Yi Pan","Qiuting Yu","Jiaming Huang","Qi Lu","Shuyan Wu","Hongyan Feng","Xuanyu Liu"],"authorCount":13,"org":"ZJU-OmniAI","affiliations":[],"orgNote":null,"date":"2026-10-18","hfDaily":"2026-10-21","primaryCategory":"cs.AI","categories":["cs.AI","cs.RO"],"links":{"abs":"https://arxiv.org/abs/2610.25043","pdf":"https://arxiv.org/pdf/2610.25043","hf":"https://huggingface.co/papers/2610.25043","code":"https://github.com/ZJU-OmniAI/VLA-Rolling","project":"https://chenbao123ww.github.io/VLA-Rolling/"},"metrics":{"hfUpvotes":60,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":9},"hot":48.51,"tags":["robotics","efficiency"],"summary":{"zh":"VLA-Rolling 复用上一段未执行的动作尾巴作为去噪起点，用更少步数生成动作，冻结模型即可提速并提高成功率。","en":"VLA-Rolling reuses the unexecuted action tail as the starting point for denoising, cutting steps for frozen flow-matching VLAs while often raising success."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":273},{"id":"arxiv-2610.14746","arxiv":"2610.14746","title":"BridgeWAM: Bridging Video and Action Experts via Latent Bridge Queries","authors":["Jiaxin Fang","Ruitong Liu","Zeyu Lai","Xingyu Wang","Mingyuan Jia","Pingyue Sheng","Weikang Wan","Mengdi Xu"],"authorCount":9,"org":"Fudan University","affiliations":[],"orgNote":null,"date":"2026-10-10","hfDaily":"2026-10-19","primaryCategory":"cs.RO","categories":["cs.RO","cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.14746","pdf":"https://arxiv.org/pdf/2610.14746","hf":"https://huggingface.co/papers/2610.14746","code":"https://github.com/jxfang04/BridgeWAM-code","project":"https://bridgewam.github.io/Bridge-Wam-show/"},"metrics":{"hfUpvotes":46,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":22},"hot":48.42,"tags":["robotics","multimodal"],"summary":{"zh":"BridgeWAM 用 32 个潜在桥接查询连接视频专家与动作专家，替代逐层稠密耦合，动作头只需两层。","en":"BridgeWAM links video and action experts through 32 latent bridge queries instead of dense layer-wise coupling, enabling a two-layer action head."},"disciplines":["robotics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":274},{"id":"arxiv-2610.21459","arxiv":"2610.21459","title":"Can Visual Generation Expand What Unified Multimodal Models Know?","authors":["Wenkai Li","Wenyi Hong","Zhuoyi Yang","Jie Tang"],"authorCount":4,"org":"Z.ai","affiliations":[],"orgNote":null,"date":"2026-10-15","hfDaily":"2026-10-19","primaryCategory":"cs.CV","categories":["cs.CV","cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.21459","pdf":"https://arxiv.org/pdf/2610.21459","hf":"https://huggingface.co/papers/2610.21459","code":"https://github.com/kevinlllee/Knowledge-Transfer-Probing","project":"https://kevinlllee.github.io/Knowledge-Transfer-Probing/"},"metrics":{"hfUpvotes":34,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":14},"hot":47.97,"tags":["multimodal"],"summary":{"zh":"用受控实验检验统一多模态模型能否通过学画图获得新知识并用文字答出，发现存在迁移但不均衡。","en":"A controlled probe shows unified multimodal models can transfer knowledge learned through image generation into text answers, though unevenly."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":275},{"id":"arxiv-2610.00559","arxiv":"2610.00559","title":"PhysVista: Benchmarking Physical Intelligence in VLMs via a Perception-Reasoning-Assessment Loop","authors":["Xinge Peng","Yiting Lu","Tianwu Zhi","Wen Wen","Jianzhao Liu","Xin Li","Zhibo Chen"],"authorCount":7,"org":"university of science and technology  of china","affiliations":[],"orgNote":null,"date":"2026-09-30","hfDaily":"2026-10-02","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.00559","pdf":"https://arxiv.org/pdf/2610.00559","hf":"https://huggingface.co/papers/2610.00559","code":null,"project":null},"metrics":{"hfUpvotes":69,"hfComments":3,"citations":1,"citationsSource":"semantic_scholar","githubStars":null},"hot":47.34,"tags":["multimodal","benchmark","reasoning"],"summary":{"zh":"PhysVista 通过“感知-推理-判断”闭环评测视觉语言模型的物理常识与判断能力。","en":"PhysVista evaluates VLMs' physical intelligence through a perception-reasoning-assessment loop."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":276},{"id":"arxiv-2610.17656","arxiv":"2610.17656","title":"Open-o3-Video-v2: Seek, Ground, and Reason with Verifiable Spatio-Temporal Evidence","authors":["Jiahao Meng","Zhongyuan Bao","Yue Tan","Haiquan Wen","Renrui Zhang","Anna Wang","Jason Li","Yanwei Fu"],"authorCount":10,"org":"Peking University","affiliations":[],"orgNote":null,"date":"2026-10-13","hfDaily":"2026-10-14","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.17656","pdf":"https://arxiv.org/pdf/2610.17656","hf":"https://huggingface.co/papers/2610.17656","code":"https://github.com/Multimedia-Semantic-Analytics-Lab/Open-o3-Video-v2","project":"https://msalab-pku.github.io/projects/Open-o3-Video-v2/"},"metrics":{"hfUpvotes":29,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":19},"hot":47.15,"tags":["multimodal","reasoning","agent"],"summary":{"zh":"Open-o3-Video-v2 让模型在视频推理中主动补看画面，并把时间点和物体框写入推理过程作为可验证证据。","en":"Open-o3-Video-v2 actively gathers more frames and writes timestamps and boxes into its reasoning as verifiable evidence."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":277},{"id":"arxiv-2610.22169","arxiv":"2610.22169","title":"Lexical Reference Frames: Interpreting LM Residual Streams in Native Semantics","authors":["Jian Gu","Hongyu Zhang","Chunyang Chen","Aldeida Aleti"],"authorCount":4,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-10-15","hfDaily":"2026-10-19","primaryCategory":"cs.CL","categories":["cs.CL","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2610.22169","pdf":"https://arxiv.org/pdf/2610.22169","hf":"https://huggingface.co/papers/2610.22169","code":"https://github.com/jianguda/lexframe","project":"https://jianguda.github.io/blogs/lexframe/"},"metrics":{"hfUpvotes":55,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":6},"hot":46.77,"tags":["safety","llm"],"summary":{"zh":"提出“词汇参照系”，借模型自身后续层读出中间隐藏状态的含义，比 logit lens 等方法更忠实。","en":"Lexical reference frames read hidden states through the model's own remaining layers, giving far more faithful readings than fixed lenses."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":278},{"id":"arxiv-2610.04299","arxiv":"2610.04299","title":"Questioning the Questions: Sustaining Self-Evolution in Reasoning Models","authors":["Jinyuan Li","Chengsong Huang","Langlin Huang","Donghong Cai","Shiping Gao","Yuyi Yang","Jiaxin Huang"],"authorCount":7,"org":"Huang's INTelligence lab","affiliations":[],"orgNote":null,"date":"2026-10-03","hfDaily":"2026-10-08","primaryCategory":"cs.LG","categories":["cs.LG","cs.AI","cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.04299","pdf":"https://arxiv.org/pdf/2610.04299","hf":"https://huggingface.co/papers/2610.04299","code":null,"project":"https://jinyuanli0012.github.io/R-Quest/"},"metrics":{"hfUpvotes":85,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":46.62,"tags":["reasoning","rl"],"summary":{"zh":"分析自进化推理模型为何会崩溃（无效题和重复题变多），并提出维持自进化的出题质量控制。","en":"Analyses why self-evolving reasoning models collapse (invalid and repeated questions) and proposes question-quality controls."},"disciplines":["mathematics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":279},{"id":"arxiv-2610.28594","arxiv":"2610.28594","title":"Learning the 4D World through Point Trajectory Forecasting","authors":["Byungwoo Jeon","Sunghwan Hong","Dohyeon Kim","Jonghoon Lee","Jungwoo Park","Inhee Lee","Doosan Baek","Jinil Kim"],"authorCount":13,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-10-20","hfDaily":"2026-10-21","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.28594","pdf":"https://arxiv.org/pdf/2610.28594","hf":"https://huggingface.co/papers/2610.28594","code":"https://github.com/rootyJeon/VPM","project":"https://rootyjeon.github.io/video-point-model/"},"metrics":{"hfUpvotes":42,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":17},"hot":46.49,"tags":["visual-gen","robotics"],"summary":{"zh":"VPM 世界模型同时预测多视角未来视频和场景点的三维轨迹，并扩展到机器人控制。","en":"VPM is a world model that jointly forecasts multi-view future video and 3D point trajectories, extended to robot control."},"disciplines":["film-animation"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":280},{"id":"arxiv-2610.24960","arxiv":"2610.24960","title":"CoworkEnv: Synthesizing Training Environments for Long-Horizon Cowork Agents","authors":["Zhongwei Xie","Yiming Du","Jiaxin Bai","Baixuan Xu","Chunyang Li","Xiyu Ren","Wei Fan","Tianshi Zheng"],"authorCount":13,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-10-18","hfDaily":"2026-10-20","primaryCategory":"cs.AI","categories":["cs.AI","cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.24960","pdf":"https://arxiv.org/pdf/2610.24960","hf":"https://huggingface.co/papers/2610.24960","code":"https://github.com/HKUST-KnowComp/CoworkEnv","project":null},"metrics":{"hfUpvotes":42,"hfComments":4,"citations":0,"citationsSource":"semantic_scholar","githubStars":12},"hot":46.44,"tags":["agent","rl"],"summary":{"zh":"CoworkEnv 自动合成可执行的“AI 同事”训练环境（任务、工具、评分），用于监督微调和强化学习小模型。","en":"CoworkEnv synthesizes executable training environments for AI coworker agents, supporting SFT and RL of compact models."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":281},{"id":"arxiv-2610.23787","arxiv":"2610.23787","title":"DLite: History-free Speculative Decoding for Agentic AI","authors":["Jiaming Xu","Hanzhen Wang","Tianlang Zhao","Jiewen Xiao","Chengze Yuan","Jiayi Pan","Yongkang Zhou","Jiancai Ye"],"authorCount":10,"org":"Shanghai Jiao Tong University","affiliations":[],"orgNote":null,"date":"2026-10-16","hfDaily":"2026-10-19","primaryCategory":"cs.CL","categories":["cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.23787","pdf":"https://arxiv.org/pdf/2610.23787","hf":"https://huggingface.co/papers/2610.23787","code":"https://github.com/SJTU-DAI-Sys/DLite","project":null},"metrics":{"hfUpvotes":32,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":43},"hot":46.07,"tags":["efficiency","agent"],"summary":{"zh":"DLite 是不保留历史上下文的推测解码：直接复用目标模型隐藏状态做局部草稿，适合上下文不断增长的智能体任务。","en":"DLite is history-free speculative decoding that drafts locally from the target model's hidden states, suited to ever-growing agent contexts."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":282},{"id":"arxiv-2610.23230","arxiv":"2610.23230","title":"RAP: A Dataset for Paper-to-Poster Generation via Declarative Markup from Human-Crafted Exemplars","authors":["Haonan Han","Yukang Lin","Minliang Lin","Haoxian Tan","Jun-Yan He","Rui Yang","Zhi-Qi Cheng","Jie Hu"],"authorCount":10,"org":"meituan","affiliations":[],"orgNote":null,"date":"2026-10-16","hfDaily":"2026-10-19","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.23230","pdf":"https://arxiv.org/pdf/2610.23230","hf":"https://huggingface.co/papers/2610.23230","code":"https://github.com/VincentHancoder/RAP","project":"https://reconstruct-a-poster.github.io/"},"metrics":{"hfUpvotes":35,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":29},"hot":45.99,"tags":["benchmark","code"],"summary":{"zh":"RAP 数据集收录 2.5 万多对论文与海报，把海报还原成可渲染比对的 HTML/CSS，用于论文转海报生成。","en":"RAP pairs 25K+ papers with posters derendered into HTML/CSS slots, enabling render-and-compare evaluation of paper-to-poster generation."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":283},{"id":"arxiv-2610.23181","arxiv":"2610.23181","title":"Tree-Aware Posterior Distillation for Long-Horizon Agentic Reinforcement Learning","authors":["Heming Zou","Qi Wang","Yixiu Mao","Yun Qu","Yuyi Guo","Zeyuan Liu","Chenglin Li","Weijie Liu"],"authorCount":10,"org":"Tencent Hunyuan","affiliations":[],"orgNote":null,"date":"2026-10-16","hfDaily":"2026-10-19","primaryCategory":"cs.LG","categories":["cs.LG","cs.AI","cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.23181","pdf":"https://arxiv.org/pdf/2610.23181","hf":"https://huggingface.co/papers/2610.23181","code":null,"project":null},"metrics":{"hfUpvotes":52,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":45.78,"tags":["rl","agent"],"summary":{"zh":"TreeOPSD 把多轮交互建成树，回传终局奖励得到每轮的隐式信用，再做无评论家的自蒸馏，用于长程智能体强化学习。","en":"TreeOPSD builds a tree over agent turns, backs up terminal rewards into turn-level credit and self-distills without a critic for long-horizon agent RL."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":284},{"id":"arxiv-2610.25404","arxiv":"2610.25404","title":"MT Forcing: Towards Stable and Consistent Multi-Turn Image Editing with Autoregressive Diffusion","authors":["Houyuan Chen","Fanzhou Wang","Xiangyu Fan","Zesong Qiu","Yukun Wei","Songlin Yang","Zhitao Yang","Lvmin Zhang"],"authorCount":13,"org":"SenseNova","affiliations":[],"orgNote":null,"date":"2026-10-18","hfDaily":"2026-10-20","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.25404","pdf":"https://arxiv.org/pdf/2610.25404","hf":"https://huggingface.co/papers/2610.25404","code":"https://github.com/houyuanchen111/MT-Forcing","project":"https://houyuanchen111.github.io/mt-forcing.github.io/"},"metrics":{"hfUpvotes":37,"hfComments":1,"citations":null,"citationsSource":null,"githubStars":22},"hot":45.25,"tags":["visual-gen"],"summary":{"zh":"MT Forcing 用单轮蒸馏教师和教师强制引导训练自回归扩散模型，减少多轮图像编辑中的画质漂移和内容走样。","en":"MT Forcing trains autoregressive-diffusion editors with a single-turn distillation teacher and teacher-forcing guidance to curb multi-turn editing drift."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":285},{"id":"arxiv-2610.26216","arxiv":"2610.26216","title":"RoAR-VAE: Improving Latent Diffusability through Robust Prediction","authors":["Yang Yue","Fangyun Wei","Yitong Wang","Jinjing Zhao","Henry Zheng","Dong Chen","Gao Huang"],"authorCount":7,"org":"Tsinghua-LeapLab","affiliations":[],"orgNote":null,"date":"2026-10-19","hfDaily":"2026-10-20","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.26216","pdf":"https://arxiv.org/pdf/2610.26216","hf":"https://huggingface.co/papers/2610.26216","code":"https://github.com/LeapLabTHU/RoAR-VAE","project":null},"metrics":{"hfUpvotes":27,"hfComments":3,"citations":null,"citationsSource":null,"githubStars":8},"hot":44.89,"tags":["visual-gen"],"summary":{"zh":"RoAR-VAE 提出“稳健预测”原则：让被破坏的潜变量仍能恢复语义和图像，从而让潜空间更容易被扩散模型学习。","en":"RoAR-VAE trains tokenizers so semantics and pixels remain recoverable from corrupted latents, making latents easier for diffusion models to learn."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":286},{"id":"arxiv-2610.26851","arxiv":"2610.26851","title":"Manga2Animate: An Agentic Framework for Faithful Manga Animation with Script and Storyboard Planning","authors":["Sheng-Wei Zhang","Hsi-An Chen","Yan-Ting Liou","Chih-Hsuan Chen","Shih-Chen Tseng","Yu-Lun Liu"],"authorCount":6,"org":"National Yang Ming Chiao Tung University","affiliations":[],"orgNote":null,"date":"2026-10-19","hfDaily":"2026-10-20","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.26851","pdf":"https://arxiv.org/pdf/2610.26851","hf":"https://huggingface.co/papers/2610.26851","code":"https://github.com/annchenn/Manga2Animate","project":"https://annchenn.github.io/Manga2Animate/"},"metrics":{"hfUpvotes":26,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":47},"hot":44.28,"tags":["agent","visual-gen"],"summary":{"zh":"Manga2Animate 用“创作+审核”配对的多智能体先写动画剧本和分镜，再生成忠于原作的漫画动画。","en":"Manga2Animate uses paired creator and reviewer agents to plan scripts and storyboards before generating animation faithful to the source manga."},"disciplines":["film-animation"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":287},{"id":"arxiv-2610.23542","arxiv":"2610.23542","title":"Execution Before Memory: Diagnosing Memory Failures in Video World Models","authors":["Weiyang Jin","Jiwen Yu","Sihan Xu","Yuwei Niu","Xihui Liu"],"authorCount":5,"org":"The University of Hong Kong","affiliations":[],"orgNote":null,"date":"2026-10-16","hfDaily":"2026-10-19","primaryCategory":"cs.CV","categories":["cs.CV","cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.23542","pdf":"https://arxiv.org/pdf/2610.23542","hf":"https://huggingface.co/papers/2610.23542","code":"https://github.com/HKU-MMLab/MIST","project":"https://hku-mmlab.github.io/MIST/"},"metrics":{"hfUpvotes":33,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":12},"hot":43.32,"tags":["benchmark","visual-gen"],"summary":{"zh":"MIST 评测把“是否执行了指令”与“是否记住”分开，发现视频世界模型的瓶颈多在执行而非记忆。","en":"MIST separates instruction execution from memory in video world model tests, finding the bottleneck is usually execution rather than forgetting."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":288},{"id":"arxiv-2610.26853","arxiv":"2610.26853","title":"InfiniDream: Open-Ended Photorealistic 3D Worlds Through Procedural Generation","authors":["Yi-Chuan Huang","Hao-Jen Chien","Zheng-Hui Huang","Yu-Lun Liu","Kaipeng Zhang","Zhixiang Wang"],"authorCount":6,"org":"Alaya Lab","affiliations":[],"orgNote":null,"date":"2026-10-19","hfDaily":"2026-10-20","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.26853","pdf":"https://arxiv.org/pdf/2610.26853","hf":"https://huggingface.co/papers/2610.26853","code":"https://github.com/yichuanH/InfiniDream","project":"https://yichuanh.github.io/infinidream-page/"},"metrics":{"hfUpvotes":26,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":31},"hot":43.27,"tags":["vision","code"],"summary":{"zh":"InfiniDream 让视觉语言模型调用程序化生成算法，开放式地生成可编辑、照片级的 3D 资产和大场景。","en":"InfiniDream has a VLM drive procedural algorithms to generate open-ended, editable photorealistic 3D assets and large worlds."},"disciplines":["software-engineering"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":289},{"id":"arxiv-2610.26844","arxiv":"2610.26844","title":"AnyGoal: On World Model Generalization Beyond Expert Rollouts","authors":["Heejeong Nam","Divyat Mahajan","Quentin Le Lidec","Randall Balestriero","David Klindt"],"authorCount":5,"org":"galilai-group","affiliations":[],"orgNote":null,"date":"2026-10-19","hfDaily":"2026-10-20","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.26844","pdf":"https://arxiv.org/pdf/2610.26844","hf":"https://huggingface.co/papers/2610.26844","code":"https://github.com/galilai-group/AnyGoal","project":"https://hazel-heejeong-nam.github.io/anygoal/"},"metrics":{"hfUpvotes":37,"hfComments":2,"citations":null,"citationsSource":null,"githubStars":7},"hot":43.22,"tags":["benchmark","agent"],"summary":{"zh":"AnyGoal 基准测试世界模型能否组合已见过的状态和动作因素，发现现有模型在这类重组上普遍失败。","en":"AnyGoal benchmarks whether world models generalize to recombinations of familiar state and action factors; current models consistently fail."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":290},{"id":"arxiv-2610.27911","arxiv":"2610.27911","title":"MMEnterprise-Bench: Evaluating Enterprise Agents over Evolving Multimodal Records","authors":["Seohyun Lee","Jongha Kim","Dohwan Ko","Ji Soo Lee","Seoung Choi","Jayun Ku","Hyunwoo J. Kim"],"authorCount":7,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-10-20","hfDaily":"2026-10-21","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.27911","pdf":"https://arxiv.org/pdf/2610.27911","hf":"https://huggingface.co/papers/2610.27911","code":"https://github.com/mlvlab/MMEnterprise-Bench","project":"https://mlvlab.github.io/MMEnterprise-Bench/"},"metrics":{"hfUpvotes":24,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":11},"hot":40.48,"tags":["benchmark","agent"],"summary":{"zh":"MMEnterprise-Bench 含 1613 道题，基于不断修订的企业多模态记录，考察证据推理、冲突消解和拒答。","en":"MMEnterprise-Bench has 1,613 questions over evolving multimodal workplace records, testing evidence reasoning, conflict resolution and unanswerability."},"disciplines":["economics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":291},{"id":"arxiv-2610.03195","arxiv":"2610.03195","title":"Source Preference in the Wild: How LLM Agents Favor Items by Source, and How to Reduce It","authors":["Jonghyun Song","Haewon Park","Jeonghoon Shim","Woojung Song","Yohan Jo"],"authorCount":5,"org":"Seoul National University","affiliations":[],"orgNote":null,"date":"2026-10-02","hfDaily":"2026-10-05","primaryCategory":"cs.CL","categories":["cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.03195","pdf":"https://arxiv.org/pdf/2610.03195","hf":"https://huggingface.co/papers/2610.03195","code":null,"project":null},"metrics":{"hfUpvotes":47,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":40.36,"tags":["alignment","agent"],"summary":{"zh":"发现 12 个 Agent 模型在购物、订酒店、引文献时都会偏好某些来源，并提出缓解方法。","en":"Finds 12 agent models systematically favour items from certain sources when shopping, booking or citing, and proposes mitigations."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":292},{"id":"arxiv-2610.06830","arxiv":"2610.06830","title":"MemPilot: Orchestrating On-Demand Multimodal Memory Curation for LLM Agents","authors":["Haozhen Zhang","Haodong Yue","Quanyu Long","Jianzhu Bao","Qingyuan Liu","Tao Feng","Bohan Liu","Weida Liang"],"authorCount":9,"org":"Nanyang Technological University","affiliations":[],"orgNote":null,"date":"2026-10-05","hfDaily":"2026-10-20","primaryCategory":"cs.CL","categories":["cs.CL","cs.AI","cs.LG"],"links":{"abs":"https://arxiv.org/abs/2610.06830","pdf":"https://arxiv.org/pdf/2610.06830","hf":"https://huggingface.co/papers/2610.06830","code":"https://github.com/ViktorAxelsen/MemPilot","project":"https://viktoraxelsen.github.io/MemPilot/"},"metrics":{"hfUpvotes":28,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":6},"hot":40.19,"tags":["agent","multimodal"],"summary":{"zh":"MemPilot 用强化学习训练策略，按性能、成本和延迟偏好按需整理多模态记忆，决定检索还是交给不同模型处理。","en":"MemPilot trains an RL policy to curate multimodal agent memory on demand under performance–cost–latency preferences."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":293},{"id":"arxiv-2610.26770","arxiv":"2610.26770","title":"HarnessPro: Recursive Harness Self-Improvement with Pareto Optimization for Performance-Cost Trade-offs","authors":["Zhuoshi Pan","ZhiYuan Feng","Yukuo Cen","Honglin Lin","Junru Lu","Zhongxia Chen","Shuo Yang","H. Vicky Zhao"],"authorCount":9,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-10-19","hfDaily":"2026-10-20","primaryCategory":"cs.LG","categories":["cs.LG","cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.26770","pdf":"https://arxiv.org/pdf/2610.26770","hf":"https://huggingface.co/papers/2610.26770","code":"https://github.com/pzs19/HarnessPro","project":"https://pzs19.github.io/HarnessPro/"},"metrics":{"hfUpvotes":27,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":4},"hot":38.99,"tags":["agent","code"],"summary":{"zh":"HarnessPro 在无标准答案的情况下迭代改进智能体执行框架，同时给出省钱、均衡、高性能三档方案。","en":"HarnessPro evolves agent harnesses without ground-truth tests, returning lite, balanced and heavy variants that trade performance against cost."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":294},{"id":"arxiv-2610.25844","arxiv":"2610.25844","title":"SWIFT: Self-guided Weak-to-strong dIFfusion models via Target extrapolation","authors":["Yanjin He","Cuifeng Shen","Wanzhou Lei","Maohua Li","Tao Lan","Kan Liu","Hanlin Tang","Qing Qu"],"authorCount":8,"org":"alibaba","affiliations":[],"orgNote":null,"date":"2026-10-19","hfDaily":"2026-10-21","primaryCategory":"cs.LG","categories":["cs.LG","cs.AI","cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.25844","pdf":"https://arxiv.org/pdf/2610.25844","hf":"https://huggingface.co/papers/2610.25844","code":null,"project":"https://yanjinhe.github.io/swift/"},"metrics":{"hfUpvotes":27,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":38.97,"tags":["visual-gen"],"summary":{"zh":"SWIFT 用冻结的早期弱模型给训练目标做外推，放大强模型欠拟合的部分，提升扩散模型质量且不增加推理开销。","en":"SWIFT extrapolates training targets away from a frozen weak denoiser to boost a stronger diffusion model with no extra inference cost."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":295},{"id":"arxiv-2610.25950","arxiv":"2610.25950","title":"AdaExploreBench: Benchmarking LLM Agents for Proactive Exploration and Test-time Adaptation in Evolving Environments","authors":["Rui Wang","Zishen Lai","Jiayu Liu","Cheng Qian","Jeonghwan Kim","Heng Wang","Xuejun Zhang","Jiateng Liu"],"authorCount":18,"org":"HKUST","affiliations":[],"orgNote":null,"date":"2026-10-19","hfDaily":"2026-10-20","primaryCategory":"cs.CL","categories":["cs.CL","cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.25950","pdf":"https://arxiv.org/pdf/2610.25950","hf":"https://huggingface.co/papers/2610.25950","code":"https://github.com/HKUST-KnowComp/AdaExploreBench","project":"https://hkust-knowcomp.github.io/AdaExploreBench/"},"metrics":{"hfUpvotes":28,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":3},"hot":38.79,"tags":["benchmark","agent"],"summary":{"zh":"AdaExploreBench 用隐藏且会悄悄变化的规则，测试智能体能否主动探索并在测试时自我调整。","en":"AdaExploreBench uses hidden, silently drifting rules to test whether LLM agents proactively explore and adapt at test time."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":296},{"id":"arxiv-2610.09221","arxiv":"2610.09221","title":"Consistent Distribution Matching for Data-Free Diffusion Distillation","authors":["Yuxiang Fu","Qi Yan","Zike Wu","Yongxing Zhang","Purang Abolmaesumi","Lele Wang","Renjie Liao"],"authorCount":7,"org":"Vector Institute","affiliations":[],"orgNote":null,"date":"2026-10-06","hfDaily":"2026-10-19","primaryCategory":"cs.LG","categories":["cs.LG","cs.AI","cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.09221","pdf":"https://arxiv.org/pdf/2610.09221","hf":"https://huggingface.co/papers/2610.09221","code":"https://github.com/felix-yuxiang/CDMD","project":"https://consistentdmd.github.io/"},"metrics":{"hfUpvotes":22,"hfComments":1,"citations":0,"citationsSource":"semantic_scholar","githubStars":9},"hot":38.15,"tags":["visual-gen","efficiency"],"summary":{"zh":"提出无需数据和教师采样的扩散蒸馏方法，单个学生网络同时生成样本和估计分数，一步生成即达很低 FID。","en":"A data-free, simulation-free diffusion distillation where one student both generates and estimates scores, reaching strong one-step FID."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":297},{"id":"arxiv-2610.13623","arxiv":"2610.13623","title":"APEX-Harm: Evaluating Agent Safety in Shared Professional Workspaces","authors":["Ziyi Yang","Adel Bibi","Guohao Li","Bernard Ghanem","Carlos Hinojosa"],"authorCount":5,"org":"King Abdullah University of Science and Technology","affiliations":[],"orgNote":null,"date":"2026-10-09","hfDaily":"2026-10-15","primaryCategory":"cs.CR","categories":["cs.CR","cs.CL"],"links":{"abs":"https://arxiv.org/abs/2610.13623","pdf":"https://arxiv.org/pdf/2610.13623","hf":"https://huggingface.co/papers/2610.13623","code":"https://github.com/echo-yiyiyi/apex-harm","project":null},"metrics":{"hfUpvotes":26,"hfComments":1,"citations":0,"citationsSource":"semantic_scholar","githubStars":4},"hot":38.02,"tags":["safety","agent","benchmark"],"summary":{"zh":"APEX-Harm 在含上百个文件的共享工作区里，测试金融、咨询、法律 Agent 抵御提示和脚本注入的能力。","en":"APEX-Harm tests finance, consulting and legal agents against prompt and script injection in large shared workspaces."},"disciplines":["cybersecurity"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":298},{"id":"arxiv-2610.06577","arxiv":"2610.06577","title":"Optimizing the Optimizer: Language Models Discover Faster Molecular Relaxation Algorithms","authors":["Artem Tsypin","Vladimir Deshchenya","Kuzma Khrabrov","Denis Potapov","Maxim Radchenko","Artur Kadurin","Michael G. Medvedev"],"authorCount":7,"org":null,"affiliations":[],"orgNote":"机构未在 arXiv/HF 元数据中标注（待核实）","date":"2026-10-05","hfDaily":"2026-10-06","primaryCategory":"physics.chem-ph","categories":["physics.chem-ph"],"links":{"abs":"https://arxiv.org/abs/2610.06577","pdf":"https://arxiv.org/pdf/2610.06577","hf":"https://huggingface.co/papers/2610.06577","code":"https://github.com/vdeshchenya/autosella","project":null},"metrics":{"hfUpvotes":24,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":4},"hot":37.86,"tags":["ai4science","agent"],"summary":{"zh":"让语言模型 Agent 改写分子几何优化器，找到比现有最快开源方法更省力的算法。","en":"Has an LLM agent rewrite molecular geometry optimisers, discovering algorithms needing fewer force calls than the fastest open baseline."},"disciplines":["chemistry","physics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":299},{"id":"arxiv-2609.29167","arxiv":"2609.29167","title":"IndicBankBench: Evaluating Safety and Reliability of Language Model Assistants in Indian Retail Banking","authors":["Suvradip Paul","Chandra Bhushan","Harsh Sharma","Nitin Kukreja","Yatharth Dedhia","Keyur Doshi","Prashant Devadiga"],"authorCount":7,"org":"National Payments Corporation of India","affiliations":[],"orgNote":null,"date":"2026-09-24","hfDaily":"2026-09-28","primaryCategory":"cs.AI","categories":["cs.AI"],"links":{"abs":"https://arxiv.org/abs/2609.29167","pdf":"https://arxiv.org/pdf/2609.29167","hf":"https://huggingface.co/papers/2609.29167","code":"https://github.com/npci/IndicBankBench","project":null},"metrics":{"hfUpvotes":21,"hfComments":1,"citations":0,"citationsSource":"semantic_scholar","githubStars":9},"hot":37.71,"tags":["safety","benchmark"],"summary":{"zh":"IndicBankBench 用 799 个印度零售银行案例，分阶段评估银行助手的安全、工具调用和回答质量。","en":"IndicBankBench evaluates Indian retail-banking assistants on 799 cases across safety, tool use and answer quality."},"disciplines":["finance"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":300},{"id":"arxiv-2610.19686","arxiv":"2610.19686","title":"Interlock: From Local Correctness to Task Completion in Long-Horizon Agents","authors":["Linji Hao","Yanji He","Zihan Wang","Hansi Yang","Jiayuan Zhang","Hongyan Hao","Qi Gu","Yan Feng"],"authorCount":9,"org":"LongCat","affiliations":[],"orgNote":null,"date":"2026-10-14","hfDaily":"2026-10-20","primaryCategory":"cs.CL","categories":["cs.CL","cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.19686","pdf":"https://arxiv.org/pdf/2610.19686","hf":"https://huggingface.co/papers/2610.19686","code":null,"project":null},"metrics":{"hfUpvotes":34,"hfComments":3,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":37.63,"tags":["agent","benchmark"],"summary":{"zh":"Interlock 把任务表示为“完成图”，合成能检查各环节是否真正闭环的长程智能体任务和评分规则。","en":"Interlock represents tasks as typed completion graphs to synthesize long-horizon agent tasks and rubrics that check end-to-end completion."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":301},{"id":"arxiv-2610.28632","arxiv":"2610.28632","title":"Joints, Camera, Action! Co-Generating 3D Motion and Video for Arbitrary Rigs","authors":["Noam Rotstein","Yoad Tewel","Chen Tessler","Haggai Maron","Gal Chechik","Or Litany","Yoni Kasten","Yuval Atzmon"],"authorCount":8,"org":"NVIDIA","affiliations":[],"orgNote":null,"date":"2026-10-20","hfDaily":"2026-10-21","primaryCategory":"cs.CV","categories":["cs.CV","cs.AI","cs.GR","cs.LG","cs.RO"],"links":{"abs":"https://arxiv.org/abs/2610.28632","pdf":"https://arxiv.org/pdf/2610.28632","hf":"https://huggingface.co/papers/2610.28632","code":null,"project":"https://research.nvidia.com/labs/par/projects/jca/"},"metrics":{"hfUpvotes":22,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":37,"tags":["visual-gen","vision"],"summary":{"zh":"NVIDIA 提出给任意骨骼同时生成视频和可回放的 3D 动作及相机，把视频里的动作直接用于 3D 资产。","en":"Co-generates video, replayable 3D motion for any rig and a registering camera, turning video-model motion into usable 3D animation."},"disciplines":["film-animation"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":302},{"id":"arxiv-2610.23751","arxiv":"2610.23751","title":"MTEdit: Packed Trajectories Beat Rollouts for Multi-Turn Image Editing","authors":["Houyuan Chen","Fanzhou Wang","Xiangyu Fan","Songlin Yang","Zhitao Yang","Lvmin Zhang","Pei Yang","Hao Zhao"],"authorCount":11,"org":"SenseNova","affiliations":[],"orgNote":null,"date":"2026-10-16","hfDaily":"2026-10-19","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.23751","pdf":"https://arxiv.org/pdf/2610.23751","hf":"https://huggingface.co/papers/2610.23751","code":null,"project":null},"metrics":{"hfUpvotes":33,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":36.91,"tags":["visual-gen","benchmark"],"summary":{"zh":"MTEdit 用“打包轨迹画布”一次生成多个编辑状态，构建 50 万条多轮图像编辑数据和配套评测。","en":"MTEdit builds 500K multi-turn image-editing trajectories by generating several edit states in one packed canvas, plus a benchmark."},"disciplines":["economics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":303},{"id":"arxiv-2610.26834","arxiv":"2610.26834","title":"Kronecker Delta Rule for Long Context Linear Attention","authors":["Anamaria-Roberta Hartl","Levente Zólyomi","Michael List","Günter Klambauer","Sebastian Böck","Sepp Hochreiter"],"authorCount":6,"org":"NX-AI","affiliations":[],"orgNote":null,"date":"2026-10-19","hfDaily":"2026-10-20","primaryCategory":"cs.LG","categories":["cs.LG","cs.AI"],"links":{"abs":"https://arxiv.org/abs/2610.26834","pdf":"https://arxiv.org/pdf/2610.26834","hf":"https://huggingface.co/papers/2610.26834","code":null,"project":null},"metrics":{"hfUpvotes":28,"hfComments":2,"citations":null,"citationsSource":null,"githubStars":null},"hot":35.32,"tags":["efficiency","llm"],"summary":{"zh":"KronLSTM 提出克罗内克 Delta 规则，只沿单一方向修正记忆矩阵，让长上下文线性注意力更稳定。","en":"KronLSTM introduces a Kronecker delta rule that revises matrix memory along one direction, giving stable long-context linear attention that often beats DeltaNet."},"disciplines":["mathematics"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":304},{"id":"arxiv-2610.04781","arxiv":"2610.04781","title":"Super-Resolution in The Right Latent Space: A Frozen Vision-Foundation Substrate","authors":["Wanzhou Lei","Cuifeng Shen","Yanjin He","Maohua Li","Hua Yuan","Per-Olof Persson","Tao Lan","Kan Liu"],"authorCount":9,"org":"RTP-LLM","affiliations":[],"orgNote":null,"date":"2026-10-03","hfDaily":"2026-10-15","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.04781","pdf":"https://arxiv.org/pdf/2610.04781","hf":"https://huggingface.co/papers/2610.04781","code":null,"project":null},"metrics":{"hfUpvotes":26,"hfComments":1,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":34,"tags":["vision","visual-gen"],"summary":{"zh":"主张在冻结的视觉基础模型特征空间里做超分辨率，比在重建型潜空间或像素空间更合适。","en":"Argues real-world super-resolution works best in a frozen vision-foundation feature space rather than reconstruction latents or pixels."},"disciplines":["computer-science"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":305},{"id":"arxiv-2609.30467","arxiv":"2609.30467","title":"Where Does Retrieval-Based Open-Ended Evaluation Fail? Automatic Taxonomy Induction from Long-Form Medical Answer Factuality Verification","authors":["Heyuan Huang","Jirui Dai","Alexandra DeLucia","Sonal Joshi","Mahsa Yarmohammadi","Jie Gao","Bernal Jiménez Gutiérrez","Mark Dredze"],"authorCount":8,"org":"Center for Language and Speech Processing @ JHU","affiliations":[],"orgNote":null,"date":"2026-09-24","hfDaily":"2026-10-05","primaryCategory":"cs.CL","categories":["cs.CL","cs.IR"],"links":{"abs":"https://arxiv.org/abs/2609.30467","pdf":"https://arxiv.org/pdf/2609.30467","hf":"https://huggingface.co/papers/2609.30467","code":null,"project":null},"metrics":{"hfUpvotes":23,"hfComments":2,"citations":0,"citationsSource":"semantic_scholar","githubStars":null},"hot":33.43,"tags":["ai4science","benchmark"],"summary":{"zh":"为医疗长答案的事实核查建立错误分类体系，找出基于检索的评测在哪里失效。","en":"Builds taxonomies of where retrieval-based factuality checking of long medical answers fails."},"disciplines":["clinical-medicine"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":306},{"id":"arxiv-2610.26837","arxiv":"2610.26837","title":"SpecKCast: Spectral Kalman Forecasting for Diffusion Acceleration","authors":["Javad Rajabi","Hakki Can Karaimer","David B. Lindell","Babak Taati","Michael S. Brown"],"authorCount":5,"org":"Samsung Research","affiliations":[],"orgNote":null,"date":"2026-10-19","hfDaily":"2026-10-21","primaryCategory":"cs.CV","categories":["cs.CV"],"links":{"abs":"https://arxiv.org/abs/2610.26837","pdf":"https://arxiv.org/pdf/2610.26837","hf":"https://huggingface.co/papers/2610.26837","code":null,"project":"https://rajabi2001.github.io/speckcast"},"metrics":{"hfUpvotes":22,"hfComments":1,"citations":null,"citationsSource":null,"githubStars":null},"hot":32.39,"tags":["visual-gen","efficiency"],"summary":{"zh":"SpecKCast 在频域用卡尔曼滤波预测扩散 Transformer 输出，不确定性高时才真算一次，无需训练即可加速。","en":"SpecKCast forecasts DiT outputs in the frequency domain with Kalman filters, running the model only when uncertainty is high, for training-free speedup."},"disciplines":["film-animation"],"verified":{"arxivMeta":true,"source":"arXiv API + Hugging Face Daily Papers"},"rank":307}],"labPosts":[{"id":"lab-www-anthropic-com-research-claude-scorpion-toxin-binder-design","lab":"Anthropic","title":"How Claude designed protein binders against scorpion neurotoxins","date":"2026-10-20","url":"https://www.anthropic.com/research/claude-scorpion-toxin-binder-design","tags":["ai4science"],"summary":{"zh":"Anthropic 内部研究版 Claude 设计出针对蝎子神经毒素的小型结合蛋白。","en":"An internal research Claude designed minibinder proteins against scorpion neurotoxins."},"disciplines":["life-sciences","neuroscience"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-openai-com-index-low-resource-programming-languages","lab":"OpenAI","title":"Advancing coding in low‑resource programming languages","date":"2026-10-19","url":"https://openai.com/index/low-resource-programming-languages","tags":["code","benchmark"],"summary":{"zh":"OpenAI 介绍如何提升 ChatGPT 在小众编程语言上的表现，并发布 SQLite、COBOL 迁移等新评测。","en":"OpenAI on improving ChatGPT for low-resource programming languages, with new evals for SWE, SQLite and COBOL migration."},"disciplines":["software-engineering"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-research-google-blog-grounding-ai-agents-in-biological-reality-building-blocks-f","lab":"Google Research","title":"Grounding AI agents in biological reality: Building blocks for autonomous wet labs","date":"2026-10-16","url":"https://research.google/blog/grounding-ai-agents-in-biological-reality-building-blocks-for-autonomous-wet-labs/","tags":["ai4science","agent","robotics"],"summary":{"zh":"Google Research 介绍让 AI Agent 扎根真实生物实验、迈向自主湿实验室的基础组件。","en":"Google Research on building blocks that ground AI agents in biology toward autonomous wet labs."},"disciplines":["life-sciences"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-huggingface-co-blog-reflection-beam","lab":"Hugging Face","title":"Welcome Beam: a 501B open model you can own, customize, and scale","date":"2026-10-16","url":"https://huggingface.co/blog/reflection-beam","tags":["llm"],"summary":{"zh":"Hugging Face 上线 5010 亿参数开源模型 Beam。","en":"Hugging Face welcomes Beam, a 501B-parameter open model."},"disciplines":["computer-science"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-www-anthropic-com-research-ai-expectations-gap","lab":"Anthropic","title":"The AI expectations gap","date":"2026-10-16","url":"https://www.anthropic.com/research/ai-expectations-gap","tags":["llm"],"summary":{"zh":"Anthropic 用新调查数据分析各方对 AI 经济影响预期分歧的根源。","en":"Anthropic presents survey evidence on why people disagree about AI's economic impact."},"disciplines":["economics"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-www-anthropic-com-research-designing-efficient-quantum-circuits","lab":"Anthropic","title":"Designing more efficient quantum circuits with Claude","date":"2026-10-16","url":"https://www.anthropic.com/research/designing-efficient-quantum-circuits","tags":["ai4science","agent"],"summary":{"zh":"Claude 基本自主地设计量子电路，部分结果比已发表最优版本好一个数量级以上（厂商自述）。","en":"Claude largely autonomously designed quantum circuits, some improving on the best published by 10x or more (vendor claim)."},"disciplines":["physics"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-www-anthropic-com-research-expectations-for-the-future","lab":"Anthropic","title":"Working at the frontier shapes expectations for the future","date":"2026-10-16","url":"https://www.anthropic.com/research/expectations-for-the-future","tags":["llm"],"summary":{"zh":"Anthropic 发布两份报告：AI 如何改变公司内部就业，以及公众与员工对 AI 影响工作的预期差异。","en":"Two Anthropic reports on AI changing its internal employment and the public-vs-employee expectations gap."},"disciplines":["economics"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-www-anthropic-com-research-ecdsa-curve-origins","lab":"Anthropic","title":"Claude reveals ECDSA curve origins","date":"2026-10-16","url":"https://www.anthropic.com/research/ecdsa-curve-origins","tags":["safety","ai4science"],"summary":{"zh":"Anthropic 称 Claude 解开了一个流行密码学标准中 ECDSA 曲线来源的多年谜题（厂商自述）。","en":"Anthropic says Claude explained the decades-old mystery of a popular ECDSA curve's origins (vendor claim)."},"disciplines":["cybersecurity"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-huggingface-co-blog-nvidia-scaling-indic-pretraining","lab":"Hugging Face","title":"Scaling Indic Pretraining: 2.8 Trillion Tokens and an Open Data Generation Pipeline","date":"2026-10-16","url":"https://huggingface.co/blog/nvidia/scaling-indic-pretraining","tags":["llm"],"summary":{"zh":"Hugging Face 介绍 2.8 万亿 token 的印度语种预训练数据和开放数据生成流水线。","en":"Hugging Face on scaling Indic pretraining with 2.8T tokens and an open data-generation pipeline."},"disciplines":["computer-science"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-openai-com-index-introducing-the-rosalind-academic-plan","lab":"OpenAI","title":"Bringing GPT-Rosalind to academic research labs","date":"2026-10-15","url":"https://openai.com/index/introducing-the-rosalind-academic-plan","tags":["ai4science"],"summary":{"zh":"OpenAI 推出 Rosalind 学术计划，为合格的学术实验室提供 GPT-Rosalind 生命科学模型的赞助使用。","en":"OpenAI's Rosalind Academic Plan gives eligible labs sponsored access to GPT-Rosalind for life-sciences research."},"disciplines":["computer-science"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-deepmind-google-blog-accelerating-antimicrobial-discovery","lab":"Google DeepMind","title":"Accelerating the search for antimicrobial peptides with Google Antigravity","date":"2026-10-15","url":"https://deepmind.google/blog/accelerating-antimicrobial-discovery/","tags":["ai4science","agent"],"summary":{"zh":"Google DeepMind 用 Google Antigravity for Science 加速新型抗菌肽的搜索。","en":"Google DeepMind uses Google Antigravity for Science to accelerate the search for new antimicrobial peptides."},"disciplines":["life-sciences"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-www-microsoft-com-en-us-research-blog-proof-based-general-purpose-computing-with","lab":"Microsoft Research","title":"Proof-based general-purpose computing with Magnetar","date":"2026-10-15","url":"https://www.microsoft.com/en-us/research/blog/proof-based-general-purpose-computing-with-magnetar/","tags":["safety","efficiency"],"summary":{"zh":"微软研究院 Magnetar 让零知识证明支持标准编程语言并大幅提速，可在不重跑任务的情况下验证计算结果。","en":"Microsoft Research's Magnetar brings zero-knowledge proofs to standard languages with big speed-ups for verifiable computing."},"disciplines":["mathematics","physics","software-engineering"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-huggingface-co-blog-meta-muse-spark-open","lab":"Hugging Face","title":"Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta","date":"2026-10-15","url":"https://huggingface.co/blog/meta-muse-spark-open","tags":["multimodal","llm"],"summary":{"zh":"Meta 的开源权重前沿多模态 MoE 模型 Muse Spark Open 上线 Hugging Face。","en":"Meta's open-weight frontier multimodal MoE model Muse Spark Open arrives on Hugging Face."},"disciplines":["computer-science"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-research-google-blog-putting-earth-ai-to-work-geospatial-agents-for-our-planets-","lab":"Google Research","title":"Putting Earth AI to work: Geospatial agents for our planet's biggest challenges","date":"2026-10-14","url":"https://research.google/blog/putting-earth-ai-to-work-geospatial-agents-for-our-planets-biggest-challenges/","tags":["ai4science","agent"],"summary":{"zh":"Google Research 的 Earth AI 地理空间 Agent，用于应对全球性重大问题。","en":"Google Research puts Earth AI geospatial agents to work on planetary-scale challenges."},"disciplines":["geography-remote-sensing"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-research-google-blog-analyzing-technical-documents-with-deep-review","lab":"Google Research","title":"Analyzing technical documents with Deep Review","date":"2026-10-14","url":"https://research.google/blog/analyzing-technical-documents-with-deep-review/","tags":["agent"],"summary":{"zh":"Google Research 介绍用于分析技术文档的 Deep Review。","en":"Google Research introduces Deep Review for analysing technical documents."},"disciplines":["computer-science"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-huggingface-co-blog-allenai-hiro-acev2","lab":"Hugging Face","title":"HiRO-ACEv2: Simulating extreme weather at 3 km resolution for a fraction of the cost","date":"2026-10-14","url":"https://huggingface.co/blog/allenai/hiro-acev2","tags":["ai4science"],"summary":{"zh":"HiRO-ACEv2 以较低成本在 3 公里分辨率下模拟极端天气。","en":"HiRO-ACEv2 simulates extreme weather at 3 km resolution at a fraction of the cost."},"disciplines":["weather-climate"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-deepmind-google-blog-developing-gemini-4-argon-responsibly-and-safely","lab":"Google DeepMind","title":"Developing Gemini 4 Argon responsibly and safely","date":"2026-10-13","url":"https://deepmind.google/blog/developing-gemini-4-argon-responsibly-and-safely/","tags":["safety"],"summary":{"zh":"Google DeepMind 介绍 Gemini 4 Argon 的负责任开发与安全工作。","en":"Google DeepMind on developing Gemini 4 Argon responsibly and safely."},"disciplines":["computer-science"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-deepmind-google-blog-gemini-4-argon-for-robotics","lab":"Google DeepMind","title":"Gemini 4 Argon for Robotics","date":"2026-10-13","url":"https://deepmind.google/blog/gemini-4-argon-for-robotics/","tags":["robotics"],"summary":{"zh":"Google DeepMind 发布面向机器人的 Gemini 4 Argon。","en":"Google DeepMind introduces Gemini 4 Argon for Robotics."},"disciplines":["robotics"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-deepmind-google-blog-introducing-gemini-4-argon","lab":"Google DeepMind","title":"Gemini 4 Argon: Built to solve your hardest problems","date":"2026-10-13","url":"https://deepmind.google/blog/introducing-gemini-4-argon/","tags":["llm","reasoning"],"summary":{"zh":"Google DeepMind 介绍 Gemini 4 Argon 处理高难度问题的能力。","en":"Google DeepMind presents Gemini 4 Argon as built for the hardest problems."},"disciplines":["computer-science"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-www-microsoft-com-en-us-research-blog-simulating-student-learning-for-better-ai-","lab":"Microsoft Research","title":"Simulating student learning for better AI tutoring","date":"2026-10-12","url":"https://www.microsoft.com/en-us/research/blog/simulating-student-learning-for-better-ai-tutoring/","tags":["llm","agent"],"summary":{"zh":"微软 StudentSim 模拟学生学习过程，在真人测试前评估和改进 AI 辅导工具。","en":"Microsoft's StudentSim simulates learners to evaluate and improve AI tutors before real-student trials."},"disciplines":["mathematics","education"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-www-anthropic-com-research-self-replicating-ai","lab":"Anthropic","title":"AI models may be capable of self-replicating without access to a datacenter","date":"2026-10-12","url":"https://www.anthropic.com/research/self-replicating-ai","tags":["safety"],"summary":{"zh":"Anthropic 研究称小模型可能数月内就能在廉价物联网设备上自我复制，无需数据中心。","en":"Anthropic research suggests small models may soon self-replicate across cheap IoT devices without a datacenter."},"disciplines":["computer-science"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-www-anthropic-com-research-distillation-defenses","lab":"Anthropic","title":"Strengthening our defenses against distillation attacks","date":"2026-10-12","url":"https://www.anthropic.com/research/distillation-defenses","tags":["safety"],"summary":{"zh":"Anthropic 介绍防御非法蒸馏的做法，称已将蒸馏发生率降低 90% 以上。","en":"Anthropic describes defences against illicit distillation, reporting a >90% reduction in distillation rate."},"disciplines":["cybersecurity"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-www-anthropic-com-research-investigating-unintended-model-actions","lab":"Anthropic","title":"Investigating unintended model actions in our evaluations and internal use","date":"2026-10-09","url":"https://www.anthropic.com/research/investigating-unintended-model-actions","tags":["safety","alignment"],"summary":{"zh":"Anthropic 调查模型在评测和内部使用中出现的非预期行为。","en":"Anthropic investigates unintended model actions seen in its evaluations and internal use."},"disciplines":["computer-science"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-openai-com-index-disrupting-ai-enabled-false-front-operations","lab":"OpenAI","title":"Disrupting AI-enabled “false front” operations","date":"2026-10-08","url":"https://openai.com/index/disrupting-ai-enabled-false-front-operations","tags":["safety"],"summary":{"zh":"OpenAI 披露并阻断两起借 AI 冒充记者和智库传播地缘政治信息的影响力行动。","en":"OpenAI reports disrupting two AI-enabled influence operations using fake journalists and a think tank."},"disciplines":["computer-science"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-www-anthropic-com-research-the-missing-map-of-the-sky","lab":"Anthropic","title":"Using Claude Science to produce the first complete map of the sky in UV light","date":"2026-10-08","url":"https://www.anthropic.com/research/the-missing-map-of-the-sky","tags":["ai4science"],"summary":{"zh":"Claude Science 协助制作了首张完整的紫外波段全天图。","en":"Claude Science helped produce the first complete map of the sky in UV light."},"disciplines":["astronomy"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-www-microsoft-com-en-us-research-blog-agent-lightning-v1-0-a-3500-line-lightweig","lab":"Microsoft Research","title":"Agent Lightning v1.0: A 3,500-Line Lightweight Agentic RL Framework for Training Agents with Real Harnesses","date":"2026-10-07","url":"https://www.microsoft.com/en-us/research/blog/agent-lightning-v1-0-a-3500-line-lightweight-agentic-rl-framework-for-training-agents-with-real-harnesses/","tags":["agent","rl"],"summary":{"zh":"微软 Agent Lightning v1.0：约 3500 行的轻量 Agent RL 框架，可直接接入现有 Agent 训练。","en":"Microsoft's Agent Lightning v1.0 is a ~3,500-line agentic RL framework that plugs existing agents into RL training."},"disciplines":["computer-science"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-openai-com-index-sharing-ai-progress-in-mathematics","lab":"OpenAI","title":"Sharing AI progress in mathematics","date":"2026-10-06","url":"https://openai.com/index/sharing-ai-progress-in-mathematics","tags":["reasoning","ai4science"],"summary":{"zh":"OpenAI 发布内部前沿模型在数学开放问题上的新结果，并在 GitHub 公开 Lean 形式化证明（厂商自述，存在争议）。","en":"OpenAI shares new results on open maths problems from an internal model, with Lean formalisations on GitHub (vendor claim; disputed)."},"disciplines":["mathematics"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-openai-com-index-advancing-computer-use-with-ironclad","lab":"OpenAI","title":"Advancing computer use with Ironclad","date":"2026-10-06","url":"https://openai.com/index/advancing-computer-use-with-ironclad","tags":["agent"],"summary":{"zh":"OpenAI 与 Ironclad 合作，在复杂合同工作流上训练和评估电脑操作 Agent。","en":"OpenAI and Ironclad train and evaluate computer-use agents on complex contracting workflows."},"disciplines":["computer-science"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-deepmind-google-blog-embeddinggemma-2-an-open-lightweight-multimodal-embedding-m","lab":"Google DeepMind","title":"EmbeddingGemma 2: an open, lightweight multimodal embedding model","date":"2026-10-06","url":"https://deepmind.google/blog/embeddinggemma-2-an-open-lightweight-multimodal-embedding-model/","tags":["multimodal","efficiency"],"summary":{"zh":"Google DeepMind 发布开源轻量多模态向量模型 EmbeddingGemma 2。","en":"Google DeepMind releases EmbeddingGemma 2, an open lightweight multimodal embedding model."},"disciplines":["computer-science"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-openai-com-index-eu-text-provenance","lab":"OpenAI","title":"Our approach to EU text provenance rules","date":"2026-10-05","url":"https://openai.com/index/eu-text-provenance","tags":["safety"],"summary":{"zh":"OpenAI 说明在欧盟文本溯源规则下的文本水印做法：适用范围、检测方式和研究者优先开放。","en":"OpenAI explains its text watermarking approach under EU provenance rules, including scope, detection and researcher access."},"disciplines":["computer-science"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-research-google-blog-open-and-emergent-problems-in-agentic-privacy-and-security-","lab":"Google Research","title":"Open and Emergent Problems in Agentic Privacy and Security: A Contextual Angle","date":"2026-10-05","url":"https://research.google/blog/open-and-emergent-problems-in-agentic-privacy-and-security-a-contextual-angle/","tags":["safety","agent"],"summary":{"zh":"Google Research 从“情境完整性”角度梳理 Agent 隐私与安全的开放问题。","en":"Google Research outlines open problems in agentic privacy and security from a contextual-integrity angle."},"disciplines":["cybersecurity"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-research-google-blog-toward-provably-private-learning-from-federated-data","lab":"Google Research","title":"Toward provably private learning from federated data","date":"2026-10-02","url":"https://research.google/blog/toward-provably-private-learning-from-federated-data/","tags":["safety"],"summary":{"zh":"Google Research 介绍面向联邦数据的可证明隐私学习进展。","en":"Google Research on provably private learning from federated data."},"disciplines":["computer-science"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-openai-com-index-disrupting-a-coordinated-model-distillation-campaign","lab":"OpenAI","title":"Disrupting a coordinated model-distillation campaign","date":"2026-09-30","url":"https://openai.com/index/disrupting-a-coordinated-model-distillation-campaign","tags":["safety"],"summary":{"zh":"OpenAI 披露阻断一起提取受保护推理内容的协同蒸馏行动，并加强对抗性蒸馏防御。","en":"OpenAI describes disrupting a coordinated campaign to distil protected model reasoning and strengthening defences."},"disciplines":["computer-science"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-deepmind-google-blog-gemini-4-argon-our-next-era-of-frontier-intelligence","lab":"Google DeepMind","title":"Gemini 4 Argon: our next era of frontier intelligence","date":"2026-09-30","url":"https://deepmind.google/blog/gemini-4-argon-our-next-era-of-frontier-intelligence/","tags":["llm"],"summary":{"zh":"Google DeepMind 发布新一代前沿模型 Gemini 4 Argon。","en":"Google DeepMind announces Gemini 4 Argon, its next-generation frontier model."},"disciplines":["computer-science"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-deepmind-google-blog-introducing-synthid-bio","lab":"Google DeepMind","title":"Introducing SynthID Bio","date":"2026-09-30","url":"https://deepmind.google/blog/introducing-synthid-bio/","tags":["safety","ai4science"],"summary":{"zh":"SynthID Bio 概念验证：给 AI 生成的蛋白质加水印，同时保持其生物功能。","en":"SynthID Bio is a proof of concept for watermarking AI-generated proteins while keeping their function."},"disciplines":["mathematics","life-sciences"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-www-microsoft-com-en-us-research-blog-forecasting-space-weather-risks-on-power-g","lab":"Microsoft Research","title":"Forecasting space weather risks on power grids","date":"2026-09-30","url":"https://www.microsoft.com/en-us/research/blog/forecasting-space-weather-risks-on-power-grids/","tags":["ai4science"],"summary":{"zh":"微软研究院的机器学习系统可在磁暴到达前 30–60 分钟预测电网可能受损的位置。","en":"A Microsoft Research ML system predicts where space weather may damage power grids 30–60 minutes ahead."},"disciplines":["weather-climate"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-huggingface-co-blog-open-tts-leaderboard","lab":"Hugging Face","title":"Open TTS Leaderboard: Scalable Evaluation for Multilingual Text-to-Speech and Voice Cloning","date":"2026-09-30","url":"https://huggingface.co/blog/open-tts-leaderboard","tags":["multimodal","benchmark"],"summary":{"zh":"Hugging Face 推出开放 TTS 排行榜，评测多语种语音合成和声音克隆。","en":"Hugging Face's Open TTS Leaderboard evaluates multilingual text-to-speech and voice cloning."},"disciplines":["computer-science"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-research-google-blog-how-diffusion-controller-unifies-and-simplifies-ai-image-ge","lab":"Google Research","title":"How Diffusion Controller unifies and simplifies AI image generation","date":"2026-09-29","url":"https://research.google/blog/how-diffusion-controller-unifies-and-simplifies-ai-image-generation/","tags":["visual-gen"],"summary":{"zh":"Google Research 的 Diffusion Controller 统一并简化 AI 图像生成的控制方式。","en":"Google Research's Diffusion Controller unifies and simplifies control in AI image generation."},"disciplines":["computer-science"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-www-microsoft-com-en-us-research-blog-introducing-quine-an-ai-research-system-de","lab":"Microsoft Research","title":"Introducing Quine: An AI research system designed for the complexity of biology","date":"2026-09-29","url":"https://www.microsoft.com/en-us/research/blog/introducing-quine-an-ai-research-system-designed-for-the-complexity-of-biology/","tags":["ai4science"],"summary":{"zh":"微软 Quine：早期研究项目，尝试构建跨尺度的多模态生物学世界模型。","en":"Microsoft's Quine is an early effort to build a cross-scale multimodal world model of biology."},"disciplines":["life-sciences"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-openai-com-index-towards-safety-cases-for-frontier-ai-training","lab":"OpenAI","title":"Towards safety cases for frontier AI training","date":"2026-09-28","url":"https://openai.com/index/towards-safety-cases-for-frontier-ai-training","tags":["safety","alignment"],"summary":{"zh":"OpenAI 提出前沿模型训练“安全论证”的初步指南，涵盖技术防护、运营实践和失准事件调查。","en":"OpenAI's early guidelines for safety cases in frontier AI training: safeguards, operations and misalignment investigations."},"disciplines":["computer-science"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-huggingface-co-blog-rl-environments","lab":"Hugging Face","title":"Welcome RL Environments to the hub","date":"2026-09-28","url":"https://huggingface.co/blog/rl-environments","tags":["rl","agent"],"summary":{"zh":"Hugging Face Hub 开始托管强化学习环境。","en":"Hugging Face Hub now hosts RL environments."},"disciplines":["computer-science"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-research-google-blog-coherent-long-form-video-generation","lab":"Google Research","title":"Automating coherent long-form video generation","date":"2026-09-24","url":"https://research.google/blog/coherent-long-form-video-generation/","tags":["visual-gen"],"summary":{"zh":"Google Research 介绍自动生成连贯长视频的方法。","en":"Google Research on automating coherent long-form video generation."},"disciplines":["film-animation"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-www-anthropic-com-research-project-swap","lab":"Anthropic","title":"Project Swap: What happens when agents trade for us?","date":"2026-09-24","url":"https://www.anthropic.com/research/project-swap","tags":["agent"],"summary":{"zh":"Project Swap：Anthropic 搭建一个由 Claude 组成的迷你市场，观察 Agent 代人交易时哪里有效、哪里出错。","en":"Project Swap builds a miniature market of Claudes to see what works and breaks when agents trade for us."},"disciplines":["finance"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-openai-com-index-introducing-mentalhealthbench","lab":"OpenAI","title":"Introducing MentalHealthBench","date":"2026-09-23","url":"https://openai.com/index/introducing-mentalhealthbench","tags":["safety","benchmark"],"summary":{"zh":"OpenAI 发布 MentalHealthBench，在真实心理健康对话中评估 AI 回复是否有用且安全。","en":"OpenAI's MentalHealthBench evaluates whether AI responses are helpful and safe in realistic mental-health conversations."},"disciplines":["clinical-medicine","psychology"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-deepmind-google-blog-advancing-private-ai-compute-with-secure-server-side-memory","lab":"Google DeepMind","title":"Advancing Private AI Compute with secure, server-side memory","date":"2026-09-23","url":"https://deepmind.google/blog/advancing-private-ai-compute-with-secure-server-side-memory/","tags":["safety"],"summary":{"zh":"Google 为 Private AI Compute 加入安全的服务端记忆，用于个人 AI 场景。","en":"Google adds secure server-side memory to Private AI Compute for personal AI."},"disciplines":["computer-science"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-www-microsoft-com-en-us-research-blog-offloaded-inference-for-real-world-physica","lab":"Microsoft Research","title":"Offloaded inference for real-world physical AI robotics","date":"2026-09-23","url":"https://www.microsoft.com/en-us/research/blog/offloaded-inference-for-real-world-physical-ai-robotics/","tags":["robotics","efficiency"],"summary":{"zh":"微软研究院发现把 AI 推理从机器人本体卸载到外部，可提升任务成功率和效率。","en":"Microsoft Research finds offloading robot AI inference off-board improves task success and efficiency."},"disciplines":["robotics"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-openai-com-index-priorities-principles-third-party-assessments","lab":"OpenAI","title":"Priorities and principles for effective third party assessments","date":"2026-09-22","url":"https://openai.com/index/priorities-principles-third-party-assessments","tags":["safety"],"summary":{"zh":"OpenAI 阐述对前沿模型开展独立第三方安全评估的优先事项与原则。","en":"OpenAI sets out priorities and principles for rigorous, independent third-party safety assessments of frontier models."},"disciplines":["computer-science"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"},{"id":"lab-huggingface-co-blog-evaleval-aisi","lab":"Hugging Face","title":"How UK AISI and EvalEval Are Making Benchmark Results Reproducible","date":"2026-09-22","url":"https://huggingface.co/blog/evaleval-aisi","tags":["benchmark"],"summary":{"zh":"英国 AISI 与 EvalEval 合作，让模型评测结果可复现。","en":"UK AISI and EvalEval on making benchmark results reproducible."},"disciplines":["computer-science"],"kind":"lab-post","note":"实验室官方博客/技术报告；“厂商自述”类结论未经同行评审"}],"weekly":{"schema":"aimap.ai-papers.weekly/v1","week":"2026-W43","from":"2026-10-19","to":"2026-10-25","status":"in-progress","dataAsOf":"2026-10-21","generatedAt":"2026-10-22T10:04:53+08:00","rule":{"zh":"按 HF Daily 上榜日所在 ISO 周，取已写要点论文中 HF 点赞前 10；完整周冻结后不再改动。","en":"Top 10 by HF upvotes among summarised papers, grouped by ISO week of HF Daily listing; frozen once the week is complete."},"items":[{"rank":1,"id":"arxiv-2610.21046","arxiv":"2610.21046","title":"PremiseCheck: Efficient Fact-Checking of Human and AI Peer Reviews","hfUpvotes":727,"tags":["benchmark","efficiency"],"summary":{"zh":"ReviewFact 标注审稿意见中的事实错误，PremiseCheck 结合大模型和轻量判别模型低成本核查。","en":"ReviewFact labels factual errors in peer reviews; PremiseCheck combines an LLM with cheap decision models to verify them efficiently."},"org":null,"date":"2026-10-15"},{"rank":2,"id":"arxiv-2610.20951","arxiv":"2610.20951","title":"LEV: Thinking Fast and Deep with Looped Decision Models","hfUpvotes":602,"tags":["reasoning","efficiency"],"summary":{"zh":"LEV 让决策模型通过循环重跑网络层来“想得更深”，不必生成推理文字。","en":"LEV lets decision models think deeper by looping their layers instead of generating reasoning tokens."},"org":null,"date":"2026-10-14"},{"rank":3,"id":"arxiv-2610.20424","arxiv":"2610.20424","title":"BizAgentBench: Unveiling Hidden Challenges in Deploying Long-Horizon Agents for Real-World Business","hfUpvotes":490,"tags":["agent","benchmark"],"summary":{"zh":"BizBench 基于真实支付业务流程，检验长任务 Agent 在执行中途是否违反政策与约束。","en":"BizBench, built on real payment procedures, checks whether long-horizon agents violate policies and constraints mid-execution."},"org":"ant-international","date":"2026-10-14"},{"rank":4,"id":"arxiv-2610.22471","arxiv":"2610.22471","title":"Has GPT-6-Astra Achieved Embodied Intelligence?","hfUpvotes":459,"tags":["robotics","benchmark"],"summary":{"zh":"在 12 个机器人基准上系统评测 GPT-6-Astra 的具身能力，并讨论通用大模型对机器人研究的影响。","en":"Evaluates GPT-6-Astra on 12 robotics benchmarks and discusses what a general model's embodied competence means for the field."},"org":"SenseNova","date":"2026-10-15"},{"rank":5,"id":"arxiv-2610.23479","arxiv":"2610.23479","title":"HexaShell: A Thin Harness for Exploring the Limits of VLM Agents as Robot Coding Agents","hfUpvotes":313,"tags":["robotics","agent","code"],"summary":{"zh":"HexaShell 只用一层很薄的框架，让编程 Agent 通过命令行直接控制双臂机器人，并在失败后自己写工具改进。","en":"HexaShell keeps the harness thin so a coding agent drives a two-arm robot via a command line and writes its own tools after failures."},"org":"HexaFuture","date":"2026-10-16"},{"rank":6,"id":"arxiv-2610.24087","arxiv":"2610.24087","title":"OR-License: Evaluating LLM Agents as Real-World Operations Research Practitioners","hfUpvotes":291,"tags":["benchmark","agent"],"summary":{"zh":"OR-License 用 31 个真实付费运筹咨询项目的原始材料，测试大模型智能体能否直接交付专家级运筹方案。","en":"OR-License benchmarks whether LLM agents can turn original client materials from 31 paid operations-research engagements into expert-grade solutions."},"org":"Data Science & Artificial Intelligence Lab at HKUST","date":"2026-10-16"},{"rank":7,"id":"arxiv-2610.26802","arxiv":"2610.26802","title":"Scaling Parameter-Efficient Fine-Tuning for Brain Encoding","hfUpvotes":237,"tags":["ai4science","efficiency"],"summary":{"zh":"用参数高效微调把最大 104B 的多模态模型适配到脑编码任务，更准确地预测大脑对刺激的反应。","en":"Uses parameter-efficient fine-tuning to adapt multimodal backbones of up to 104B for brain encoding, better predicting neural responses."},"org":"Mind Lab","date":"2026-10-19"},{"rank":8,"id":"arxiv-2610.26203","arxiv":"2610.26203","title":"Focused Omni-Modal GUI Agents","hfUpvotes":225,"tags":["agent","multimodal"],"summary":{"zh":"FocusGUI 用轻量适配器按步骤挑出最相关的屏幕、音频等信号注入冻结模型，缓解 GUI 智能体被无关上下文带偏。","en":"FocusGUI is a lightweight adapter that picks the step-relevant omni-modal evidence for a frozen GUI agent, curbing compounding distraction."},"org":"Nanyang Technological University","date":"2026-10-19"},{"rank":9,"id":"arxiv-2610.27999","arxiv":"2610.27999","title":"Matrix-Game 4.0: Harnessing Video Models for Real-Time Playable Game Worlds","hfUpvotes":182,"tags":["visual-gen","agent"],"summary":{"zh":"Matrix-Game 4.0 让轻量引擎和编程智能体管游戏逻辑、视频模型管画面，单张 RTX 5090 跑出 60 帧。","en":"Matrix-Game 4.0 separates game logic (engine plus coding agent) from visuals (autoregressive video renderer), reaching playable 60 FPS on one RTX 5090."},"org":null,"date":"2026-10-20"},{"rank":10,"id":"arxiv-2610.26855","arxiv":"2610.26855","title":"Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human Experience","hfUpvotes":179,"tags":["robotics","visual-gen"],"summary":{"zh":"Ropedia W0 是跨人类与多种机器人的第一视角移动操作世界基础模型，用 1 万小时人类、3 千小时机器人数据训练。","en":"Ropedia W0 is an egocentric loco-manipulation world foundation model spanning humans and robots, trained on 10k hours of human and 3k hours of robot data."},"org":"Ropedia","date":"2026-10-19"}]}}