[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-a-rsi-zero-data-self-improvement":3,"topics-all":38,"news-related-23567889-4d40-4af8-a58a-ca61cdb6a8b2":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"23567889-4d40-4af8-a58a-ca61cdb6a8b2","小米A-RSI:不用外部标注,模型自己发现题库刷完了","小米研究院与特伦托大学等发布 A-RSI:递归自我改进不依赖外部标注,只读训练日志监控奖励与准确率,学习信号衰减就整池换新题,由程序库生成多模态题目并算出可验证答案。Qwen3-VL-8B 八个窗口只用 800 道新题,把十一项基准均分从 57.77 提到 61.05;审计坦承部分涨分来自学会收尾。","递归自我改进的循环,最先卡住的地方往往不是\"练什么\",而是\"旧经验什么时候失效\"。10 月 28 日挂上 arXiv 的 A-RSI(小米研究院与特伦托大学等合作)给了一套工程化的答案:让模型盯自己的训练日志,发现题池刷\"枯\"了就整池换新,全程不用外部标注。论文两天冲上 Hugging Face 每日论文榜第 2,截至发稿 155 个点赞。\n\n## 两个被忽视的老问题\n\n现有自我改进工作大多在解决\"生成什么经验\":造更难的任务、筛更好的轨迹。A-RSI 点出两个薄弱环节:一,监督信号常靠自洽性或多数投票构造,等于让模型拿自己都不确定的预测当老师;二,几乎没人回答\"当前经验何时耗尽\"——题池刷到没东西可学,训练就开始空转。\n\n## 训练日志当裁判,程序当出题人\n\nA-RSI 把两件事拆开。**换题时机**:只读模型自己的训练日志,监控奖励方差与准确率,信号留存跌破 85% 阈值就整池换新——400 题一池,按固定家族配额抽取(194 几何、131 图表、54 表格、21 计数),全程不看验证集或测试集分数。**监督来源**:一个静态的\"模型写程序\"库生成多模态题目并直接算出参考答案,答案可独立校验,不依赖模型投票。刷新规则本身经过形式化验证,32 万余个阈值与初始化用例全部匹配。\n\n## 800 道新题,涨 3.27 分\n\n主实验基于 Qwen3-VL-8B-Instruct:8 个续训窗口、累计 800 道新题,把 11 项基准的宏平均从 57.77 拉到 61.05,22,042 道测试题上所有单项全部上涨(+1.19 到 +8.33)。对照组更说明问题:老题库一直刷,到第 6 窗口只有 58.35,随后停滞;每窗全换烧掉 3,200 道题才到 60.64——A-RSI 用四分之一的新题量反超 0.40 分,每千题收益 4.09 分对 0.90 分。在小米自家 MiMo-VL-7B-SFT 上,分数从 59.08 提到 60.71,换题同样在第 3 窗自动触发。每个续训窗口包含 100 次 GRPO 更新,主轨迹与各对照臂共享同一套优化设置。\n\n## 最值钱的是\"自己拆自己\"的审计\n\n作者主动拆穿了自己收益的一部分来源:截断率从 14.34% 降到 4.16%,相当一部分涨分其实是模型学会在预算内把答案写完、正常收尾,而不是视觉推理真的变强;换成格式校正重打分,增益仍保住 +3.27。再用 Qwen2.5-14B-Instruct 当独立判官重评,判官口径准确率从 54.83 升到 59.54;把输出预算从 4096 翻倍到 8192 重测,对基座的增益仍有 3.25 分,高于先前零数据方法 MM-Zero v5 的 2.45 分——结论换几个口径都站得住。他们还承认:难度、控制器、rollout 三种选题器都没跑赢随机抽取;单次运行下,换题时机的优势结论并不成立。项目页代码标注\"即将放出\"。\n\n所以呢:零数据自提升的瓶颈,正在从\"怎么造题\"移向\"何时换题\"和\"监督独立性\"。对想在自家业务上跑自我改进循环的团队,这篇的可复用资产不是 61.05 这个数字,而是\"先形式化验证、再独立审计、最后承认局限\"的动作——先抄审计,再抄方法。\n\n论文: https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.38568 ;Hugging Face 页面: https:\u002F\u002Fhuggingface.co\u002Fpapers\u002F2610.38568","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.38568","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"c4651eed-32a1-438b-ad4f-8b20508971c4","en","Xiaomi A-RSI: Zero-Data Self-Improvement, Auto Task Refresh","A-RSI refreshes task pools from training-log signals; program-computed answers replace labels. Qwen3-VL-8B: 57.77 to 61.05 with 800 fresh tasks.","Recursive self-improvement loops usually stall on a question nobody optimizes for: when has the current experience run dry? A-RSI, posted to arXiv on October 28 by Xiaomi Research with the University of Trento and others, offers an engineering-grade answer — watch the model's own training log, and swap the whole task pool when the learning signal fades. No external labels anywhere in the loop. The paper reached #2 on Hugging Face Daily Papers, with 155 upvotes at writing time.\n\n## Two old problems\n\nMost self-improvement work focuses on what experience to generate next. A-RSI highlights two gaps: supervision is often built from self-consistency or majority voting — the model grading its own uncertain guesses — and almost nothing decides when fresh experience should be introduced, so training spins on a spent pool.\n\n## Training log as referee, programs as question writers\n\nA-RSI separates the two. **Refresh timing**: it monitors reward variation and accuracy in training logs only; when signal retention drops below an 85% threshold, the entire 400-task pool is replaced, drawn under fixed family quotas (194 geometry, 131 chart, 54 table, 21 counting). No validation or test scores are consulted. **Supervision**: a static library of model-authored programs generates multimodal questions and computes reference answers that are independently checkable — no voting. The refresh rule itself is formally verified across 321,605 threshold and initialization cases.\n\n## 800 fresh tasks, +3.27 points\n\nOn Qwen3-VL-8B-Instruct, eight continuation windows and 800 fresh tasks lift the Macro11 average from 57.77 to 61.05, with all 22,042 test items improving (+1.19 to +8.33 per benchmark). The controls tell the story: continuing on the old pool plateaus at 58.35 by window 6; refreshing every window burns 3,200 tasks to reach 60.64. A-RSI beats that by 0.40 with a quarter of the tasks — 4.09 points gained per thousand tasks versus 0.90. On Xiaomi's MiMo-VL-7B-SFT, scores rise from 59.08 to 60.71, and the detector fires at window 3, as on Qwen.\n\n## The honest audit is the best part\n\nThe authors puncture part of their own gain: truncation falls from 14.34% to 4.16%, so a chunk of the improvement is the model learning to finish within budget rather than genuinely stronger visual reasoning — format-corrected rescoring still preserves +3.27. An independent Qwen2.5-14B-Instruct judge puts accuracy at 59.54 versus 54.83 at the start, and doubling the output budget from 4096 to 8192 tokens leaves a 3.25-point gain over base, ahead of the prior zero-data method MM-Zero v5 at 2.45. Each continuation window runs 100 GRPO updates. They also report that difficulty, controller, and rollout-based selectors fail to beat random draws, and that single-run margins do not establish a timing advantage. The project page lists code as coming soon.\n\nThe bottleneck for zero-data self-improvement is moving from generating tasks to knowing when to refresh and keeping supervision independent. For teams running self-improvement loops on their own stacks, the reusable asset is the sequence — formal verification, independent audits, stated limits — not the 61.05 number.\n\nPaper: https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.38568 ; Hugging Face: https:\u002F\u002Fhuggingface.co\u002Fpapers\u002F2610.38568","a-rsi-zero-data-self-improvement","2026-10-30T19:09:58Z","2026-10-30T19:10:00.675255Z","2026-10-30T19:10:00.675265Z",true,"agent",81,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"bd050dd6-4d85-4616-a004-55c23c533a24","腾讯混元合并大语言模型与多模态团队，成立基础模型部探索全模态统一","tencent-hunyuan-foundation-model-dept","2026-07-24T03:00:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"63506aeb-566f-47dc-a59f-d4f0197e4545","面壁 MiniCPM 拿下三星旗舰端:7 款端侧 LLM 同日备案,标志端侧 AI 进入合规量产节奏","minicpm-samsung-end-side","2026-07-15T14:01:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"098ef0da-3801-419a-9e33-855d6c6d7ee6","Trinity:递归自我改进,训练基准平均涨 34 分","trinity-capability-driven-recursive-self-improvement","2026-10-28T13:16:03+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"5480e0ca-7b17-481c-8a24-8498f1a0aecd","8B模型OST涨13.41分:空间推理的接口锁定","chronoreplay-spatial-interface-lock-in","2026-10-28T05:09:43+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"edc863dc-1918-4fdf-8306-640d913e15e2","Google OPST:零人类习题,模型自出题自提升","google-opst-data-free-self-training","2026-10-26T21:10:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"f728cc32-7d43-4f75-b4fa-5d76a6115069","全模态幻觉的病根在融合层:OWP 免训练重平衡证据与先验","omni-woodpecker-training-free-hallucination-fix","2026-10-26T05:12:35+00:00"]