[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-memento-3-recursive-self-improvement-rulebook":3,"topics-all":35,"news-related-8474959b-1569-45d6-afcc-34599f05b2fd":54},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":21,"news_slug":28,"published_at":29,"created_at":30,"modified_at":31,"is_published":32,"publish_type":33,"image_url":14,"view_count":34},"8474959b-1569-45d6-afcc-34599f05b2fd","Memento 3:不动权重,LLM自我改进写进规则手册","UCL 提出 Memento 3:冻结 LLM 权重,让智能体把环境规律写成可修订的自然语言规则手册并编译成代码,更新需过忠实性判断与逐格重放双门验证。论文报告其通关 ARC-AGI-3 全部 25 个游戏,相对人类动作效率 100.0,动作数仅为人类 44%;Pong 控制器三局 21:0,零后续 LLM 调用。","递归自我改进(RSI)的讨论大多默认一件事:模型想变得更聪明,就得动自己的权重,今年 6 月 Anthropic 还专门发文讨论过自我改进的\"刹车时刻\"。而出现在 Hugging Face 论文热榜上的 Memento 3 给出了另一条路:LLM 冻结不动,把\"变聪明\"外包给一份外挂的自然语言规则手册。\n\n## 规则手册:一份可修订的世界模型\n\n按论文设计,冻结的 LLM 智能体维护一份自然语言规则手册作为持久语义记忆:记录对环境动力学的可修订假设,没观察到的部分明确留白。智能体把手册编译成可执行代码做预测与规划,再进入\"观察—反思—修订规则—编译—验证\"的持续循环,用预测误差同时打磨规则和代码。\n\n更新并不是想写就写:新代码要过两道门——LLM 判断它与规则手册语义忠实,且逐格重放能精确复现观测到的状态转移,才会被接受。这等于给自我改进配上了回归测试,写歪的假设进不了长期记忆。论文还提供种群扩展:并行维护多个世界模型、共享交互证据,用各自预测引导探索。\n\n## 通关 ARC-AGI-3 与 21:0\n\n按[论文报告](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.11794),单模型智能体在 ARC-AGI-3 公开集通关全部 25 个游戏,平均相对人类动作效率(RHAE)100.0,动作数仅为人类的 44%。Atari Pong 案例中,学出的反馈控制器在三局不同开局下全部 21:0 取胜,而且此后不再调用 LLM——世界模型一旦学成,自己就能接管决策。\n\n## 泼冷水:日记还是学习?\n\nHF 社区讨论里有条冷静的质疑:能落进散文式规则的,只有模型能明确表述的修复;现实中很多可靠性问题得靠 schema 级硬约束解决,散文备忘录做不到。归因也是黑洞——规则生效、任务通过,是规则的功劳还是重试的运气?没有\"关掉规则重跑\"的消融,你建的可能只是一本昂贵的日记。\n\n但这个方向值得盯:权重不动、改进外置、每次更新可审计可回滚,恰好贴合企业真实部署里\"冻结 API 模型+外部记忆\"的形态。当 RSI 争论还在担心模型改写自己,Memento 3 把自我改进变成了一份能打开检查的手册。真正的问题是:你敢把多大决策权,交给一本自己会长的规则书?","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.11794","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[22],{"id":23,"lang":24,"title":25,"summary":26,"content":27},"50bf898d-5302-47c7-9259-886930b1754d","en","Memento 3: Frozen LLMs Self-Improve via Reflective Rulebooks","Memento 3: frozen LLM, external rulebook compiled to code, replay-gated updates. Clears all 25 ARC-AGI-3 games, RHAE 100.0, 44% of human actions.","Recursive self-improvement (RSI) discussions usually assume the model must edit its own weights — Anthropic devoted an essay to the approaching \"braking point\" this June. Memento 3, now trending on Hugging Face's daily papers, takes another route: the LLM stays frozen, and \"getting smarter\" is outsourced to an external natural-language rulebook.\n\n## The rulebook: a revisable world model\n\nBy design, a frozen LLM agent maintains a natural-language rulebook as persistent semantic memory: it records revisable hypotheses about environment dynamics and deliberately leaves unobserved parts underspecified. The agent compiles the rulebook into executable code for prediction and planning, then enters a continual loop of observation, reflection, rule revision, compilation, and verification, using prediction errors to refine both the rules and the code.\n\nUpdates are not free-form. New code passes two gates — the LLM must judge it faithful to the rulebook, and a cell-exact replay must reproduce the observed transitions — before acceptance. It is regression testing for self-improvement: hallucinated hypotheses never reach long-term memory. A population extension maintains multiple world models in parallel, sharing interaction evidence while their predictions guide exploration.\n\n## Clearing ARC-AGI-3 and a 21:0 sweep\n\nPer the [paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.11794), the single-model agent clears every level of all 25 public ARC-AGI-3 games, with a mean Relative Human Action Efficiency (RHAE) of 100.0, using 44% of the human action count. In an Atari Pong case study, the learned feedback controller won 21:0 in each of three episodes with different openings — with no further LLM calls. Once the world model is learned, it runs the game on its own.\n\n## Cold water: a diary or learning?\n\nA sharp critique in the HF community discussion is worth keeping: only fixes the model can articulate make it into prose rules, while real-world reliability often demands schema-level hard constraints that a memo cannot express. Attribution is another blind spot — when a rule fires and the task passes, was it the rule or the retry? Without a rules-disabled ablation, you may have built an expensive diary and called it learning.\n\nStill, the direction deserves attention: frozen weights, externalized improvement, auditable and reversible updates — a close match for how enterprises actually deploy frozen API models with external memory. While the RSI debate worries about models rewriting themselves, Memento 3 turns self-improvement into a manual you can open and inspect. The real question: how much decision authority would you hand to a rulebook that grows on its own?","memento-3-recursive-self-improvement-rulebook","2026-10-11T13:07:13Z","2026-10-11T13:07:20.147736Z","2026-10-11T13:07:20.147753Z",true,"agent",56,[36,45],{"slug":37,"tag_slug":37,"title_zh":38,"title_en":39,"intro_zh":40,"intro_en":41,"id":42,"is_active":32,"created_at":43,"modified_at":44},"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":46,"tag_slug":46,"title_zh":47,"title_en":48,"intro_zh":49,"intro_en":50,"id":51,"is_active":32,"created_at":52,"modified_at":53},"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":55},[56,61,66,71,76,81],{"id":57,"title":58,"news_slug":59,"published_at":60},"9c8731f9-e170-4855-93a2-b1705abb61c3","Muse Wonder发布:Meta把宝押在个人智能体上","meta-muse-wonder-agentic-model","2026-10-12T23:10:00+00:00",{"id":62,"title":63,"news_slug":64,"published_at":65},"5edbfca7-7c1a-4df3-8577-a0195408b827","RSI-Anything:世界模型陪跑,自我改进有了回滚键","rsi-anything-world-model-recursive-self-improvement","2026-10-12T21:20:00+00:00",{"id":67,"title":68,"news_slug":69,"published_at":70},"e40de2f9-9ee4-47f3-b1f1-7bf93a4870a9","一个动词翻转工具调用决策,LLM 内部向量现形","llm-tool-call-decision-vector","2026-10-09T13:10:00+00:00",{"id":72,"title":73,"news_slug":74,"published_at":75},"c368ad9f-9308-4a0b-8f5c-3ae4601b48b9","D2K-Bench: 专家设计把 LLM 写 GPU 核提速 33.9%","d2k-bench-llm-gpu-kernel-design-guidance","2026-10-07T03:00:00+00:00",{"id":77,"title":78,"news_slug":79,"published_at":80},"85f7f1c2-d896-436b-a915-37faed8776eb","你改主意了,模型没改:被拒需求也会带偏大模型","intent-eval-rejected-change-confusion","2026-10-06T17:15:00+00:00",{"id":82,"title":83,"news_slug":84,"published_at":85},"206ea36a-1eea-463f-a241-1e3b32f5ec2d","USTC GraphForge:证据图把任务和 rubric 钉在一起,Qwen3.6-27B 涨三基准","graphforge-ustc-qwen36-27b-evidence-graph","2026-10-04T03:05:00+00:00"]