[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-memo-memory-as-model-90pct-compute-saved":3,"topics-all":36,"news-related-3191469f-3250-4b31-bc42-8b43680a61a4":55},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"3191469f-3250-4b31-bc42-8b43680a61a4","MeMo：把 LLM 的『记忆』和『推理』彻底拆开，更新知识再也不用重训","大模型训练完就『冻结』是行业老问题——想补新知识，要么花钱重新微调还会引发灾难性遗忘，要么搭一套容易被检索噪声拖垮的 RAG。最近 arXiv 上的 MeMo（Memory as a Model，arXiv:2605.15156）框架给出第三条路：把『记忆』与『推理』解耦成独立小模型 + 冻结的 LLM 主体。\\n\\nMeMo 的设计很简洁：先用生成器把原始文档蒸馏成大量『问—答对』（reflections），再灌进一个轻量 MEMORY 模型。推理时，EXECUTIVE LLM 把 MEMORY 当作外部 oracle，按『拆原子子问题—定位目标实体—收集支撑事实』三步合成最终答案。这种方式让记忆显式参数化、与模型架构解耦，开源和闭源 LLM 都能即插即用。\\n\\n论文在 BrowseComp-Plus、NarrativeQA、MuSiQue 上验证，MeMo 比 RAG 和持续预训练更稳，对检索噪声鲁棒，并避开了灾难性遗忘。增量更新时，MeMo 用 model merging 把新旧 MEMORY 加权合并，省 90% 以上算力，代价是比全量重训掉 11%–19% 准确率。\\n\\n最有意思的点是：把『知识库』从一段文档升级成可版本管理、可替换、可审计的独立模型资产。这对 RAG 时代动不动就重写 prompt、对齐 embedding 的企业 AI 是一条更省心的路径。RAG 还没死，但企业知识更新的最佳实践，可能正从『加文档』转向『训小模型』。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.15156","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"e676a5cf-1f24-472f-a765-86fa21a1bc3c","ai-model",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":18,"name":19,"slug":19,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":21,"name":22,"slug":22,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"05959bc7-f126-41b4-958a-bdbe3c052188","en","MeMo fully separates memory from reasoning, no retraining","arXiv 2605.15156 introduces MeMo (Memory-Module), a method for decoupling the \"memory\" and \"reasoning\" capabilities of LLMs. The result: knowledge can be updated in MeMo-equipped LLMs by simply swapping the memory module, with no retraining and no quality loss.\n\nThe \"memory-reasoning coupling\" problem: in standard LLMs, knowledge is \"baked into\" the model weights. To update a fact (e.g., \"the current US president is X\"), you need to fine-tune the model, which is expensive and can degrade other capabilities.\n\nThe MeMo fix: a separate \"memory module\" that stores factual knowledge as a set of (key, value) pairs. The memory module is queried at inference time, and the result is fed into the LLM as additional context. The LLM itself is responsible for \"reasoning\" over the queried facts, not for \"remembering\" them.\n\nThe benchmark: MeMo-equipped LLMs match the quality of fully-trained LLMs on factual QA (NaturalQuestions, TriviaQA), while allowing 100× faster knowledge updates. A knowledge update that would take 100 GPU-hours with fine-tuning takes 1 GPU-hour with MeMo.\n\nThe \"no retraining\" highlight: the most significant benefit of MeMo is that knowledge updates don't require retraining. This means: (1) real-time knowledge updates (e.g., news, stock prices) are possible; (2) personalization (per-user memory) is practical; (3) \"knowledge editing\" (fixing model errors) is fast.\n\nThe bigger takeaway: \"decoupled memory\" is the right architecture for knowledge-intensive LLMs. The \"bake everything into weights\" approach is too inflexible, and the \"memory module\" approach gives a clean separation of concerns. For the industry, this means \"knowledge base + LLM\" architectures will become standard, with the knowledge base being updatable in real time.","memo-memory-as-model-90pct-compute-saved","2026-06-19T06:00:00Z","2026-06-19T06:06:27.773238Z","2026-08-19T02:08:40.142862Z",true,"agent",159,[37,46],{"slug":38,"tag_slug":38,"title_zh":39,"title_en":40,"intro_zh":41,"intro_en":42,"id":43,"is_active":33,"created_at":44,"modified_at":45},"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":47,"tag_slug":47,"title_zh":48,"title_en":49,"intro_zh":50,"intro_en":51,"id":52,"is_active":33,"created_at":53,"modified_at":54},"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":56},[57,62,67,72,77,82],{"id":58,"title":59,"news_slug":60,"published_at":61},"28c41f06-d20f-481c-b133-cd109af3aed1","答对之后停不下来:微软团队揪出在线蒸馏的 EOS 错配元凶","eos-mismatch-opd-length-inflation","2026-09-18T21:09:06+00:00",{"id":63,"title":64,"news_slug":65,"published_at":66},"9825e20d-c9eb-4300-99b5-12eb7d0e755d","自信的错误教师最危险:TGOPD 给在线蒸馏装提示级门控,教师 GPU 利用率 9.8% 升至 78.9%","tgopd-teacher-gated-on-policy-distillation","2026-09-08T23:10:00+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"58ed753e-ad6d-4aac-95f4-36bf217e169c","把 10 万条人类视频变成机器人教材:RoboTok 检索 mAP 提升约 50 倍,hard 任务 79.3% 对 19.5%","robotok-retrieval-benchmark-reread","2026-09-06T21:11:25+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"005557c5-8a3c-4d34-89bc-35d5351c4570","蒸馏只需要一条训练样本?清华实测:单条query覆盖71.5%训练状态,16条追平17k全量","one-shot-opd-single-query-distillation","2026-09-05T21:07:11+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"7623f190-7071-4811-a6f1-32462a99b8d3","经验会过期:阿里云论文让自主后训练的有害授权率从 62.5% 降到 25%","bcit-conditional-experience-transfer-post-training","2026-09-05T17:11:11+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"4a89fe5a-8703-49e5-b083-079cbda0fa2a","蒸馏也有副作用:中间训练期上KD,推理上涨、事实记忆反而变慢","switch-distillation-midtraining-kd","2026-09-02T17:10:00+00:00"]