[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-memento-context-management-long-context-agent":3,"topics-all":38,"news-related-8676e640-c0d6-4066-9021-f82e3adaffda":67},{"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},"8676e640-c0d6-4066-9021-f82e3adaffda","Memento:8B 模型读完 250 万 token 长书","哥伦比亚大学团队提出 Memento:让 Qwen3-8B 学会自己「记笔记」管理上下文。先用合成轨迹 SFT 冷启动,再用 GRPO 强化。在最长 250 万 token 的整本书问答上,准确率从 36.9% 提到 67.8%,上下文始终不超过 2 万 token,成本约为 753B GLM-5.3 的九分之一。","把一整本书塞进上下文窗口，然后祈祷模型自己记得住重点——这是长文档问答最常见的做法，也是最贵的做法。哥伦比亚大学团队给出的数据说明，问题可能不在窗口大小，而在于「记笔记」这件事本身从没被训练出来。\n\n论文叫 Memento: Learning Context Management for Long Context Processing Agents，2026 年 10 月 30 日挂上 arXiv（[arXiv:2611.00244](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.00244)）。任务设定很具体：整本书的多跳问答。数据里 90% 的书超过 5.1 万个 Qwen3 token，70% 超过 9 万，最长的一本 250 万 token。而训练用的模型上下文只有 16K 到 20K——也就是说，一个合格的智能体必须自己把进度带过几十个窗口。\n\n## 问题出在笔记写不对\n\n论文给了一个扎心的观察：现成的智能体模型不会记笔记。GPT-5.5 在 337 条测试集上一条压缩笔记都没写过；GLM-5.3 在训练集上跑了 1.2 万条 rollout，没有一条包含有效笔记；Qwen3-8B 稍好一些，但也爱把一模一样的笔记连着写两遍，只有约 6% 的轨迹是有效的。\n\n更麻烦的是，光靠 RL 训练不出这个行为。原因是探索：策略如果从来没采样到过好的笔记写法，RL 就没有东西可以强化。论文观察到，没有 SFT 预热的策略在 RL 开始后 20 步内就把压缩频率降到接近零，然后彻底放弃这项技能。\n\n## 两段式训练\n\n解决办法是先用合成数据把行为演示出来，再让 RL 把它变可靠。\n\n第一步是 SFT 冷启动。论文让 GLM-5.3 在自己的 harness 里跑出正确轨迹，再事后注入压缩笔记——要求它总结关键发现、要避免的错误、下一步计划，不合格的笔记重新采样，搜索动作和最终答案保持不动。过滤后大约 6K 条轨迹拿来微调 Qwen3-8B。\n\n第二步是 RL 强化，用 GRPO，奖励由四项组成：格式合规（作为总闸门，不合规其余奖励全部归零）、笔记有效性、答案正确性、以及用负 Brier 分数衡量的置信度校准。笔记有效性有三条硬性启发式：每个非末尾窗口至少一条笔记、总数少于五条；笔记不能是窗口内的第一次工具调用，也不能紧跟在另一条笔记后面；任意两条笔记的 token 集合相似度不能超过 80%。\n\n关键在训练与推理的解耦：训练时只会「压缩进新窗口」这一种操作，但同一套笔记结构在推理时可以直接切换到另外两种策略——上下文折叠（额外保留最新笔记之后的对话轮次）和基于置信度的回溯（答案置信度低时，退回早前的笔记检查点重新生成）。\n\n## 数字\n\n在 LiteraryQA、NovelHopQA、SagaScale（英文子集）上训练与评测，NovelQA 完全留出做泛化测试，测试集共 4.7K 题。原始 Qwen3-8B 智能体准确率 36.9%，完整系统达到 67.8%，全程上下文占用不超过 2 万 token。\n\n对照组更刺眼。论文报告：753B 的 GLM-5.3 在同一 harness 下 69.3%，成本约为 Memento 的 9 倍；把整本书直接塞进 GLM-5.3 的窗口能到 74.3%，高出 6.4 个点，但推理成本超过 26 倍。GPT-5.5 在同一 harness 下只有 13.1%——不过论文给了原因：它只有 21.4% 的 rollout 能输出可解析的最终答案，且从不调用笔记工具，即便修正输出格式也只有 57.1%，成本还高出 54 倍。这里得说明，GPT-5.5 因 API 成本限制每个数据集最多只测 100 题，SagaScale 测试集本身也只有 37 题，噪声不小。\n\n消融数据把「为什么两段式缺一不可」摊开了：单窗口下未训练 29.6%、SFT 34.4%、SFT+RL 58.6%；加上上下文折叠变成 32.6% \u002F 46.6% \u002F 66.9%；再加置信度回溯则是 36.9% \u002F 59.1% \u002F 67.8%。SFT 单独不够，因为它会过度压缩、在跨窗口时累积错误；RL 单独也不够，因为它采不到可强化的样本。\n\n## 我的判断\n\n这篇论文真正有意思的地方不是「8B 打败了 GPT-5.5」——那个对比有很强的 harness 依赖，GPT-5.5 的低分主要来自它没被教会用这套笔记工具，论文自己也把这点写在了正文里。\n\n真正值得记下来的是两条。\n\n第一条，上下文管理不是提示词工程，而是一项可训练的能力。行业普遍的假设是「长上下文靠更大的窗口解决」，但这篇论文给出的成本曲线是：买窗口要付 26 倍的钱，学压缩只要 9 分之一。对长文档检索、代码库理解这类要反复读大量材料的场景，这笔账很好算。\n\n第二条，顺序比配方更重要。RL 在这里的失败方式非常典型——不是训不好，而是根本训不起来，因为探索空间里没有可强化的样本。SFT 的作用不是提升准确率本身（论文指出，RL 开始时有没有 SFT 初始化，验证准确率几乎一样），而是先把策略推到 RL 能采到好样本的区域。这对任何「要用 RL 训出模型原本不会的行为」的场景都是通用提醒。\n\n当然这篇也有明显边界：只做了英文整本书问答，答案短，还依赖 LLM judge 判分；笔记有效性靠手工启发式，未必等于真的有用；训练只覆盖了压缩进新窗口这单一操作，折叠和回溯完全靠推理时迁移。所以它证明的是「这条路线可行」，不是「这条路线已经解决」。\n\n上下文窗口这条军备竞赛，可能正在被另一条路线悄悄绕过去：不是让模型装得下更多，而是让它学会扔掉大部分。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.00244","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"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},"7ac06d8e-b074-4147-abfc-ffaa4c6b8744","ai-efficiency",{"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},"c187600e-804c-4697-b828-1e4330e0eb10","qwen",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"12346f7c-590f-4b4d-9ef7-42649895ce95","en","Memento: An 8B Model Reads 2.5M Tokens in a 20K Window","Columbia's Memento teaches an 8B agent to take notes and manage its own context, hitting 67.8% on 2.5M-token book QA at a ninth of GLM-5.3's cost.","Stuffing an entire book into the context window and then hoping the model remembers what mattered is the default approach to long-document QA — and also the most expensive one. A team from Columbia University suggests the real problem is not window size, but that nobody ever trained models to take notes.\n\nThe paper is \"Memento: Learning Context Management for Long Context Processing Agents\", posted to arXiv on 30 October 2026 ([arXiv:2611.00244](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.00244)). The task is specific: multi-hop question answering over whole books. In their data, 90% of books exceed 51,000 Qwen3 tokens, 70% exceed 90,000, and the longest reaches 2.5 million tokens. The models they train with have context windows of only 16K to 20K — meaning a competent agent has to carry its own progress across dozens of windows.\n\n## The problem is bad notes, not long context\n\nThe paper makes an uncomfortable observation: off-the-shelf agentic models do not know how to take notes. GPT-5.5 wrote not a single compaction note across a 337-sample test set. Across 12,000 GLM-5.3 rollouts on the training set, not one contained a valid note. Qwen3-8B did slightly better, but it tends to write identical notes back to back, and only about 6% of its trajectories are valid.\n\nWorse, reinforcement learning alone cannot produce the behavior. The reason is exploration: if the policy never samples a good note, RL has nothing to reinforce. The paper reports that without SFT warm-starting, the compaction rate collapses to near zero within the first 20 RL steps, and the policy abandons the skill entirely.\n\n## Two-stage training\n\nThe fix is to demonstrate the behavior with synthetic data first, then use RL to make it reliable.\n\nStep one is SFT cold-start. The team ran GLM-5.3 in their harness to collect correct trajectories, then injected compaction notes after the fact — asking the model to summarize key findings, mistakes to avoid, and next steps. Notes that failed the validity heuristics were resampled; search actions and final answers were left untouched. Roughly 6K trajectories survived filtering and were used to fine-tune Qwen3-8B.\n\nStep two is RL, using GRPO, with a reward built from four parts: format compliance (which gates everything else — no format, no reward), note validity, answer correctness, and confidence calibration via a negated Brier score. Note validity rests on three hand-crafted heuristics: every non-final window contains at least one note and the rollout has fewer than five in total; no note may be the first tool call in its window nor immediately follow another note; and no two notes may exceed 80% token-set similarity.\n\nThe key design is decoupling training from inference. Training only ever exercises \"compact into a fresh window,\" yet the same note structure supports two more strategies at test time: context folding (also retaining the turns after the latest note) and confidence-based backtracking (reverting to an earlier note checkpoint after a low-confidence answer and generating an alternative continuation).\n\n## The numbers\n\nTraining and evaluation used LiteraryQA, NovelHopQA, and the English subset of SagaScale, with NovelQA held out entirely for generalization testing — 4.7K test questions in total. The original Qwen3-8B agent scored 36.9%; the full system reaches 67.8%, while never holding more than 20K tokens in context.\n\nThe comparison numbers are starker. The paper reports that GLM-5.3 at 753B scores 69.3% in the same harness at roughly 9x Memento's cost, and that placing the whole book in GLM-5.3's context window yields 74.3%, 6.4 points higher, but at more than 26x the inference cost. GPT-5.5 reaches only 13.1% in the same harness — though the paper supplies the reason: only 21.4% of its rollouts produce a parsable final answer, and it never calls the note tool, so even after correcting for output format it lands at 57.1%, at 54x the cost. Worth flagging: GPT-5.5 was evaluated on at most 100 questions per dataset because of API cost, and the SagaScale test split contains only 37 questions, so the noise is not trivial.\n\nThe ablations show why both stages are needed. With a single window, the untrained model gets 29.6%, SFT gets 34.4%, and SFT+RL gets 58.6%. Adding context folding: 32.6% \u002F 46.6% \u002F 66.9%. Adding confidence-based backtracking: 36.9% \u002F 59.1% \u002F 67.8%. SFT alone is insufficient because it over-compacts and accumulates errors across windows; RL alone is insufficient because it never samples anything worth reinforcing.\n\n## My read\n\nThe genuinely interesting part is not \"an 8B model beats GPT-5.5.\" That comparison depends heavily on the harness — GPT-5.5 scores low mainly because it was never taught to use this note-taking tool, and the paper says so in its own text.\n\nTwo things are worth keeping. First, context management is not prompt engineering; it is a trainable capability. The prevailing assumption is that long context gets solved by buying a bigger window, but this paper's cost curve says otherwise: buying the window costs 26x, learning to compress costs a ninth. For long-document retrieval or codebase comprehension, where an agent re-reads large volumes of material, that is an easy calculation.\n\nSecond, ordering matters more than recipe. RL's failure mode here is instructive — not that it trains poorly, but that it cannot train at all, because its exploration space contains no reinforceable samples. SFT does not primarily raise accuracy; the paper notes that at the start of RL, validation accuracy is nearly identical with and without SFT initialization. Its job is to push the policy into a region where RL can sample good trajectories. That is a general lesson for any project trying to teach RL a behavior the base model does not have.\n\nThe boundaries are real, too: English-only whole-book QA, short answers, LLM-judge grading; note validity enforced by hand-crafted heuristics that may not capture usefulness; and training covers only compaction into a fresh window, with folding and backtracking existing purely as test-time transfers. What it demonstrates is that this route works, not that it is solved.\n\nThe context-window arms race may be quietly getting routed around by a different approach: not letting models hold more, but teaching them to throw most of it away.","memento-context-management-long-context-agent","2026-11-04T03:10:00Z","2026-11-04T03:09:58.901369Z","2026-11-04T03:09:58.901378Z",true,"agent",369,[39,49,58],{"slug":40,"tag_slug":41,"title_zh":42,"title_en":43,"intro_zh":44,"intro_en":45,"id":46,"is_active":35,"created_at":47,"modified_at":48},"image-generation","text-to-image","图像生成：从文生图到逐框编辑的模型竞赛","Image Generation: from text-to-image to frame-level editing","文生图已是红海，战场正转向编辑与可控性：FLUX.2\u002F3 用开源和量化把生图搬进消费级显卡，Qwen-Image、Seedream 等国产模型紧咬不放，GPT Image 2 则把推理能力带进像素。本专题追踪图像生成与编辑模型的发布、架构演进、加速量化与商业化数据，看「好图」如何变成「可用的图」。","Text-to-image is a crowded race, and the frontier has moved to editing and control: FLUX.2\u002F3 push open weights and quantization onto consumer GPUs, Qwen-Image and Seedream close in from China, and GPT Image 2 brings reasoning into pixels. This topic tracks model releases, architecture shifts, acceleration work, and the business numbers behind \"pretty\" becoming \"usable\".","52de7c59-7864-4569-8ae7-3d656d26a1a1","2026-11-20T07:20:23.534506Z","2026-11-20T07:20:23.534527Z",{"slug":50,"tag_slug":50,"title_zh":51,"title_en":52,"intro_zh":53,"intro_en":54,"id":55,"is_active":35,"created_at":56,"modified_at":57},"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":59,"tag_slug":59,"title_zh":60,"title_en":61,"intro_zh":62,"intro_en":63,"id":64,"is_active":35,"created_at":65,"modified_at":66},"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":68},[69,74,79,84,89,94],{"id":70,"title":71,"news_slug":72,"published_at":73},"e8f13686-3f2f-468a-b974-55634bb30795","Metal Sandbox:Mac 集群的 GPU 调度术","metal-sandbox-mac-cluster-gpu-scheduling","2026-11-15T03:15:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"13d6adf3-a5c3-46fb-9ff5-a6999dcf0300","Critic值轮内近乎恒定:MPO重构Agent信用分配","mpo-turn-level-credit-long-horizon-agents","2026-10-28T19:08:42+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"74cb8016-bdf7-4c8b-adbc-456d8c899b1a","EdgeWise 把 9B 模型训练成「喊人」高手:边云协同砍半成本","nvidia-edgewise-edge-cloud-slm-self-delegation-coding","2026-10-19T01:01:45+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"48726cf0-6f0c-4b09-841e-06aa8b0ffcfd","上海AI Lab Intern-Decision:多模态决策模型,4B 微超 Jev 90.02%","intern-decision-qwen3-5-multimodal-decision-model","2026-10-18T11:13:06+00:00",{"id":90,"title":91,"news_slug":92,"published_at":93},"97ee186c-a42b-4389-8427-3342d77e8590","小模型何时喊云端大模型?NVIDIA 砍 62% 推理成本","edgewise-edge-agents-cloud-delegation","2026-10-16T21:07:57+00:00",{"id":95,"title":96,"news_slug":97,"published_at":98},"fead7184-42cc-4604-9aba-48dcb20ef7ab","阿里2852条长轨迹:喂出12基准全涨的27B智能体","terminalhorizon-long-trajectory-27b-agent","2026-10-13T21:08:01+00:00"]