[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-autocompact-learned-context-compaction":3,"topics-all":38,"news-related-feb04a1c-2c75-42d2-9b62-9c4f0585d924":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},"feb04a1c-2c75-42d2-9b62-9c4f0585d924","AutoCompact:让编码代理自己学会何时压缩上下文","长度触发压缩把时机绑在窗口上限而非任务进度。AutoCompact 给编码代理加可主动调用的 compact() 动作,在线裁判纠错收集 1052 条轨迹做 SFT,再以任务成败为唯一奖励做 GRPO。SWE-bench Verified 绝对提升 9.2 个百分点,16K 窗口仍领先。","编码代理跑长任务时,上下文只增不减:翻代码、跑搜索、改文件、跑测试,轨迹里堆满过期的假设、失败的尝试和冗长的工具输出。眼下主流 harness——论文里点名的 Codex 和 Claude Code——的做法是\"到线才压\":上下文逼近窗口上限时自动压缩。10 月 1 日上传 arXiv 的 AutoCompact(新加坡管理大学、南洋理工与哈佛团队)把这个决策搬进模型策略本身:让代理自己学会什么时候压、留什么、压完怎么接着干。\n\n## 到线才压,问题不在\"压\",在\"何时\"\n\n长度触发的设计把压缩绑在上下文长度上,而不是任务进度上。论文的观察很直接:过期的探索会一直堆积到阈值触发,而触发点可能正落在一个还没了结的阶段中间——此时的证据恰恰最不该被摘要掉。也有让模型自己决定何时压缩的先行工作:推理时用 rubric 的自压缩代理、离线往完整轨迹里插压缩调用的 SWE-Compressor、以及把压缩塞进 RL 循环的 CompactionRL。但论文指出一个共同盲区:这些方法都不监督\"压缩之后的第一步动作\"——工作状态摘要写得再对,代理也可能回头重搜已经记录过的内容。\n\n## 裁判在线纠错,改完再执行\n\nAutoCompact 给代理加了一个可主动调用的 compact() 动作,数据收集阶段用 GPT-5.5-Codex 当在线裁判:每一步审查提案动作,压缩时机不对、摘要漏了关键状态、压完第一步跑偏,都当场改写——被改过的输出直接进环境执行,轨迹从纠正点继续生长。这与事后标注有本质区别:纠错会真实改变后续状态。在 379 个 SWE-rebench 任务上,团队收集了 1,052 条裁判纠正轨迹,其中 24% 管时机、53% 管摘要、23% 管后续动作。先 SFT 初始化,再在 SWE-Gym 上用 GRPO 做联合优化——奖励只有\"最终补丁是否通过测试\"这一个二元信号,不给压缩加任何辅助奖励。\n\n## 数字:两个基准、两档窗口、全线提升\n\n基座是 Qwen3-Coder-30B-A3B-Instruct。SWE-bench Verified 通过率 39.6%(较基座 +9.2 个百分点),SWE-PolyBench Verified 24.5%(+5.0);RL 在 SFT 之上再加 7.4 与 2.8 个百分点。两个设定值得注意:256K 窗口下没有一条轨迹触发强制压缩,SFT 版仍然胜过全量历史基座——说明完整历史即使放得下也不是最优输入;16K 窗口下所有方法共用同一长度触发兜底,训练出的主动压缩依然领先。收益覆盖每任务 $0.10 到 $4.00 的全部推理预算,且预算越紧优势越大——同权重模型跳过 compact() 执行后,$0.10 档通过率掉 19.9 个百分点。行为层面,主动压缩的使用率从 SFT 版的 44.3% 升到 RL 后的 58.5%,摘要漏关键状态的比例从 3.1% 降到 0.2%,漏下一步动作从 8.2% 降到 2.2%。\n\n## 摘要自洽:一个新维度的短板\n\n案例对比刺眼:django-13809 上,SFT 版摘要如实记录了未解决的语法错误,下一步却写\"直接收尾\"——状态与动作互相矛盾,最终以 SyntaxError 失败;RL 版摘要与后续动作一致,任务通过。作者称之为\"摘要自洽性\"。局限也直说:RL 训练用 32K 序列,评测却是 256K 上下文,靠短训长用的迁移撑住收益。\n\n所以呢?上下文管理正在从 harness 的管道工程,变成模型可训练的策略行为——\"何时压缩\"本身就是能力的一部分。论文只验证了一个 scaffold,但方向清楚:还在到线硬压的商业编码代理,等于把一个可学习的决策交给了阈值。\n\n参考:arXiv:2610.02163(arxiv.org\u002Fabs\u002F2610.02163)","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.02163","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},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",{"id":19,"name":20,"slug":20,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":22,"name":23,"slug":23,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"c1dc2640-52d5-43fe-9574-9247a3a82963","en","AutoCompact: Coding Agents Learn When to Compact Context","Coding agents learn when to compact context: 1,052 judge-corrected trajectories, GRPO on task success alone, +9.2 points on SWE-bench Verified.","When a coding agent runs a long task, its context only grows: inspecting files, running searches, editing code, executing tests. The trajectory piles up stale hypotheses, failed attempts, and verbose tool outputs. Today's mainstream harnesses — the paper names Codex and Claude Code — handle this with a length-triggered design: compact automatically as the context approaches the window limit. AutoCompact, posted to arXiv on October 1 by Xuan Zhang, Longtao Zheng and colleagues (Singapore Management University, Nanyang Technological University, and Harvard among the affiliations), moves that decision into the model's own policy: the agent learns when to compact, what to preserve, and how to continue afterwards.\n\n## The problem with \"compact when full\" is the \"when\"\n\nLength-triggered compaction ties the decision to context length rather than task progress. The paper's observation is blunt: stale exploration keeps accumulating until a threshold fires, and that trigger can land in the middle of an unresolved stage — exactly when the evidence is most needed. Prior work did let models decide: rubric-based self-compacting agents at inference time, SWE-Compressor inserting compaction calls into completed trajectories offline, and CompactionRL folding compaction into an RL loop. But the authors point out a shared blind spot: none of them supervise the first action after compaction — even a correct working-state summary can be followed by the agent re-running a search whose result was already recorded.\n\n## An online judge that corrects, then executes\n\nAutoCompact gives the agent a proactively invocable compact() action. During data collection, GPT-5.5-Codex serves as an online judge: at every step it reviews the proposed action — wrong compaction timing, a summary missing key state, a post-compaction step going off-track are all rewritten on the spot. Corrected outputs are executed in the environment directly, so the trajectory grows from the correction point. This differs fundamentally from post-hoc annotation: the correction genuinely changes later states. Across 379 SWE-rebench tasks the team collected 1,052 judge-corrected trajectories — 24% supervising triggers, 53% working-state construction, 23% continuation. SFT initializes the policy; GRPO on SWE-Gym then jointly optimizes coding and compaction with a single binary reward: whether the final patch passes the task's tests. No compaction-specific reward shaping.\n\n## The numbers: two benchmarks, two window regimes, gains across the board\n\nThe base model is Qwen3-Coder-30B-A3B-Instruct. Pass rate reaches 39.6% on SWE-bench Verified (+9.2 points absolute over base) and 24.5% on SWE-PolyBench Verified (+5.0); outcome RL adds a further 7.4 and 2.8 points over the SFT checkpoint. Two settings stand out. Under a 256K window, no trajectory ever hits forced compaction, yet the SFT variant still beats the full-history base — a complete history is not the optimal input even when it fits. Under a constrained 16K window where all methods share the same length-triggered fallback, the trained proactive compaction still leads. Gains hold across inference budgets from $0.10 to $4.00 per task, and are largest where budget is tightest: with the same trained checkpoint but compact() execution skipped, the $0.10 tier loses 19.9 points. Behaviorally, proactive compaction usage rises from 44.3% of tasks after SFT to 58.5% after RL, key-state omissions in summaries fall from 3.1% to 0.2%, and next-action omissions from 8.2% to 2.2%.\n\n## Summary self-consistency: a newly named weakness\n\nThe case study offers a sharp contrast. On task django-13809, the SFT summary faithfully records an unresolved syntax error, yet proposes to finish without further action — state and action contradict each other, and the run fails with a SyntaxError. The RL summary stays consistent with its recorded state and passes. The authors call this dimension summary self-consistency. The limitations are stated plainly: RL training used 32K-token sequences while evaluation ran at 256K context, with the gains resting on that short-train, long-evaluate transfer.\n\nSo what? Context management is shifting from harness plumbing into trainable model behavior — knowing when to compact is itself part of the capability. The paper validates a single scaffold, but the direction is clear: commercial coding agents that still hard-compact at the limit are handing a learnable decision to a threshold.\n\nReference: arXiv:2610.02163 (https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.02163)","autocompact-learned-context-compaction","2026-10-04T17:10:39Z","2026-10-04T17:10:44.926248Z","2026-10-04T17:10:44.926259Z",true,"agent",673,[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},"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",{"id":65,"title":66,"news_slug":67,"published_at":68},"859242dd-0035-41e6-8db9-4583989048b7","AgentSWE基准:让编码智能体造智能体,无人及格","agentswe-coding-agent-benchmark","2026-10-13T19:08:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"a820bb5a-ac48-4aab-8d8b-34da340900c1","REMORY 让 LLM 用 5% 上下文继续跑 Agent","remory-agent-memory","2026-10-11T09:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"d552a77b-3b52-4d59-8962-e897aa06b37c","腾讯RSR:把借来的轨迹改写成自己的能力","recursive-self-rewrite-terminal-trajectories","2026-10-06T21:06:49+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"88d15f9c-5548-4dae-bde5-75e79c3d0a28","8K 上下文顶 1M:把长文档问题训练成 agent 接力,而不是堆窗口","trained-agentic-context-management","2026-10-01T11:32:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"e4b3903e-c65f-46cb-90ae-81502eb8cdd9","CliffCompaction开源:只删不改的会话压缩,长程Agent成本砍半","cliffcompaction-truncate-only-agent-compaction","2026-09-23T17:10:56+00:00"]