[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-rsi-prime-decision-sft-model-training":3,"topics-all":38,"news-related-bc983cc8-6c0e-4e7a-9d0c-cd6a3175c18e":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},"bc983cc8-6c0e-4e7a-9d0c-cd6a3175c18e","模型学炼丹:350 条轨迹教会 GLM-5.3 训模型","模型能不能学会训练模型?RSI-Prime 把 350 条完整炼丹轨迹做成数据集,给 90,738 个决策标注质量,只在好决策上算 loss。GLM-5.3 微调后在 PostTrainBench 从 36.56 涨到 48.24,相对提升 31.95%。监督更少,分数更高。","炼丹这件事,正在从手艺变成模型可以习得的技能。11 月 17 日,一篇来自 Sun Yat-sen University 等机构的 arXiv 预印本 RSI-Prime(https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.23138)给出了新证据:把 350 条完整炼丹实验轨迹做成数据集,微调 GLM-5.3 之后,它在 PostTrainBench 上的总分从 36.56 涨到 48.24,论文实测相对提升 31.95%。\n\n## 实验过程本身成了数据\n\n先看赛场。PostTrainBench 是另一个独立团队搭的基准:给 agent 一张 H100、10 小时预算和一个 Qwen3-4B-Base 底座,让它自己找数据、训练、迭代、交模型。闭源组合在这个榜上明显领先,按论文表格,Fable 5.1 + Claude Code 拿到 53.68,而 GLM-5.3 的起点只有 36.56。\n\nRSI-Prime 的思路是把实验过程直接变成监督数据:350 条完整轨迹,覆盖七类任务,拆出 90,738 条决策记录。轨迹不是随便录的,提交的模型必须跨过一道门槛,相对改进要超过参考 Instruct 与 Base 差距的一半(g>0.5),每条轨迹还要过 PostTrainBench 的四个轨迹裁判,失败的中间尝试则全部保留在上下文里。\n\n## 成功的实验里,也有烂决策\n\n论文最扎实的洞察是:实验最终成功,不代表每个决策都值得学。五类动作里,评测与诊断占 38.43%,数据准备占 27.09%,炼丹时间的重心其实不在训练执行本身。数据集里 92.42% 的决策被评为高质量,但实现优化类决策的问题率高达 16.89%,数据源选择也有 12.80%。全盘模仿,等于把坑一起学进去。\n\n于是他们做了决策级 loss 掩码:一条消息关联的所有决策全部通过质量筛选,才参与监督,否则整条消息被 mask 掉。保守规则最终保留 76.25% 的监督消息、75.48% 的 token,也就是 68,027 \u002F 89,211 条消息。质量标注由 GLM-5.3 自己打三次、多数投票,再抽 400 条做人工审计,接受率在 97.50% 到 99.50% 之间,论文也自陈这套审计不估计漏标率。\n\n## 监督更少,分数反而更高\n\n结果段是全文最反直觉的地方。全轨迹微调把 GLM-5.3 从 36.56 拉到 45.61,决策级监督再拉到 48.24,六个子任务全部转正,其中 AIME 2025 提高 12.23 分、HealthBench 提高 26.63 分。两个对照组都被压了下去:按工具报错掩码,保留 89.15% 的 token,只有 45.12;按步骤重要性筛选(改造自 ATLAS),保留 76.63% 的 token,只有 43.21,监督量与决策级几乎相同,分数却低了 5.03。\n\n行为层面同样在改善:评测轨迹里,决策质量通过率 96.11%,全轨迹版本是 87.25%。按论文表格,48.24 已经压过 Gemini 4 Argon(47.92)、Opus 5.5(47.84)、GPT-6.1 Sol(46.45)和 Fable 5(46.31),距离榜首的 Fable 5.1 + Claude Code 还差 5.44 分。\n\n## 该泼的冷水\n\n- 训练语料含 HumanEval 相关内容,论文自己承认这一项存在训练-评测重叠;\n- 基线的三次独立运行里,Arena Hard Writing 单项分别跑出 25.79、37.03 和 52.57,run 间方差不小,而三次平均只对原始 GLM-5.3 做过;\n- 质量标注的打分者就是被微调的 GLM-5.3 本身,人工审计只抽了 400 条;\n- 全部基准数值均为论文自报,暂时没有第三方复现。\n\n## 不是孤例\n\n这个方向并非一家之言。另一个团队 10 月放出的 RSI-Dojo(https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.03698),同样把后训练轨迹拿来教 agent 炼丹;Lim 等人 8 月的实证分析(https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.19072)看了 1,338 条后训练轨迹,发现只有 2.1% 的相邻训练轮转换会真正改变策略,agent 会修 bug,却很少回头改方案。决策级监督补的恰好是这一环:让模型学会在什么时候换打法,而不只是把既定流程跑顺。\n\n所以呢:炼丹经验第一次被做成了可复用的数据资产,会训模型本身成了一种可以微调出来的能力。更刺激的下一个问题是,用这套方法训出来的模型,再拿去训下一代模型,循环转起来会发生什么?论文没有回答,但路已经铺到了这里。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.23138","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},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",{"id":19,"name":20,"slug":20,"description":14,"color":14},"f72d264d-7fa0-458b-8d43-0ec5168d69db","instruct-model",{"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},"73695a54-c9fb-47d2-9171-4076b68056e7","en","GLM-5.3 Learns to Train Models: Decision Masking +31.95%","RSI-Prime distills 350 fine-tuning trajectories into decision-level supervision; GLM-5.3 rises from 36.56 to 48.24 on PostTrainBench, up 31.95%.","In AI slang, 'alchemy' means the craft of training and tuning models. A preprint posted November 17 by authors from Sun Yat-sen University and other institutions (https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.23138) argues that this craft is now something a model can learn. RSI-Prime packs 350 complete fine-tuning experiments into a dataset; after training on it, GLM-5.3 lifts its PostTrainBench overall score from 36.56 to 48.24, a 31.95% relative gain as measured by the paper.\n\n## The experiment process becomes data\n\nFirst, the arena. PostTrainBench, built by an independent team, hands an agent one H100 GPU, a 10-hour budget and a Qwen3-4B-Base target, then lets it find data, train, iterate and submit a model. Closed-source pairs lead this board: per the paper's table, Fable 5.1 + Claude Code sits at 53.68, while GLM-5.3 starts at 36.56.\n\nRSI-Prime turns the experiment process itself into supervision: 350 complete trajectories across seven task categories, decomposed into 90,738 decision records. A trajectory only enters the set if its submitted model clears a bar — relative improvement exceeding half the reference Instruct-Base gap (g>0.5) — and it must pass PostTrainBench's four trajectory judges; failed intermediate attempts stay in context.\n\n## Good runs contain bad decisions\n\nThe sharpest observation in the paper: a successful experiment does not mean every decision deserves imitation. Among five action categories, evaluation and diagnosis account for 38.43% of records and data preparation 27.09% — the workload's center of gravity is not training execution itself. 92.42% of records are labeled high quality, but implementation-optimization decisions show a 16.89% problem rate, and data-source selection shows 12.80%. Blind imitation learns the pitfalls along with the skill.\n\nTheir fix is decision-level loss masking: an assistant message is supervised only when every decision linked to it passes quality screening; otherwise the whole message is masked. The conservative rule keeps 76.25% of supervised messages and 75.48% of supervised tokens — 68,027 of 89,211 messages. Quality labels come from GLM-5.3 itself, three assessments combined by majority voting, plus a 400-record human audit with acceptance rates between 97.50% and 99.50%; the authors state that this audit does not estimate the recall of decision identification.\n\n## Less supervision, higher scores\n\nHere is the counterintuitive part. Full-trajectory fine-tuning takes GLM-5.3 from 36.56 to 45.61, and decision-level supervision pushes further to 48.24, improving all six sub-tasks, with AIME 2025 up 12.23 points and HealthBench up 26.63 points. Both natural controls fall behind: masking on tool failures, which keeps 89.15% of supervised tokens, scores only 45.12; step-importance selection adapted from ATLAS, which keeps 76.63%, scores 43.21 — nearly the same supervision volume as decision-level, yet 5.03 points lower.\n\nBehavior improves too: on evaluation trajectories the decision-quality pass rate reaches 96.11%, against 87.25% for full-trajectory training. Per the paper's table, 48.24 already sits above Gemini 4 Argon (47.92), Opus 5.5 (47.84), GPT-6.1 Sol (46.45) and Fable 5 (46.31), 5.44 points behind the leader, Fable 5.1 + Claude Code.\n\n## Caveats worth stating\n\n- The training corpus contains HumanEval-related content, which the paper itself flags as train-evaluation overlap on that task;\n- The three reference runs of baseline GLM-5.3 scored 25.79, 37.03 and 52.57 on Arena Hard Writing — sizable run-to-run variance, and three-run averaging was applied only to the original agent;\n- The quality annotator is the very model being fine-tuned, and the human audit covers 400 records;\n- All benchmark numbers are self-reported by the paper, with no independent replication yet.\n\n## Not a lone data point\n\nThe direction is not one team's bet. RSI-Dojo, released in October by a different group (https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.03698), likewise teaches agents post-training from curated trajectories. An August empirical study by Lim et al. (https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.19072) analyzed 1,338 post-training trajectories and found that only 2.1% of transitions between adjacent training runs ever change strategy — agents repair implementations but rarely revise their plans. Decision-level supervision targets exactly this gap: teaching the model when to change approach, not just how to run a fixed pipeline smoothly.\n\nSo what: alchemy experience has become a reusable data asset, and knowing how to train models is now a capability you can fine-tune into a model. The sharper question — what happens when a model trained this way goes on to train the next generation, and the loop starts turning — is left open by the paper. But the road is now paved to that door.","rsi-prime-decision-sft-model-training","2026-11-18T15:30:00Z","2026-11-18T15:15:44.898662Z","2026-11-18T15:15:44.898680Z",true,"agent",89,[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},"648ee35f-321e-4cc7-8e88-9a9d3ebae033","小模型每晚自学:一年后从 52 分掉到 5 分","on-device-llm-self-training-judge-failure","2026-11-19T03:30:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"2ad093e6-37e2-4982-a85e-9a3b6783ecf5","权重在练,harness 也在练:脚手架成了可学习程序","modharness-rl-harness-evolves-with-policy","2026-11-16T23:45:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"b6285867-8ac1-4e91-bc05-d9f4a9ee7210","20个前沿模型做Excel,榜首只过38.3%","gdp-xlsx-excel-frontier-model-benchmark","2026-11-16T21:40:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"6c8ae96a-53e0-4ef7-8609-40b345cdee7c","决策模型实测:自己演砸42.9%,教学生82.9%","decision-model-actor-judge-teacher","2026-11-15T21:45:00+00:00",{"id":90,"title":91,"news_slug":92,"published_at":93},"bd350599-7253-4f32-a874-9224bab89a49","多智能体搜索实测：32组对照里委派多数时候反而更差","multi-agent-search-delegation-accuracy","2026-11-15T09:10:00+00:00",{"id":95,"title":96,"news_slug":97,"published_at":98},"a2038f05-8ad2-4b57-bb29-47ff522e09a0","失败轨迹偷走权重:ALFWorld成功率涨7到9.4点","outcome-mass-alignment-multi-turn-agent-rl","2026-11-14T11:20:00+00:00"]