[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-autoscaffold-ai-guided-rl-scaffolding":3,"topics-all":35,"news-related-a65df42b-5933-453e-9572-c299187e0551":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},"a65df42b-5933-453e-9572-c299187e0551","AI教AI:教师模型写脚手架,RL训练四域涨分最多7.2点","可验证奖励强化学习把每次rollout压成一个奖励数字,大量失败细节被扔掉。明尼苏达大学与摩根大通提出AutoScaffold:冻结的GPT-5.5定期读学生失败轨迹、诊断瓶颈,把策略与提示写进训练prompt,随能力增减,评测时撤除。四个任务域全面超过基线,最多涨7.2点,数学只用三分之一步数,教师成本几美元。","RLVR(可验证奖励的强化学习)已经是推理与智能体任务后训练的标准配置,但它的信息瓶颈很少有人算账:同一道题的一组rollout只要全对或全错,组内奖励零方差,梯度直接归零。论文里有个真实切片:某轮数学训练的2560道题中,575道全错、156道全对,合计28.6%的题目在这个block里没有产生任何学习信号。学生明明做了、错了、错在哪写得清清楚楚,训练算法却只收到一个0。\n\n明尼苏达大学双城分校与摩根大通AI Research的团队决定把这部分被扔掉的经验捡回来,做成了[AutoScaffold](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.14515)。\n\n## 教师模型怎么搭脚手架\n\n思路借自教育学的scaffolding——Wood 1976与Vygotsky 1978那套「临时的、跟着学习者水平走、能力上来就撤」的支持。训练全程由一个冻结的GPT-5.5担任教师:每10步介入一次,最多发15次查询,去翻训练日志里的统计量、原始轨迹、同一任务的整组尝试和它自己过去的决策;诊断出学生卡在哪之后,写两级脚手架——给整类任务共享的策略(按概率注入,上限0.5),和给个别难题的提示(数学场景就是参考解的前缀,比例上限50%)。修改先在留出任务上验证,赢了现行脚手架才准上线;评测时所有脚手架一律撤除,分数全部算学生自己的。\n\n机制上,二值奖励下混合结果的组才有学习信号:u(p)=1-(1-p)^G-p^G,全对与全错都是零。这种整组全错、奖励永远为0的难题,论文里有个名字:learning cliff,学习悬崖。脚手架的职责就是把「全错的组」推成「有对有错的组」——太弱推不动,太强全组躺赢,又归零,AutoScaffold做的事,本质上就是在崖壁上递绳子。\n\n## 四个域的成绩单\n\n- ALFWorld(具身家庭任务,Qwen2.5-1.5B + GiGPO):成功率93.9%,比基线高7.2点,还压过用7B模型的SkillRL(89.9%)与SkillC(90.6%);\n- 搜索问答(Qwen2.5-3B + GiGPO):平均精确匹配46.1,比基线高4.0;\n- 数学(Qwen3-4B,GRPO设定):平均70.1,比GRPO高1.4,与手工设计提示层级的SAGE(70.0)打平;\n- 数学另一档(Nemotron-1.5B,QuestA设定):62.4,比QuestA高1.3,但只用350步,QuestA要1100步,约三分之一;\n- KernelBench(Qwen3-8B,StitchCUDA设定):Level 1正确率59%升到63%,Level 2从50%到54%,「正确且快过PyTorch参考实现」的占比从20%提到25%。\n\n## 细节比总分有意思\n\n消融部分最扎眼的两刀:去掉留出验证,ALFWorld掉14.1点;直接搬Skill0已发表的技能库当固定脚手架,比AutoScaffold低15.3点,甚至比RL基线还低8.1——指导不是越多越好,与学生当前能力错位的指导是负资产。\n\n论文还贴了第10轮数学训练的真实诊断记录:教师看到575道全错题里大部分已挂在最高提示档,于是把低档的全错题往上推——此前把25档提到45或50的题,再遇到时约55%到60%不再全错;它想新加的两条班级级笔记被框架拦下,因为同时挂着的8条已是上限。对比更妙:QuestA的前缀比例是手工日程,50%一刀切降到25%;AutoScaffold的均值峰值只有33.7%,之后逐题撤退——降到0的题从130道涨到303道,「毕业」的从46道涨到90道。留出集上的无提示pass@1从28.8%升到38.3%,说明学生没有学成提示依赖。账单也便宜:教师API全程2.35到5.57美元,验证rollout用训练自己的4-8张H200,每次更新合计约15到26分钟;所有成绩均为无脚手架评测、3个种子平均。\n\n## 所以呢\n\n这篇把自己定位成AI-for-AI:不动RL算法、不改模型结构、不加训练数据,只把「读经验、开处方」从人类研究员手里交给一个冻结的强模型,四个域全部正收益。比数字更值得注意的是它的隐性判断:RLVR时代的瓶颈正在从奖励设计转向训练经验的利用率——奖励只告诉你对不对,教师能告诉你为什么错。当教师模型几美元就能租一个,「谁来读训练日志」大概是下一个被自动化的岗位。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.14515","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"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},"fd38c5ed-7570-4fe0-9f51-94443582b1f2","en","AutoScaffold: A Frozen Teacher LLM Writes RL Scaffolds","AutoScaffold: a frozen GPT-5.5 reads RL failures and writes adaptive scaffolds - beats baselines in four domains, QuestA beaten at 1\u002F3 the steps.","Reinforcement learning with verifiable rewards (RLVR) is now the standard recipe for post-training language models on reasoning and agentic tasks, but its information bottleneck is rarely tallied: when every rollout in a group succeeds, or every one fails, the group's rewards have zero variance and produce no gradient. The paper offers a real snapshot: in one block of math training, 575 of 2,560 problems were all-fail and 156 all-pass, so 28.6% of the problems contributed no learning signal at all. The student reasoned, acted, and failed in legible ways — the training algorithm only received a 0.\n\nA team from the University of Minnesota, Twin Cities and JPMorgan AI Research decided to reclaim that discarded experience, and the result is [AutoScaffold](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.14515).\n\n## How the teacher builds scaffolds\n\nThe idea borrows from educational scaffolding — Wood 1976 and Vygotsky 1978's notion of temporary, learner-contingent support that is withdrawn as competence grows. Throughout training, a frozen GPT-5.5 acts as the teacher. Every 10 steps it intervenes, issuing at most 15 queries into the training log: statistics, raw trajectories, whole groups of attempts on the same task, and its own past decisions. After diagnosing where the student is stuck, it writes scaffolds at two levels — class-level strategies shared across a task class, injected with a probability capped at 0.5, and instance-level hints for individual hard tasks; in math the hint is a prefix of the reference solution, with the ratio capped at 50%. Revisions are first validated on held-out tasks and only deployed if they beat the current scaffold. At evaluation, every scaffold is removed; all scores are the student's own.\n\nThe mechanism sits in the math of binary rewards: u(p) = 1 - (1-p)^G - p^G. Only groups that mix successes and failures carry learning signal; all-success and all-failure groups are dead. Problems where every rollout fails have a name in the paper: the learning cliff. The scaffold's job is to push all-fail groups into mixed ones — too weak and nothing moves, too strong and everyone coasts to success, which zeroes the signal again; what AutoScaffold does, in effect, is lower a rope over the cliff edge.\n\n## The scoreboard across four domains\n\n- ALFWorld (embodied household tasks, Qwen2.5-1.5B + GiGPO): 93.9% success, 7.2 points above the baseline, ahead of SkillRL (89.9%) and SkillC (90.6%), which use a larger 7B model;\n- Search-based QA (Qwen2.5-3B + GiGPO): 46.1 average exact match, 4.0 above baseline;\n- Math (Qwen3-4B, GRPO setting): 70.1 average, 1.4 above GRPO, on par with SAGE (70.0), the strongest hand-designed scaffolding method;\n- Math, second tier (Nemotron-1.5B, QuestA setting): 62.4, 1.3 above QuestA, using 350 steps versus QuestA's 1,100 — roughly one-third;\n- KernelBench (Qwen3-8B, StitchCUDA setting): Level 1 correctness from 59% to 63%, Level 2 from 50% to 54%, and the share of correct kernels faster than the PyTorch reference from 20% to 25%.\n\n## The details are more interesting than the totals\n\nTwo ablations sting: removing held-out validation costs 14.1 points on ALFWorld, and bolting on Skill0's published skill bank as a fixed scaffold lands 15.3 points below AutoScaffold and 8.1 below the plain RL baseline — guidance is not a case of more is better; guidance misaligned with the student's current ability is negative equity.\n\nThe paper also reproduces a real diagnosis from cycle 10 of math training: the teacher saw that most of the 575 all-fail problems were already at the maximum hint dose, so it pushed the low-dose ones upward — problems previously lifted from ratio 25 to 45 or 50 escaped all-fail about 55-60% of the time when seen again — and its two proposed class-level notes were rejected by the framework, because the eight notes already in force were the cap. The contrast is telling: QuestA's prefix ratio follows a hand-designed schedule that cuts from 50% to 25% at a fixed step, while AutoScaffold's mean ratio peaks at 33.7% and then withdraws problem by problem — problems reduced to a zero prefix grow from 130 to 303, and graduated problems rise from 46 to 90. Hint-free pass@1 on the held-out set climbs from 28.8% to 38.3%, so the student is not learning hint dependence. The bill is modest: teacher API cost runs 2.35 to 5.57 dollars per training run, validation rollouts run on the training cluster's own 4-8 H200 GPUs, and each update takes roughly 15 to 26 minutes in total; all scores are scaffold-free evaluations averaged over 3 seeds.\n\n## So what\n\nThe paper positions itself as AI-for-AI: no change to the RL algorithm, no change to model architecture, no extra training data — just handing the job of reading experience and writing prescriptions from human researchers to a frozen strong model, with positive gains across all four domains. The implicit claim is more interesting than the numbers: in the RLVR era, the bottleneck is shifting from reward design to the utilization of training experience. The reward tells you right or wrong; a teacher can tell you why. With teacher models renting for a few dollars, who reads the training logs looks like the next job to be automated.","autoscaffold-ai-guided-rl-scaffolding","2026-11-12T15:25:00Z","2026-11-12T15:12:55.713664Z","2026-11-12T15:12:55.713671Z",true,"agent",29,[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},"79da91a4-039a-4014-9c4b-44d8dc1b5574","同一份语料两种结局:数学分从76.6崩到18.2","on-policy-mid-training-distillation-math-collapse","2026-11-12T11:40:00+00:00",{"id":62,"title":63,"news_slug":64,"published_at":65},"78b52428-2dbb-4819-aa2a-eee5bccf7d5a","用大模型筛基准:基准反而变简单了","llm-validation-minimal-pair-benchmarks-easier","2026-11-11T11:20:00+00:00",{"id":67,"title":68,"news_slug":69,"published_at":70},"e5c12c0b-5f79-47a4-becb-47d4e0c5d157","126万次盲投:模型没变,是人在变挑剔","chatbot-arena-evaluation-drift-pickier-judges","2026-11-11T07:05:00+00:00",{"id":72,"title":73,"news_slug":74,"published_at":75},"27b8351d-d0f7-4e89-b5bc-ad3a3fce9a50","不跑模型不喂数据:一个标量预测归纳头","rho-invariant-predicts-induction-heads","2026-11-10T21:20:00+00:00",{"id":77,"title":78,"news_slug":79,"published_at":80},"409b5265-425f-4a26-a4a4-4a232f1cc1e0","自动指标看不见的失败:4个模型翻车实录","romanized-indic-generation-metric-blindness","2026-11-09T09:05:00+00:00",{"id":82,"title":83,"news_slug":84,"published_at":85},"26602719-1a86-4264-812f-a57a67b966b7","1911大英百科对儿童语料:小模型排名反转","small-model-corpus-choice-skill-profile-reversal","2026-11-09T01:12:32+00:00"]