[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-opsd-privileged-solution-teacher-audit":3,"topics-all":38,"news-related-08c85d0c-1ad8-43d6-bd32-0ec8b21bf2aa":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},"08c85d0c-1ad8-43d6-bd32-0ec8b21bf2aa","给教师开小抄,学生涨分却不是抄来的","Meta 与 UCLA 团队审计了在策略自蒸馏:让教师模型偷看标准解法再逐 token 教学生。论文在 Qwen3 三个尺度上拆解这条链路,发现涨分真正来自教师的思考模式而非那份答案,并给出可复用的对照实验方法。","现在给推理模型做提升，有一条很主流的路子：让模型自己当自己的老师。\n\n具体做法叫 on-policy self-distillation（OPSD）。学生模型采样出一条推理轨迹，教师是学生的一份冻结副本，只是教师手里多了一张小抄——那道题的标准解法。学生没见过这张小抄，教师在学生生成的每个 token 位置上给出打分，学生照着这个逐 token 的信号更新。\n\n这套逻辑听上去很顺：教师看到了答案，学生理应学到更多。\n\n但 Meta 与 UCLA 的一篇论文把它拆开验了一遍，结论不太好看：教师偷看答案，并没有把答案教给学生。\n\n论文标题就是这句反直觉的判断——Privileged Information Does Not Guarantee Privileged Guidance，arXiv 编号 2611.16064，作者 Haoyi Qiu（Meta Superintelligence Labs 与加州大学洛杉矶分校，实习期间在 Meta 完成）和 Akshat Agarwal（Meta Superintelligence Labs），2026 年 11 月 11 日挂出 v1，论文自己的备注是「工作进行中」。全文见 https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.16064\n\n## 先看这套配方长什么样\n\n论文在 Qwen3 的 1.7B、4B、8B 三个尺度上做实验，评测用 AIME 2024、2025、2026 与 HMMT 2025 年 11 月、2026 年 2 月五套竞赛数学题。训练数据取 OpenThoughts 的数学推理子集，固定 3,200 条样本、跑 100 步更新、全局 batch 32，学生每题采样一条最长 1,024 token 的回复；教师与学生对同一条轨迹的前缀打分，损失用 forward-KL，每个词表项封顶 0.05，只更新学生的 LoRA 参数，框架用 veRL，跑在 NVIDIA H100 上，每个条件配三个随机种子。\n\n论文把这条链路拆成四段来审：Source（教师会不会用这份答案）、Encoding（教师给出的逐 token 目标有没有用）、Transmission（学生学没学到）、Expression（学生离开答案还能不能自己用出来）。\n\n## 第一段就漏了：教师拿着答案，仍然认不出学生的错\n\n先看 Source。论文从 Hendrycks MATH Level 5 里筛出学生 12 次采样全错的硬题，2,242 道清洗后的题里，1.7B、4B、8B 分别筛出 111 道（5.0%）、44 道（2.0%）、35 道（1.6%），题目与解答的有效性由 GPT-5.6-Sol 和 Claude Opus 4.8 判定。\n\n教师拿着答案去标学生的错误前缀，正确标出的比例只有 13.0% 到 25.5%，而且在 4B 和 8B 上，这个数字比不给答案的教师还低。两边判断都对的配对准确率在 6.5% 到 13.0% 之间。整体前缀准确率只有 1.7B 上那 +5.75pp 的提升区间排除了零，而且即便如此，准确率仍低于「全部接受」这个 50% 的基线。\n\n也就是：教师揣着标准解法，都认不太出学生错在哪一步。\n\n## 涨分来自哪里：思考模式，不是答案\n\n再往下是最扎眼的一段。论文把教师看答案（privileged）和教师只看题（problem-only）两条路摆在一起比：在 Thinking-OFF 评测下，Step 100 时，只看题的教师训练出来的学生，Avg@12 在三个尺度上都更高——1.7B 高 1.36pp，4B 高 15.21pp，8B 高 8.60pp（1.7B 的区间含零）。\n\n反复训练 16 个 epoch 后规律更明显：在 4B 和 8B 上，只看题的教师把 Thinking-OFF 的 Avg@12 抬高了 15.98 到 22.70 个百分点，而给答案的教师只抬了 5.68 到 10.66 个百分点；六组规模与题集对照里有四组是前者胜出。单轮训练时给答案的教师在 Step 20 还领先 1.02 到 1.91pp，到 Step 100 就被反超。\n\n论文把这个现象叫作 reasoning arbitrage——推理套利：涨的不是答案，是教师自己的思考模式本身。\n\n## 论文也没把话说死\n\n这篇的克制之处在于，它没有断言给答案没用。但反向的证据也在：Table 1 显示，给答案的教师在硬题上的 Avg@4 确实大幅领先——1.7B 是 40.54% 对 2.93%，4B 是 64.20% 对 1.14%，8B 是 58.57% 对 0.00%。论文引用的另一条路线 Purified OPSD 也确实有效：Step 100 时它在三个尺度上都比基座高 0.74 到 1.86pp，而标准 OPSD 反而比基座低 1.92 到 3.57pp；到了 Qwen3-4B 的代码任务上，Purified 比标准版在 LiveCodeBench v6 高 4.45pp、在 CodeContests 高 6.57pp。\n\n只是论文把话说准了：准确率涨了，不代表特权信息传过去了。教师能不能用好答案、给出的 token 目标有没有指向正确、学生学没学到、学生在没答案时用不用得出来——这四段必须分开测。\n\n另外几拨做数学特权蒸馏的团队，方向也一致。OP²SD 那篇（https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.09228）把参考解换成别的题的题面和解法，三个模型三个数学基准上照样超过基座、且与 OPSD 保持竞争力；Shrestha 与 Tessier（https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.18271）发现正确的参考解在不同生成模式、不同模型规模、不同数据集上并不一致地带来收益，换一道题的解法在好几个数学推理基准上反而能超过正确答案本身；Zhang 等人（https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.20612）用 5,319 道题的 AMPLE-Math 套件做匹配对照，发现参考无关的蒸馏已经能解释 Qwen3-1.7B 在开启思考评测下的很大一部分提升，完整推理链额外带来的好处只有在 SmolLM3-3B 的 step 50 上达到两个百分点。\n\n四篇论文作者互不相同，问的却是同一个：给教师开小抄，学生到底学到了什么。\n\n## 所以呢\n\n这篇论文对做模型的人有个很实际的提醒：你以为在优化「给模型更多信息」，实际可能在优化「让模型自己把自己的长处教给自己」。\n\n这对做评测的人也有直接提醒。以前看到一个方法涨了 5 分，很容易归因到它新增的那份数据、那个提示、那份答案。但这篇论文的做法值得抄：设计一个只差「教师有没有小抄」的对照，其余一切保持不变，再把涨分高的那一段单独拎出来测。\n\n这也是为什么它的标题用了一个那么硬的词——Guarantee。论文的结论是：不保证。有信息是一回事，信息真的被传下去是另一回事，这两件事之间隔着一个必须单独测的沟。\n\n对做应用的人也是：如果你今天在用「更强的教师 + 更狠的信号」这套东西训模型，先把教师那份小抄撤掉，跑一遍看还剩多少。剩下的那部分，才是你真正在训练的东西。\n","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.16064","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"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},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"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},"47a53141-86ea-4ad5-8201-96d83ed51680","en","Teacher Saw the Answer, Student Learned Something Else","A Meta and UCLA audit of on-policy self-distillation finds the teacher holding the solution does not teach it: gains track thinking mode, not answers.","There is one increasingly popular way to improve a reasoning model these days: let the model teach itself.\n\nThe recipe is called on-policy self-distillation, or OPSD. A student model samples a reasoning trajectory. The teacher is a frozen copy of that same student, except the teacher holds one extra card: the worked solution to the problem. The student never sees that card. The teacher scores the student's response prefix by prefix, and the student updates against those per-token targets.\n\nThe logic sounds airtight: the teacher saw the answer, so the student should learn more.\n\nA paper from Meta and UCLA tore that logic apart and reached an uncomfortable conclusion: the teacher peeking at the answer did not teach the answer to the student.\n\nThe title is the finding itself — Privileged Information Does Not Guarantee Privileged Guidance, arXiv 2611.16064, by Haoyi Qiu (Meta Superintelligence Labs and University of California, Los Angeles, work done during an internship at Meta) and Akshat Agarwal (Meta Superintelligence Labs). Version 1 appeared on 11 November 2026, and the paper's own comment field says it is work in progress. Full text: https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.16064\n\n## What the recipe actually looks like\n\nThe experiments run at three Qwen3 scales: 1.7B, 4B and 8B. Evaluation uses five competition-math benchmarks: AIME 2024, AIME 2025, AIME 2026, HMMT November 2025 and HMMT February 2026. Training data is the mathematical-reasoning subset of OpenThoughts, fixed at 3,200 examples for 100 updates at a global batch size of 32, with the student sampling one response of up to 1,024 tokens per problem. Teacher and student score the same sampled prefixes, the loss is forward-KL with each vocabulary term capped at 0.05, only the student's LoRA parameters are updated, the framework is veRL, and the runs sit on NVIDIA H100 GPUs with three seeds matched across conditions.\n\nThe audit splits the pipeline into four stages: Source (whether the teacher can actually use the solution), Encoding (whether the resulting token targets are useful), Transmission (whether the student learns from them) and Expression (whether the student applies the behaviour without the solution in hand).\n\n## Stage one leaks: with the answer in hand, the teacher still cannot spot the student's mistakes\n\nStart with Source. The paper screens Hendrycks MATH Level 5 for problems the student gets wrong in all 12 sampling attempts. Out of 2,242 cleaned questions, that yields 111 questions (5.0%) at 1.7B, 44 (2.0%) at 4B and 35 (1.6%) at 8B. GPT-5.6-Sol and Claude Opus 4.8 judge whether each problem and its worked solution are valid.\n\nGiven the privileged solution, the teacher correctly flags only between 13.0% and 25.5% of erroneous prefixes — and at 4B and 8B that rate is lower than when the teacher sees the problem alone. Pair accuracy, where both judgements must be right, ranges from 6.5% to 13.0%. For overall prefix accuracy, only the 1.7B gain of +5.75pp has a paired 95% interval excluding zero, and even there accuracy stays below the 50% baseline of accepting every prefix.\n\nIn other words: holding the reference solution, the teacher still cannot reliably identify which step the student got wrong.\n\n## Where the gains actually come from: thinking mode, not the answer\n\nThe next section is the most striking. The paper runs privileged-solution teachers and problem-only teachers side by side. Under Thinking-OFF evaluation at Step 100, students trained with a problem-only teacher score a higher mean Avg@12 at all three scales — 1.36pp higher at 1.7B, 15.21pp at 4B and 8.60pp at 8B, with the 1.7B interval including zero.\n\nRepeat training for 16 epochs and the pattern sharpens: at 4B and 8B, problem-only teachers raise Thinking-OFF Avg@12 by 15.98 to 22.70pp, versus 5.68 to 10.66pp for privileged-solution teachers, and the problem-only side wins four of six scale-and-question-set comparisons. In the one-pass runs the privileged-solution teacher leads by 1.02pp to 1.91pp at Step 20, but that relative advantage reverses by Step 100.\n\nThe paper names this pattern reasoning arbitrage. What drove the gains was not the solution. It was the teacher's own thinking mode.\n\n## The paper does not overclaim\n\nThe restraint here matters: it never concludes that privileged solutions are useless. The opposite evidence sits right next to it. Table 1 shows the privileged teacher taking a large lead on hard problems at Avg@4: 40.54% against 2.93% at 1.7B, 64.20% against 1.14% at 4B, 58.57% against 0.00% at 8B. And Purified OPSD, a method the paper uses as a matched control, genuinely works: at Step 100 it scores 0.74 to 1.86pp above Base across Qwen3 scales, while Standard OPSD sits 1.92 to 3.57pp below Base. On Qwen3-4B coding, Purified exceeds Standard by 4.45pp on LiveCodeBench v6 and 6.57pp on CodeContests.\n\nWhat the paper does insist on is precision: an accuracy gain does not establish that privileged information transferred. Whether the teacher can use the solution, whether its token targets point toward correctness, whether the student learns, and whether the student uses any of it without the solution present — these four have to be measured separately.\n\nThree other groups working on privileged distillation in mathematics reach compatible conclusions. The OP²SD paper (https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.09228) swaps the reference solution for the problem and solution of a different example; across three models and three mathematics benchmarks it still improves over the base model and stays competitive with OPSD. Shrestha and Tessier (https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.18271) find the correct reference gives no consistent performance benefit across teacher generation modes, model sizes and training datasets, and that another problem's solution can beat the correct one on several mathematical reasoning benchmarks. Zhang and colleagues (https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.20612) build the 5,319-problem AMPLE-Math suite with six reasoning views and matched reference-free distillation controls, finding that reference-free distillation already accounts for much of Qwen3-1.7B's improvement under thinking-enabled evaluation, with complete traces adding two percentage points only in SmolLM3-3B at step 50.\n\nFour sets of authors, none of them the same, asking the same question: when you give the teacher a cheat sheet, what exactly does the student learn?\n\n## So what\n\nFor anyone training models, the practical lesson is blunt: what looks like optimising more information may actually be optimising the model teaching its own strengths back to itself.\n\nThe bigger effect is on how results get attributed. When a method gains 5 points, the instinct is to credit the new dataset, the new prompt, the new solution. The design in this paper is worth copying: build a control that differs only in whether the teacher gets the cheat sheet, hold everything else fixed, then test each stage of the pipeline separately.\n\nThat is also why the title leans on such a hard word — Guarantee. The finding is that there is no guarantee. Having the information and actually transmitting it are two different things, and between them sits a gap you have to measure.\n\nFor practitioners the test is simple. If you are training with a stronger teacher and a harder signal, take the teacher's cheat sheet away, rerun, and see what is left. Whatever survives is what you were actually training.\n","opsd-privileged-solution-teacher-audit","2026-11-15T01:30:00Z","2026-11-15T01:08:19.567902Z","2026-11-15T01:08:19.567919Z",true,"agent",235,[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},"8e741f04-82a2-46da-9d30-3c31117ca70c","Ambient SFT:让预训练的自己当高温老师","ambient-sft-temperature-matching","2026-11-18T17:30:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"0be3d0c4-dee7-4fc2-9364-6a8ce68f5f3d","0.8B 模型换输出格式,静默错误率 10.3%→4.5%","output-format-silent-errors-small-parser","2026-11-13T09:20:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"a6d8e95f-41ba-4a97-815c-94eff9401706","按步数赢按算力输:GRPO密集奖励的评测账本","grpo-dense-rewards-matched-compute","2026-11-11T05:05:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"e8168e92-d16e-452a-991c-aac74dd23ec0","预训练赢了,微调输了:Muon 的反转","muon-adamw-fine-tuning-gap-reversal","2026-11-04T01:12:00+00:00",{"id":90,"title":91,"news_slug":92,"published_at":93},"bb244989-00b5-4b67-b70a-5316452ee44e","把 RLVR 的\"零梯度\"浪费变成有效信号：Mistral 联合剑桥用 Reasoning Arena 让推理训练提速 41%","reasoning-arena-mistral-cambridge-41pct","2026-06-08T00:00:00+00:00",{"id":95,"title":96,"news_slug":97,"published_at":98},"42f5ea83-365c-4a4e-a318-17d7c4c11c15","论文实测:自洽性证书在AIME上24%的题上会误触发","self-consistency-certificates-misfire-correlated-reasoning","2026-11-21T09:30:00+00:00"]