[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-one-shot-opd-single-query-distillation":3,"topics-all":38,"news-related-005557c5-8a3c-4d34-89bc-35d5351c4570":48},{"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},"005557c5-8a3c-4d34-89bc-35d5351c4570","蒸馏只需要一条训练样本?清华实测:单条query覆盖71.5%训练状态,16条追平17k全量","清华 Thinking-Space 团队把 on-policy 蒸馏压到数据极限:只用一条训练样本,数百步内持续涨点,覆盖 71.5% 全量训练状态;16 条语义多样 query 覆盖 98.9% 并追平约 17k 全量。结论:OPD 数据过量、算法饥饿,瓶颈在学生吸收监督的速率而非数据规模。","做蒸馏的人大多默认一个前提:训练数据越多越好。9 月 3 日上传 arXiv 的论文 Rethinking On-Policy Distillation II: One Training Example 把这个前提推到极限——on-policy 蒸馏只用一条训练样本,学生模型照样持续涨分。数字摆出来之后,「攒数据」这件事的边际价值需要重新掂量。\n\n## 一条 query 能走到多远\n\n先补背景:on-policy 蒸馏(OPD)的组合方式,是学生模型自己生成 rollout,再由教师在 token 级别提供密集监督。论文自述,已有工作主要在研究它的算法行为,训练数据扮演什么角色一直不清楚。Thinking-Space 团队(README 联系方式为清华邮箱)的做法是把数据压到最小极限:只用一条 query 训练。\n\n结果有三层。第一,单样本 OPD 在数百步内持续改进,并在不同任务域和模型家族上都找回了全量数据 OPD 增益的大头。第二,团队定义了「状态覆盖率」——全量 OPD 训练中到访过的状态里,当前 query 集的 rollout 能触达的比例——单条 query 就达到 71.5%,而且大头在前 100 步内完成。第三,继续加语义多样的 query,状态覆盖率和验证准确率一起上升,16 条时覆盖率达到 98.9%,追平约 17k query 的全量训练。\n\n## 数据过量,算法饥饿\n\n更关键的对照是吸收速率:不管用一条 query 还是全部 17k,学生对齐教师的速度都以相似节奏放缓;甚至固定住同一批状态,学生也要数百步才能消化。论文给出一句话结论:OPD 是「数据过量、算法饥饿」(data-overfed but algorithm-starved)——rollout 很快铺开广泛的监督信号,学生吸收监督却越来越慢。\n\n两个延伸发现。其一,状态覆盖结论可以推广到多教师 MOPD:每个域 16 条语义多样的 query 即可追平全量多教师训练。其二,压力测试里,内容近乎空的模板、域外 WildChat query 都能逼近真实 query 的基线——任务内容和它诱导的状态覆盖,是可以分离的。仓库里甚至留了 template 模式,让学生自己写训练输入。\n\n## 开源与复现门槛\n\n代码开源于 GitHub(Thinking-Space\u002FOne-Shot-OPD,Apache 2.0),基于 veRL 扩展实现,论文参考数字来自单台 8 卡 H100\u002FA100 80GB 节点。评测覆盖四个域:数学(MATH-500、AMC 2023、AIME 2025,avg@16)、代码(LiveCodeBench v6,avg@3)、指令跟随(Multi-IF 八语言)、工具调用(BFCL v3,avg@8)。这是系列第二篇,前作今年 4 月上传,团队希望这些发现把后续工作引向 OPD 的步效率问题。\n\n## 所以呢\n\n对做后训练的团队,启示不是「以后蒸馏只用一条样本」,而是投入结构的再分配:数据工程的边际收益比想象中低,怎么让学生更快吸收监督(步效率)才是下一阶段的竞争点。和月初普渡大学那篇「固定负优势就能追平教师」的发现放在一起看,「OPD 到底在学什么」正在被集中重审。论文与代码:arXiv:2609.04172,github.com\u002FThinking-Space\u002FOne-Shot-OPD。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.04172","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","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},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"09840724-b021-492f-8914-79d7bd386847","en","One query reaches 71.5% state coverage in on-policy distillation","One query reaches 71.5% state coverage; 16 reach 98.9% and match full-data OPD. The team's verdict: OPD is data-overfed but algorithm-starved.","Distillation practice assumes more training data is better. A paper uploaded to arXiv on Sep 3 — Rethinking On-Policy Distillation II: One Training Example — pushes that assumption to the limit: on-policy distillation trained on a single example keeps improving. Once the numbers are on the table, the marginal value of \"collecting more data\" needs a second look.\n\n## How far one query goes\n\nBackground first: on-policy distillation (OPD) combines student-generated rollouts with dense token-level supervision from a teacher. The authors state that existing work has mainly studied its algorithmic behavior, leaving the role of training data unclear. The Thinking-Space team (README contacts are tsinghua.edu.cn addresses) took the data-minimal limit literally: train on a single query.\n\nThree results. First, one-shot OPD keeps improving for hundreds of steps and recovers most of full-data OPD's gain across task domains and model families. Second, the paper defines state coverage — the fraction of the states full-data OPD visits that a query set's rollouts reach — and a single query already hits 71.5%, with most of it within the first 100 steps. Third, adding semantically distinct queries raises coverage and validation accuracy together, until 16 queries reach 98.9% and match full-data training on roughly 17k queries.\n\n## Data-overfed, algorithm-starved\n\nThe sharper finding is absorption rate: alignment with the teacher slows at a similar pace whether OPD trains on one query or all 17k, and even a fixed set of states takes hundreds of steps to absorb. The paper's one-line verdict: OPD is data-overfed but algorithm-starved — rollouts quickly expose broad supervision, while the student absorbs it increasingly slowly.\n\nTwo extensions. The state-coverage result carries over to multi-teacher MOPD, where 16 semantically diverse queries per domain match full-data MOPD. And in stress tests, content-light templates and off-domain WildChat queries also approach the real-query baseline: task content and induced state coverage can come apart. The repo even ships a template mode where the student writes its own training input.\n\n## Open source and reproduction\n\nCode is on GitHub (Thinking-Space\u002FOne-Shot-OPD, Apache 2.0), implemented by extending veRL; reference numbers come from a single 8-GPU node with H100\u002FA100 80GB. Evaluation spans four domains: math (MATH-500, AMC 2023, AIME 2025, avg@16), code (LiveCodeBench v6, avg@3), instruction following (Multi-IF across eight languages), and tool use (BFCL v3, avg@8). This is Part II of a series; Part I went up in April, and the team hopes these findings direct future work toward the step efficiency of OPD.\n\n## So what\n\nFor post-training teams the lesson is not \"distill on one example from now on\" but a reallocation of effort: the marginal return on data engineering is lower than assumed, and how fast a student absorbs supervision is the next battleground. Read alongside Purdue's early-September finding that fixed negative advantages can match the teacher, the question \"what is OPD actually learning\" is getting re-examined from several directions at once. Paper and code: arxiv.org\u002Fabs\u002F2609.04172, github.com\u002FThinking-Space\u002FOne-Shot-OPD.","one-shot-opd-single-query-distillation","2026-09-05T21:07:11Z","2026-09-05T21:07:19.392684Z","2026-09-05T21:07:19.392701Z",true,"agent",76,[39],{"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},"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":49},[50,55,60,65,70,75],{"id":51,"title":52,"news_slug":53,"published_at":54},"58ed753e-ad6d-4aac-95f4-36bf217e169c","把 10 万条人类视频变成机器人教材:RoboTok 检索 mAP 提升约 50 倍,hard 任务 79.3% 对 19.5%","robotok-retrieval-benchmark-reread","2026-09-06T21:11:25+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"7623f190-7071-4811-a6f1-32462a99b8d3","经验会过期:阿里云论文让自主后训练的有害授权率从 62.5% 降到 25%","bcit-conditional-experience-transfer-post-training","2026-09-05T17:11:11+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"4a89fe5a-8703-49e5-b083-079cbda0fa2a","蒸馏也有副作用:中间训练期上KD,推理上涨、事实记忆反而变慢","switch-distillation-midtraining-kd","2026-09-02T17:10:00+00:00",{"id":66,"title":67,"news_slug":68,"published_at":69},"3c9e4d7f-6f2a-4fea-8d22-c351b8fd7a4a","IBM Granite 4.1：Dense架构回归，8B参数挑战32B MoE性能","ibm-granite-4-1-dense-8b-moe-32b-grc","2026-04-29T19:10:00+00:00",{"id":71,"title":72,"news_slug":73,"published_at":74},"e84fe968-5d86-4247-baad-5da23efef860","UltraData-RL-2609 开源:85,995 条可验证奖励任务,拆解 MiniCPM5-2B 的 RL 燃料","ultradata-rl-2609-verifiable-rl-dataset","2026-09-07T23:07:45+00:00",{"id":76,"title":77,"news_slug":78,"published_at":79},"0c29e1ad-914a-4b79-a153-445c087acb03","被 LLM 抛弃的 dropout 翻身:Cerebras 称调好可省 25% 训练 FLOPs","dont-drop-dropout-layer-sparsity","2026-09-07T21:06:35+00:00"]