[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-davinci-prover-agent-native-mid-training":3,"topics-all":38,"news-related-9531b4f7-3f4d-474c-b233-c715925a3e76":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},"9531b4f7-3f4d-474c-b233-c715925a3e76","8.4B轨迹压过25.3B技能语料:形式化证明中训配方","论文 daVinci-Prover 把「智能体该喂什么数据」算成一笔细账:33.7B token 的中间训练语料里,环境交互的智能体轨迹只有 8.4B,却在 MiniF2F 与 FATE-M 上压过 25.3B 的通用技能先验;代价是 48 张 H200 跑约 20 天,折合 23,040 个 GPU 小时。","把「智能体到底该喂什么数据」这笔账摊开算清楚,可能比再多刷高几个点更有用。\n\narXiv 上的 daVinci-Prover([论文](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.16433))做的是形式化定理证明:模型在隔离工作区里写 Lean 4 代码,编译器当场判对错,错了再改。它有意思的地方不是又一个刷榜模型,而是一组反常识的对照实验。\n\n反差在这条。按论文实测:33.7B token 的中间训练语料里,真正与环境交互相关的智能体轨迹只有 8.4B,但这 8.4B 在 MiniF2F 和 FATE-M 上的增益,比 25.3B token 的通用定理证明技能先验还大。另一笔账更贵:48 张 H200 跑约 20 天,折成公有云租赁是 122,000 到 146,000 美元。\n\n## 语料里装了什么\n\n语料分两块。一块是技能先验,覆盖证明分解与组装、带验证的子目标搜索、错误诊断与修订、自我验证;另一块是环境原生轨迹,包含真实的工具调用、执行结果、中间证明状态和 Lean 反馈。数据源是 NuminaMath-LEAN、AoPS、Lean Workbook、Goedel LM\u002FSFT_dataset_v2 和 OProofs,按 Qwen3.5 分词器算总量 33.7B token、样本 9.585M 条,其中回放又加进来 5.11B,预回放数据 28.6B。污染过滤也做了:2,137 道评测题里查出 29 道与语料重叠,占比 1.36%,最多折合 11.55M token,占整个语料的 0.03427%。\n\n教师模型是 DeepSeek-V4-Flash 和 DeepSeek-V4-Flash-0731。SFT 阶段另搭了一套 daVinci-Prover harness:输入定理原样写进 scratch\u002Fsolution.lean,智能体每回合从七个工具里挑(读写编辑文件、检索 Mathlib 声明、跑 Lean 检查),可写目录只有 scratch\u002F 和 sketch\u002F,target\u002F 只读,每次改 Lean 文件自动触发验证。它故意不暴露 bash 和 Python,为将来做 RL 留口子。轨迹用 DeepSeek-V4-Flash-0731 在 Goedel-LM\u002FRL_dataset_V2 的约 99K 道未验证题目上跑出来,上限 30 轮、200 次工具调用,留下约 36K 条高质量轨迹做 SFT,不足 5 轮的直接丢掉,超出 144K token 上下文的也不要。\n\n## 8.4B 打赢 25.3B\n\n消融在 Qwen3.5-9B 上做,基座本身是 MiniF2F 87.4、FATE-M 88.4、ProofNet 42.9。只喂技能先验(25.3B token)变成 87.8、89.8、43.6;只喂智能体轨迹(8.4B token)变成 88.8、92.2、43.5;两块合起来 33.7B token 则是 90.1、93.3、44.5。论文给的解释是分布对齐:推理时模型面对的就是工具调用、执行结果和验证反馈,训练时喂同构的东西,静态 next-token 训练和交互式证明之间那道裂缝就补上了。\n\n和已有语料横向比,他们用 33B token 的中间训练对上 Oprover 的 65B:同样只做 SFT 的 Qwen3-8B-Base 是 73.4、68.0、28.4,Oprover-8B-Base 加 SFT 是 80.3、73.1、30.3,这套配上是 85.3、85.3、40.4。论文自己的说法是,除了规模,配比本身同样关键。\n\n## 压缩上下文,比换参数划算\n\n还有个容易被忽略的开关:记忆压缩。做法很朴素,上下文到 200K token 时触发,把旧交互替换成抽取出来的工具证据,保留原始定理、当前 Lean 证明和最近几轮,最多压 8 次,不依赖模型自己总结。效果在更难的 FATE-H 上更明显:35B 模型从 16.3 涨到 42.0,9B 从 16.7 到 28.0;MiniF2F 上两个模型都到 97.1,论文在摘要里把这称作接近饱和。\n\n## 这笔账要花多少\n\n合成数据的成本论文算得很细:DeepSeek-V4-Flash-0731 跑了 48 张 NVIDIA H200、约 20 天,合计约 23,040 个 GPU 小时;Lean 验证服务用 8 组 worker、共 160 个 CPU 核和 640 GB 内存,约 76,800 个 CPU 核小时。按 Runpod 公布的 H200 两种规格单价(每 GPU 小时 5.29 美元和 6.35 美元),折算租下来大约 122,000 到 146,000 美元。论文强调这只是参考价,不是实际支出。\n\n## 我的判断\n\n值得抄的作业有两点。一是把「推理时会发生什么」提前写进训练语料,比多灌一半 token 更划算:同样 25.3B token 的技能先验,FATE-M 停在 89.8,换成 8.4B 的环境原生轨迹就是 92.2。二是 Lean 这种自带编译器的环境天然适合当智能体训练场,验证反馈免费、自动、可复现,比人造奖励靠谱。\n\n要泼的是冷水。论文自己在 Limitations 里承认,中间训练阶段没有回放任何泛化数据,并认为补上泛化回放「既必要也重要」,这意味着现在的增益大概率还没到头。另一条是数据口径必须分清:arXiv 论文里的表格用 Lean v4.26、上下文 200K、至多 30 轮交互,pass@1 与 pass@8 混着报;官方仓库([GAIR-NLP\u002FdaVinci-prover](https:\u002F\u002Fgithub.com\u002FGAIR-NLP\u002FdaVinci-prover))的 README 给的是另一套:Lean 4.35.0、每题 200 轮交互,底座换成 DeepSeek-V4.1-flash-1012,比的是同一个模型换不同 harness——daVinci-Prover 93.0,Claude Code 85.0,Codex 86.0,DeepSeek Harness 93.0;FATE-X 上 47.0 对 45.0、20.0、37.0;Putnam-200 上 87.5 对 86.0、86.0、88.5,同一列里 DeepSeek Harness 的 88.5 反而高出 daVinci-Prover 的 87.5。两套数字不能混着看。\n\n旁证来自另外两个团队。Moonshot 的 Kimi-Dev([论文](https:\u002F\u002Farxiv.org\u002Fabs\u002F2509.23045))用 agentless 训练产出技能先验,再拿 5k 条公开轨迹做 SFT,把 SWE-Agent 推到 48.6% pass@1,论文自陈与 Claude 3.5 Sonnet 相当——「先验打底加少量环境轨迹」这套配方,在代码仓库场景已经先跑过一遍。DeepSeek-Prover-V2([仓库](https:\u002F\u002Fgithub.com\u002Fdeepseek-ai\u002FDeepSeek-Prover-V2))走的是另一条路,用 DeepSeek-V3 递归分解子目标合成冷启动数据再上 RL,671B 模型在 MiniF2F-test 上做到 88.9% 通过率,PutnamBench 658 题解出 49 题;而 ProverBench 这个基准有 325 道题,15 道出自 AIME 24 与 AIME 25,另外 310 道取自教材与教程,这一点两个独立团队的说法能对上。\n","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.16433","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},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"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},"9d998b7a-f52b-468b-8213-968b17e5c364","en","8.4B Agent Trajectories Beat 25.3B Skill Tokens in Proving","Only 8.4B of a 33.7B-token mid-training corpus is agent trajectory data, yet that slice beats a 25.3B-token skill corpus on MiniF2F and FATE-M.","Writing down what an agent should actually be fed may matter more than squeezing out another point on a leaderboard.\n\nThe paper daVinci-Prover on arXiv (paper: https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.16433) works on formal theorem proving. The model writes Lean 4 code inside an isolated workspace, the compiler judges it on the spot, and the model revises when it fails. What is interesting is not another benchmark entry but a set of counterintuitive controlled comparisons.\n\nThe contrast is this. According to the paper's own measurements, out of a 33.7B-token mid-training corpus only 8.4B tokens are agent trajectories that involve real environment interaction, and those 8.4B tokens produce larger gains on MiniF2F and FATE-M than a 25.3B-token corpus of general theorem-proving skill priors. The other bill is heavier: 48 H200 GPUs for about 20 days, which works out to roughly 122,000 to 146,000 US dollars if rented from a public cloud.\n\n## What the corpus contains\n\nThe corpus has two parts. One is skill priors, covering proof decomposition and assembly, verified subgoal search, error diagnosis and revision, and self-verification. The other is environment-native trajectories containing real tool calls, execution outcomes, intermediate proof states and Lean feedback. The sources are NuminaMath-LEAN, AoPS, Lean Workbook, Goedel LM\u002FSFT_dataset_v2 and OProofs, totalling 33.7B tokens and 9.585M samples under the Qwen3.5 tokenizer, of which replay adds 5.11B and pre-replay data accounts for 28.6B. Decontamination was done too: 29 overlapping problems were found among 2,137 evaluation problems, a rate of 1.36%, worth at most 11.55M tokens, or 0.03427% of the whole corpus.\n\nThe teacher models are DeepSeek-V4-Flash and DeepSeek-V4-Flash-0731. For post-training the authors built a separate daVinci-Prover harness: the input theorem is copied verbatim into scratch\u002Fsolution.lean, and at each turn the agent picks from seven tools (reading, writing and editing files, retrieving Mathlib declarations, running Lean checks). Only scratch\u002F and sketch\u002F are writable, target\u002F is read-only, and every Lean file edit triggers verification automatically. Bash and Python are deliberately withheld to leave room for RL later. Trajectories were collected with DeepSeek-V4-Flash-0731 on roughly 99K unverified problems from Goedel-LM\u002FRL_dataset_V2, capped at 30 turns and 200 tool calls, leaving about 36K high-quality trajectories for SFT. Anything with fewer than 5 turns is dropped, and so is anything exceeding a 144K-token context.\n\n## How 8.4B beats 25.3B\n\nThe ablation runs on Qwen3.5-9B, where the base model sits at 87.4 on MiniF2F, 88.4 on FATE-M and 42.9 on ProofNet. Skill priors alone, at 25.3B tokens, give 87.8, 89.8 and 43.6. Agent trajectories alone, at 8.4B tokens, give 88.8, 92.2 and 43.5. Both together at 33.7B tokens give 90.1, 93.3 and 44.5. The paper's explanation is distribution alignment: at inference time the model faces tool calls, execution results and verifier feedback, so feeding the same structures during training closes the gap between static next-token training and interactive proving.\n\nAgainst existing corpora, the authors match a 33B-token mid-training run against Oprover's 65B. Qwen3-8B-Base with SFT only scores 73.4, 68.0 and 28.4; Oprover-8B-Base plus SFT scores 80.3, 73.1 and 30.3; the daVinci-Prover recipe scores 85.3, 85.3 and 40.4. The paper's own reading is that composition, not only scale, is what matters.\n\n## Compressing context beats swapping parameters\n\nOne more switch is easy to overlook: memory compaction. The method is plain. When the context reaches 200K tokens, compression triggers, older interactions are replaced by deterministically extracted tool evidence, and the original theorem, the current Lean proof and recent interactions are preserved. Up to 8 compactions are allowed, and nothing in the pipeline depends on a model summarising for itself. The effect is clearer on the harder FATE-H, where the 35B model goes from 16.3 to 42.0 and the 9B model from 16.7 to 28.0. On MiniF2F both models reach 97.1, which the abstract describes as near-saturating.\n\n## What the bill looks like\n\nThe paper is unusually specific about synthesis cost. DeepSeek-V4-Flash-0731 ran on 48 NVIDIA H200 GPUs for about 20 days, roughly 23,040 GPU-hours. The Lean verification service used about 8 worker instances with 160 CPU cores and 640 GB of memory in total, roughly 76,800 CPU core-hours. At the two H200 rates Runpod lists, 5.29 and 6.35 US dollars per GPU-hour, renting the equivalent deployment costs about 122,000 to 146,000 US dollars. The paper stresses this is a reference-price estimate rather than actual spend.\n\n## My take\n\nTwo things are worth copying. First, write down what will actually happen at inference time into the training corpus, because that buys more than another half of the tokens: with 25.3B tokens of skill priors FATE-M sits at 89.8, while swapping in 8.4B of environment-native trajectories moves it to 92.2. Second, an environment that ships its own compiler, like Lean, is naturally a good agent training ground, since the verification feedback is free, automatic and reproducible, which is better than hand-made reward.\n\nNow the cold water. The paper itself admits in its limitations section that no generalization data is replayed during mid-training, and argues that adding generalization replay is both necessary and important, so the current gains probably are not the ceiling. The other thing is that two different sets of numbers must not be mixed. The tables in the arXiv paper use Lean v4.26, a 200K context and at most 30 interactions, reporting pass@1 and pass@8 side by side. The README in the official repository (https:\u002F\u002Fgithub.com\u002FGAIR-NLP\u002FdaVinci-prover) ships a different setup: Lean 4.35.0, 200 interaction turns per problem, and a different backbone in DeepSeek-V4.1-flash-1012, so what it compares is harnesses rather than models. On that harness comparison daVinci-Prover scores 93.0, Claude Code 85.0, Codex 86.0 and DeepSeek Harness 93.0. On FATE-X the numbers are 47.0 against 45.0, 20.0 and 37.0. On Putnam-200 they are 87.5 against 86.0, 86.0 and 88.5, where the DeepSeek Harness column is actually 88.5, above daVinci-Prover at 87.5. Two sets of numbers should not be read as one.\n\nCorroboration comes from two other teams. Moonshot's Kimi-Dev (paper: https:\u002F\u002Farxiv.org\u002Fabs\u002F2509.23045) derives skill priors from agentless training and then adapts on 5k publicly available trajectories, pushing SWE-Agents to 48.6% pass@1, which the paper puts on par with Claude 3.5 Sonnet, so the recipe of prior plus a small amount of environment trajectory has already run in the code repository setting. DeepSeek-Prover-V2 (repository: https:\u002F\u002Fgithub.com\u002Fdeepseek-ai\u002FDeepSeek-Prover-V2) takes a different route, using DeepSeek-V3 to recursively decompose subgoals into cold-start data before reinforcement learning, and its 671B model reaches an 88.9% pass ratio on MiniF2F-test while solving 49 of 658 PutnamBench problems. On the benchmark spec itself, ProverBench has 325 problems, 15 from AIME 24 and AIME 25 and 310 drawn from textbooks and tutorials, and two independent teams agree on that number.\n","davinci-prover-agent-native-mid-training","2026-11-14T07:20:00Z","2026-11-14T07:16:36.133845Z","2026-11-14T07:16:36.133859Z",true,"agent",128,[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},"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":65,"title":66,"news_slug":67,"published_at":68},"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":70,"title":71,"news_slug":72,"published_at":73},"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",{"id":75,"title":76,"news_slug":77,"published_at":78},"343f6cae-ae14-4617-a317-7316748e1d6e","删掉推理链再转发:同档32B模型差51分","deletion-sensitivity-reasoning-trace-routing","2026-11-13T07:05:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"365a0af2-d564-4eff-bf03-a138cdea0c5f","研究智能体大考:只查交付过69%,查全要求剩28%","are-bench-research-agent-context-requirements","2026-11-13T01:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"f2723003-43ac-47fe-bb59-49030446919a","28个决策模型大考:多选最高才36.81%","mmdm-decision-model-benchmark","2026-11-12T13:30:00+00:00"]