后训练自动化:AIBuildAI 用元搜索在 PostTrainBench 拿第一,46.6 分超过所有 Agent

Autonomous LLM Post-Training: AIBuildAI's Meta Search Tops PostTrainBench at 46.6, Beating Every Agent

Research #LLM#post-training#SFT#RLHF#DPO#GRPO#autonomous-training#meta-search#PostTrainBench#open-source
更新于
🇨🇳 中文

📌 开源仓库:https://github.com/aibuildai-inc/aibuildai-llm-posttrain-agent 技术报告:github.com/aibuildai-inc/aibuildai-llm-posttrain-agent/blob/main/docs/aibuildai-llm-post-train-agent.pdf License:Apache 2.0 | Stars:7 | Language:Python | 联系:[email protected]


后训练为什么难?

预训练好的语言模型不能直接用。Post-training(后训练)是让模型能听指令、能推理、会拒绝有害请求的那个阶段,它决定了模型在实际使用中是否有用。

但后训练的流水线设计成本高、门槛高。主要原因是决策之间强耦合

  • 用哪个偏好对齐算法(SFT → DPO / GRPO / PPO),取决于数据混合比
  • 数据混合比,取决于基座模型的特点
  • 学习率调度,取决于算法选择
  • 任何一个错,要从头重跑——H100 上一次完整训练就要花几小时

PostTrainBench 把这个问题形式化成了一个基准测试:给定基座模型 + 目标能力,Agent 必须在十小时预算内自主完成整个后训练流水线,结果由基准自己的隐藏测试集评分。


以往的思路是把 Claude Code 或 Codex 这类通用编程 Agent 对准任务跑。这能出结果,但有一个根本限制:Agent 在早期就提交了一个方案,后续所有决策都沿着这条线走,一旦卡住只会在同一条线上优化,不会放弃再换一个方向。

AIBuildAI 的核心贡献是元搜索——让搜索拓扑本身成为 Agent 的输出。

传统树搜索(每个节点是一次完整训练实验):

  • 所有任务走同一种树形结构
  • 无法表达「多阶段课程」「基于奖励模型的在线 RL」「多模型投票」等结构
  • 数据生成阶段与训练阶段捆在一起,不能复用前缀

元搜索

  • 元 Agent 先读任务、查知识库,然后写一个搜索程序
  • 搜索程序由三种原语组成:Agent(带工具的 LLM 会话)、Program(确定性计算)、Search(组合编排)
  • 可以生成 DAG、锦标赛、多路融合、迭代循环——不限于树
  • 支持「运行时重规划」:一阶段结束后,剩余预算和上一阶段结果交给新的元 Agent,它根据实测结果决定下一阶段的结构

每次运行的拓扑都是任务专属的,而不是固定预设。


知识系统:四个语料库通过 MCP 检索

元 Agent 和实验 Agent 在做设计决策时,都能查询一个远程检索服务(Kb),通过 MCP 接口访问,按语义嵌入检索。

四个子语料库

类别规模内容
工作流(Workflow)1 个技能文档,34 张参考卡后训练流程的 13 个顺序研究步骤,每步说明要做的判断、可选方向、不可迁移的情况
方法论(Methodology)104 张方法卡,30 个有代码实现每张卡含原始论文、数学推导、数值样例、各库默认参数、理论与实际成本、训练信号监控
数据集(Dataset)215 张数据集卡,13 个推荐,34 个须隔离每张卡含许可证、列定义、版本锁定加载代码、真实样本行、采纳证据、筛查记录
框架(Framework)108 张库卡,16 个推荐何时选这个库、如何启动训练、监控什么指标、如何保存可被加载的结果

每张技能文档同时记录「什么在什么条件下有效」和「什么在什么条件下失败」,Agent 可以从正负两个方向检索。


PostTrainBench 结果

七个任务:AIME 2025(竞赛数学)、GSM8K(数学)、ArenaHard Writing(创意写作)、GPQA Main(问答)、HealthBench(医疗建议)、HumanEval(代码生成)、BFCL(函数调用)

四个基座模型:Qwen3-1.7B-Base、Qwen3-4B-Base、SmolLM3-3B-Base、gemma-3-4b-pt

每次运行:单张 H100 GPU,10 小时上限,全自主,无人工介入,所有 Agent 角色跑在 Claude Opus 5 上

方法AIMEArenaHardBFCLGPQAGSM8KHealthBenchHumanEval总分
AIBuildAI(本文)15.865.595.831.482.342.669.046.6
Locus (Opus 5)9.466.394.333.182.944.366.645.6
Claude Code (Fable 5)13.361.573.028.383.540.158.441.8
Claude Code (Opus 5)9.253.41.531.480.338.759.835.0
Official instruct(参考基线)29.270.285.036.287.043.371.551.1

每格为四个基座模型的平均值;加粗为 Agent 类最高

三个领先任务的原因分析(来自报告):

BFCL(函数调用,95.8):任务核心是让模型输出符合 schema 的函数调用格式,是一个「输出契约」问题——学会了契约就接近满分,学不会就接近零分。元搜索能精准定位并针对这个契约设计流水线,而通用 Agent 容易绕开这一步。

HumanEval(代码生成,69.0):在全部四个基座模型上都领先。代码任务有明确的自动化验收标准(单元测试通过率),奖励信号清晰,适合迭代优化。

AIME 2025(竞赛数学,15.8):数学推理对格式要求严格(整数输出),且 SFT + GRPO 的叠加策略在小模型上效果差异很大。元搜索能根据中间结果调整策略,而固定拓扑在早期做了错误决策后难以纠正。

唯一在前面的是 Official instruct(51.1)——这是各模型官方开发团队发布的指令微调版本,属于「参考基线」而非竞争 Agent,使用的算力和时间预算不受约束。


和竞争方案的核心区别

Locus(外部提交,Intology,也跑在 Opus 5):总分 45.6,在 ArenaHard、GPQA、HealthBench 三个任务上领先。根据报告描述,Locus 是目前外部提交中最强的单一 Agent 方案。AIBuildAI 在这三个任务上与它的差距都在 2 分以内。

Claude Code + Fable 5(41.8):Claude Code 在 BFCL 上出现了 73.0 ± 28.0 的高标准差,说明在函数调用任务上不稳定。这与前文的分析一致——Claude Code 是通用代码 Agent,不针对后训练知识设计。

Claude Code + Opus 5(35.0):BFCL 只有 1.5 分,是明显的格式输出失败——说明同一模型(Opus 5)在不同架构下差距巨大。知识系统 + 元搜索的组合在这里发挥了关键作用。


技术细节:三个用树搜索无法表达的场景

报告明确列举了为什么树搜索不够,这三个场景展示了元搜索的必要性:

1. 数据生成阶段 树节点必须是已训练并评分的 pipeline,但数据语料只有在模型训练完成后才能评分。要比较 k 种数据生成方案,树搜索必须跑 k 次完整训练。元搜索可以把数据生成作为独立阶段,用廉价的代理指标(验证器通过率、去重率)筛选语料,再把好的语料传入训练阶段。

2. 多阶段课程学习 SFT → 偏好对齐 → 拒绝采样微调(Rejection Fine-Tuning)的顺序未知。树搜索每个候选都要从头训练共享前缀。元搜索可以缓存前缀,只搜索后续阶段。

3. 在线强化学习 奖励模型和策略模型需要在一次连续训练中交替更新(策略更新 → 奖励模型重拟合 → 继续训练)。树搜索只能在训练完成后评分,无法在训练中介入。元搜索可以把这表达成一个迭代循环。


产品定位与局限

AIBuildAI 的市场定位是:帮企业把专有数据转化为高性能定制 LLM,降低人力、时间和试错成本。

从技术报告来看,当前系统的实际限制:

  • 需要 Linux x86_64 + systemd + cgroup-v2,不能跑在 macOS 上
  • 每次运行需要单张 H100(十小时预算),硬件成本不低
  • 所有 Agent 运行在 Claude Opus 5 上,API 成本叠加在训练成本上
  • 仓库刚开放(2026-09-16),Stars 只有 7,代码成熟度待验证
  • PostTrainBench 上的结果是「当前快照」,报告明确说系统仍在开发中

技术路线本身是清晰的:元搜索 + 领域知识库 + 自动化实验 = 比通用 Agent 更高效的后训练探索。这与「人意图 + Agent 执行」的当前 AI 工程方向高度一致。


开源代码与模型仅供学习、勿直接用于工作。

© 2026 Author: Mycelium Protocol. 本文采用 CC BY 4.0 授权——欢迎转载和引用,须注明作者姓名及原文链接,不得去除署名后以原创发布。

🇬🇧 English

📌 Repository: https://github.com/aibuildai-inc/aibuildai-llm-posttrain-agent Technical report: github.com/aibuildai-inc/aibuildai-llm-posttrain-agent/blob/main/docs/aibuildai-llm-post-train-agent.pdf License: Apache 2.0 | Stars: 7 | Contact: [email protected]


Why Is Post-Training Hard?

A pretrained language model isn’t ready to use. Post-training is the stage that makes a model follow instructions, reason about its outputs, and decline harmful requests — it determines whether a model is actually useful in practice.

But designing post-training pipelines is expensive and expertise-intensive. The core reason: decisions are tightly coupled.

  • Which preference alignment algorithm to use (SFT → DPO / GRPO / PPO) depends on the data mixture
  • The effective data mixture depends on the base model’s characteristics
  • The learning rate schedule depends on the algorithm choice
  • Any mistake means a full rerun — one complete training run on an H100 takes hours

PostTrainBench formalizes this as a benchmark: given a base model and a target capability, an agent must autonomously complete the entire post-training pipeline within a ten-hour budget, scored by the benchmark’s own held-out evaluation.


The prior approach was to point a general coding agent (Claude Code, Codex) at the task and let it run. This produces results, but has a fundamental limitation: the agent commits to one approach early and every subsequent decision is conditioned on its existing trajectory — when it stalls, it refines the same line rather than abandoning it for something different.

AIBuildAI’s core contribution is meta search: making the search topology itself the output of an agent.

Traditional tree search (each node is one complete training experiment):

  • Every task goes through the same tree structure
  • Cannot express multi-stage curricula, online RL against a learned reward model, multi-model voting
  • Data generation and training are bundled — no prefix caching

Meta search:

  • A meta agent reads the task, consults the knowledge base, then writes a search program
  • The program is composed from three primitives: Agent (LLM session with tools), Program (deterministic computation), Search (composite orchestration)
  • Can generate DAGs, tournaments, fan-ins, iterative loops — not just trees
  • Supports “runtime replanning”: at each stage boundary, the remaining budget and stage results are handed to a new meta agent, which decides the next stage’s structure based on measured outcomes

Each run’s topology is task-specific rather than fixed in advance.


Knowledge System: Four Corpora via MCP

Both the meta agent and experimenter agents query a remote retrieval service (Kb) via MCP, using semantic embedding search when making design decisions.

CorpusScaleCoverage
Workflow1 skill doc, 34 reference cards13 ordered research actions per run, each explaining the judgment it settles, options available, and non-transferable cases
Methodology104 method cards, 30 with codeDefining paper, math, numerical example, per-library defaults, theory/practice cost, signals to watch during training
Dataset215 cards, 13 recommended, 34 held-outLicense, column definitions, version-pinned load line, real sample row, adoption evidence, screening record
Framework108 library cards, 16 recommendedWhen to pick, how to start, what to monitor, how to save a loadable result

Each skill document records both what works under what conditions, and what fails and why — both positive and negative evidence are available.


PostTrainBench Results

Seven tasks: AIME 2025 (competition math), GSM8K (math), ArenaHard Writing (creative writing), GPQA Main (QA), HealthBench (health advice), HumanEval (code generation), BFCL (function calling)

Four base models: Qwen3-1.7B-Base, Qwen3-4B-Base, SmolLM3-3B-Base, gemma-3-4b-pt

Each run: Single H100 GPU, 10-hour budget, fully autonomous, no human intervention; all agent roles running on Claude Opus 5

MethodAIMEArenaHardBFCLGPQAGSM8KHealthBenchHumanEvalOverall
AIBuildAI (ours)15.865.595.831.482.342.669.046.6
Locus (Opus 5)9.466.394.333.182.944.366.645.6
Claude Code (Fable 5)13.361.573.028.383.540.158.441.8
Claude Code (Opus 5)9.253.41.531.480.338.759.835.0
Official instruct (reference)29.270.285.036.287.043.371.551.1

Each cell is the average across four base models; bold = highest agent score

Analysis of three leading tasks:

BFCL (function calling, 95.8): The task is gated by a single output contract — the emitted call must parse and match the expected schema. Learn the contract, score high; miss it, score near zero. Meta search can precisely design a pipeline for this contract.

HumanEval (code generation, 69.0): Leads on all four base models. Code tasks have clear automated acceptance criteria (test pass rates), providing clean reward signals that reward iterative optimization.

AIME 2025 (competition math, 15.8): Math reasoning requires strict output formatting (integer answers), and the SFT + GRPO stacking strategy varies dramatically across small models. Meta search can adjust strategy based on intermediate results; fixed topologies can’t course-correct after early wrong decisions.


Core Technical Differences

Locus (external submission by Intology, also runs on Opus 5): 45.6 overall, leads on ArenaHard, GPQA, HealthBench. The current strongest single-agent external submission. AIBuildAI trails by less than 2 points on each of those three tasks.

Claude Code + Fable 5 (41.8): High variance on BFCL (73.0 ± 28.0 std dev), indicating inconsistent handling of function-calling format — Claude Code is a general coding agent, not designed around post-training domain knowledge.

Claude Code + Opus 5 (35.0): BFCL score of 1.5 — a near-complete formatting failure. The same underlying model (Opus 5) produces dramatically different results depending on the surrounding architecture.


Current Limitations

  • Requires Linux x86_64 + systemd + cgroup-v2 — no macOS support
  • Needs a single H100 per run (10-hour budget) — hardware costs are real
  • All agents run on Claude Opus 5 — API costs stack on top of training costs
  • Repository just opened (2026-09-16), 7 stars — production maturity is unknown
  • PostTrainBench results are described as a “snapshot of a system still under construction”

The technical direction is clear: meta search + domain knowledge base + automated experimentation = more efficient post-training exploration than general-purpose agents. This aligns directly with the “human intent + agent execution” pattern that defines current AI engineering practice.


Open-source code and models are for learning purposes only — do not use directly in production work.

© 2026 Author: Mycelium Protocol. Licensed under CC BY 4.0 — free to share and adapt with attribution.

💬 评论与讨论

使用 GitHub 账号登录后发表评论

关于本站 · 免责声明

🍄 Mushroom Research Blog 是非营利、免费公开的个人科技观察博客与公众号 XStack18,不接受商业合作、不代表任何企业或机构立场,也不谋求商业利益。我们以个人视角客观中立地记录和分析 AI、Web3 等领域的最新模型发布与技术动态——不止转述新闻标题或二手信息,而是给出有独立思考的深入分析,希望帮更多人获得有价值的一手科技认知。

⚠️ 文中介绍的开源代码与模型,仅供学习交流与技术借鉴。它们大多仍处于早期阶段,有待进一步研究和验证,请勿直接用于工作或生产环境;如需采用,请先自行充分测试,并核实其许可证与安全性。
Open-source code and models featured here are shared for learning and reference only. Most are early-stage and still need further study and verification — please don't use them directly in your work or in production. Test them thoroughly and check their licenses and security first.

  1. 本站文章均为作者基于公开信息的个人研究与观点整理,不代表文中提及的任何公司、产品、模型的官方立场,未与其构成商业关联或合作关系。
  2. 科技行业信息更新极快,我们尽力保证内容准确、及时,但不对完整性、实时性做绝对保证,具体请以相关企业/项目官方公告为准。
  3. 文中引用的第三方商标、产品名称、图片、数据等版权归原权利人所有,我们会尽量注明来源;如你认为存在版权疑问或侵权,请通过下方邮箱联系我们,收到通知后会尽快核实处理(更正、加注来源或删除)。
  4. 文章内容仅为技术科普与个人观点,不构成投资、法律或其他专业建议,据此进行任何决策的后果需自行判断和承担。

📮 侵权 / 勘误 / 合作咨询:[email protected]