Stanford CS329A：《自我改进 AI Agent》，斯坦福研究生研讨课，9 集视频全公开
Stanford CS329A: Self-Improving AI Agents — 9 Lectures Now Free on YouTube
📌 课程官网：https://cs329a.stanford.edu/ GitHub 组织：https://github.com/stanford-cs329a YouTube 全集（9集）：https://www.youtube.com/playlist?list=PLangBM27OtEA 社区中文转录：https://github.com/flowioo/stanford-cs329a-self-improving-ai-agents
斯坦福 2025 年秋季开了一门研究生研讨课：CS329A「Self-Improving AI Agents」。授课人是 Azalia Mirhoseini 和 Aakanksha Chowdhery，两位都有 Google DeepMind/Brain 背景。
课程 9 集视频在 2026 年 8 月全部公开上线，可以免费看。作业仓库也在 GitHub 开放。
这门课在讲什么
一句话概括：Agent 如何利用自己的执行经验，持续改进自身能力。
课程的核心论点是：
“推理时算力，制造下一个模型所需的训练数据。” Test-time compute manufactures the training data that improves the next model.
这不只是一个优化技巧，而是对当前 AI 能力演进路径的一个结构性判断——模型在推理时产生的轨迹数据，正在成为训练更好模型的原材料。
课程覆盖的核心技术
自我改进机制：
| 技术 | 作用 |
|---|---|
| Constitutional AI | 模型根据原则对自身输出进行自我批评和修正 |
| STaR（Self-Taught Reasoner） | 用模型自己生成的推理链条作为训练数据 |
| DAPO | 通过强化学习在训练阶段实现自我改进 |
| 领域特定 verifier | 用可验证的外部信号（代码运行结果、数学验证）提供精准奖励 |
推理时扩展：
| 技术 | 作用 |
|---|---|
| Test-time compute scaling | 推理阶段投入更多算力换取更好的输出 |
| Search-o1 类方法 | 把搜索与 LLM 结合，让模型在推理时主动查找信息 |
| Multi-step reasoning | 多步规划和执行，而不是单步生成 |
工具与记忆：
| 技术 | 作用 |
|---|---|
| ReAct | 将推理和行动交替进行，工具调用的基础框架 |
| MemGPT | 超出上下文窗口的长期记忆管理 |
| Code execution | 让模型写代码并执行，用结果验证推理 |
应用领域：coding agents（代码生成和修复）、STEM 研究助手、机器人控制。
论文清单
课程公布了完整的论文阅读清单，覆盖：
- ReAct（Yao et al. 2022）——工具调用 Agent 的基础
- Constitutional AI（Anthropic）——自我对齐与自我批评
- STaR（Zelikman et al. 2022）——自我生成推理链作为训练数据
- DAPO——基于 RL 的自我改进
- MemGPT——外部记忆管理
- Search-o1——搜索增强推理
- AlphaCode——代码生成的大规模 RL 训练
这份清单是独立学习的一个有价值的入口——按课程顺序读完这些论文，可以建立起 Agent 自我改进方向的完整知识图。
开放资源
YouTube 全集（9 集，Autumn 2025）： 课程 9 集完整视频在 2026 年 8 月上线，全部免费。内容包括每个主题的讲授，格式是研讨课风格，不是传统讲座。
https://www.youtube.com/playlist?list=PLangBM27OtEA
GitHub 组织： https://github.com/stanford-cs329a
有 6 个仓库，含 Fall 2025 和上一学期的作业题目，可以直接 clone 做练习。
课程网站： https://cs329a.stanford.edu/
列出了完整的课程日历和每节课对应的论文，是结构化论文阅读的导航地图。没有公开 slides 或讲义，但论文链接完整。
社区中文转录： https://github.com/flowioo/stanford-cs329a-self-improving-ai-agents
社区整理的 9 集视频转录 + 中文导读，对不适应英文听力速度的学习者更友好。
和 CS329Z 的区别
斯坦福还有另一门相近的课 CS329Z「AI Agent Engineering」，两者容易混淆：
| 维度 | CS329A（本课） | CS329Z |
|---|---|---|
| 核心问题 | Agent 如何持续自我改进 | 如何工程化地构建和部署 Agent |
| 侧重 | 自我改进机制、强化学习、推理时扩展 | 系统设计、可靠性、部署、工具链 |
| 受众 | 研究向 | 工程向 |
CS329A 是在问”Agent 能变多好”，CS329Z 是在问”怎么把 Agent 做出来”。
适合谁
- 想系统理解 Agent 自我改进方向的研究者或工程师
- 想了解 test-time compute scaling 为什么重要的人
- 需要一份结构化论文阅读清单的人
9 集视频加上论文清单，是目前这个方向上结构最完整的公开材料之一。
© 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。
📌 Course website: https://cs329a.stanford.edu/ GitHub org: https://github.com/stanford-cs329a YouTube playlist (9 lectures): https://www.youtube.com/playlist?list=PLangBM27OtEA Community Chinese transcription: https://github.com/flowioo/stanford-cs329a-self-improving-ai-agents
Stanford’s Autumn 2025 graduate seminar CS329A “Self-Improving AI Agents” is taught by Azalia Mirhoseini and Aakanksha Chowdhery, both with Google DeepMind/Brain backgrounds.
All 9 lecture videos went public on YouTube in August 2026 — free to watch. Homework repos are open on GitHub.
What This Course Covers
One sentence: how AI agents use their own execution experience to continuously improve their capabilities.
The course’s central thesis:
“Test-time compute manufactures the training data that improves the next model.”
This isn’t just an optimization trick — it’s a structural claim about how AI capability evolves. The trajectories models generate during inference are becoming the raw material for training better models.
Core Techniques
Self-improvement mechanisms:
| Technique | Role |
|---|---|
| Constitutional AI | Model self-critiques and revises its outputs against principles |
| STaR (Self-Taught Reasoner) | Uses model-generated reasoning chains as training data |
| DAPO | Self-improvement through RL at train time |
| Domain-specific verifiers | Precise reward signals from verifiable external feedback (code execution, math verification) |
Inference-time scaling:
| Technique | Role |
|---|---|
| Test-time compute scaling | More inference compute → better outputs |
| Search-o1-style methods | Combines search with LLMs for active information retrieval during reasoning |
| Multi-step reasoning | Multi-step planning and execution rather than single-shot generation |
Tools and memory:
| Technique | Role |
|---|---|
| ReAct | Interleaves reasoning and action — the foundation for tool-calling agents |
| MemGPT | Long-term memory management beyond the context window |
| Code execution | Model writes and runs code, using results to validate reasoning |
Applications: coding agents, STEM research assistants, robotics.
Paper List
The course publishes a complete reading list including:
- ReAct (Yao et al. 2022) — foundation of tool-calling agents
- Constitutional AI (Anthropic) — self-alignment and self-critique
- STaR (Zelikman et al. 2022) — self-generated reasoning chains as training data
- DAPO — RL-based self-improvement
- MemGPT — external memory management
- Search-o1 — search-augmented reasoning
- AlphaCode — large-scale RL training for code generation
Reading through these papers in course order provides a complete knowledge map of the agent self-improvement landscape.
Open Materials
YouTube playlist (9 lectures, Autumn 2025): All 9 sessions published free in August 2026. Seminar-style format rather than traditional lectures.
https://www.youtube.com/playlist?list=PLangBM27OtEA
GitHub organization: https://github.com/stanford-cs329a
6 repositories including homework assignments from Fall 2025 and the previous quarter — available to clone and work through.
Course website: https://cs329a.stanford.edu/
Complete course calendar with linked papers for each session. No public slides or lecture notes, but the paper links are complete — it’s a structured reading map.
Community Chinese transcription: https://github.com/flowioo/stanford-cs329a-self-improving-ai-agents
Community-compiled transcriptions + Chinese summaries for all 9 lectures.
CS329A vs. CS329Z
Stanford also offers CS329Z “AI Agent Engineering” — easy to confuse with this one:
| Dimension | CS329A (this course) | CS329Z |
|---|---|---|
| Core question | How do agents keep improving themselves? | How do you engineer and deploy agents reliably? |
| Focus | Self-improvement, RL, inference-time scaling | Systems design, reliability, tooling |
| Audience | Research-oriented | Engineering-oriented |
CS329A asks “how good can agents get,” CS329Z asks “how do you actually build them.”
Who It’s For
- Researchers or engineers who want a systematic grounding in agent self-improvement
- Anyone trying to understand why test-time compute scaling matters
- People who need a structured research reading list in this area
9 lectures plus a curated paper list — currently one of the most complete publicly available resources on this topic.
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. You must credit the author and link to the original; removing attribution and republishing as original is not permitted.
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.
- 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。
- 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。
- 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）。
- 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]
💬 评论与讨论
使用 GitHub 账号登录后发表评论