Stanford AI 课程地图 EN GitHub ↗
← 返回课程地图
Stanford · 2026—27 CS 372 LLM 与 Agent ★★★★☆

核心课程 · LLM 与 Agent

AGI 的推理、规划与决策

Artificial General Intelligence for Reasoning, Planning, and Decision Making

AGI 的推理、规划与决策:从 System 1 到 System 2、多 LLM 协作、时序推理。

3 学分 2027 冬 Letter or Credit/No Credit ★★★★☆
AGI推理规划多 Agent

2026-27 学年未在 ExploreCourses 排课;课程官网暂未找到稳定链接。

课程定位

CS 372 由 Edward Chang 讲授,是一门研究导向的课程,讨论如何让 LLM 从“模式匹配”走向“有意识的推理与规划”。内容包含多 LLM 协作、熵引导的信息交换、constitutional AI、时序推理,并以医疗为应用案例。

主要内容

  1. System 1 vs System 2:直觉与审慎推理
  2. 多 LLM 协作:MACI 框架、角色分工、辩论
  3. 信息交换:熵引导、不确定性传播
  4. Constitutional AI:原则约束、自我修正
  5. 时序推理与规划:时间感知、长程任务
  6. 应用:诊断、治疗规划等
  7. 项目:实现一个 AGI 风格的推理系统

先修

机器学习、深度学习基础。

考核与工作量

讲座/讨论 + hands-on project。具体以当季官网为准。

适合谁

对 AGI、推理、规划、多 agent 系统感兴趣的研究生;能接受较思辨内容的人。

学习建议

带着批判视角听:部分概念(如“意识”)争议较大。重点抓可验证的推理/规划方法。

相关课程

官方课程描述(英文原文)

Course Description:Large Language Models (LLMs) have revolutionized AI through remarkable pattern matching capabilities. However, the path to Artificial General Intelligence (AGI) requires advancing beyond unconscious (System 1) to conscious (System 2) processing. This research-oriented course explores fundamental approaches to elevate LLMs toward AGI capabilities through conscious reasoning, planning, and decision-making. Core Research Questions: 1. How can we enable LLMs to transition from pattern matching to conscious deliberation? 2. What frameworks support robust reasoning and verifiable decisions? 3. How do we implement planning and temporal awareness in LLM systems? 4. What role does multi-LLM agent collaboration play in advancing toward AGI capabilities? The course examines: 1. Theoretical foundations of consciousness in AI 2. Multi-LLM Agent Collaborative Intelligence (MACI) frameworks 3. Entropy-guided information exchange 4. Constitutional AI principles 5. Temporal reasoning and planning architectures. Through lectures, discussions, and hands-on projects, students will explore practical implementations across various domains. While healthcare provides immediate applications (diagnosis, treatment planning), the principles apply broadly to any field requiring AGI-level reasoning capabilities. Prerequisites: Machine Learning, Deep Learning