跳到主要内容Skip to main content

ReActReAct

ReAct

  • 智能体Agents
  • 第三站 · 用深Station 3 · Going Deeper
  • 引用核验 2026-11-19refs checked 2026-11-19
解决什么问题What problem it solves

让模型交替「想一步、做一步、看结果」,把推理和工具调用编成循环。Make the model alternate between thinking, acting and observing, folding tool use into a reasoning loop.

这是什么What it is

模型先输出一段推理(Thought),再选择一个动作(Action),例如搜索或调用 API。

执行动作后观察结果(Observation),把结果写回上下文,再决定下一步。

循环直到给出最终答案;推理不是一次性写完,而是被工具结果不断修正。

The model writes a Thought, then chooses an Action such as search or an API call.

The result comes back as an Observation and is appended to context before the next step.

The loop ends with a final answer; reasoning is revised by tool results, not written once.

为什么需要知道Why it matters

它是大多数 agent 框架的默认骨架。理解 ReAct,才能看懂工具调用、轨迹评测和长时任务为什么容易跑偏。It is the default skeleton of most agent frameworks. Understand ReAct and you can see why tool calls, trajectory evals and long tasks drift.

出处Sources
  • ReAct 论文(arXiv)ReAct: Synergizing Reasoning and Acting in Language Models (arXiv 2210.03629) ↗