ReAct (Reasoning + Acting)
定义
ReAct 是一种范式，其中模型交替进行推理（接下来做什么、为什么）和行动（工具调用). 来自环境的观察 feeds back into the next 推理 step, forming a loop until the task is done.
它是 the standard pattern for agents that use tools: each action is preceded by a thought, which reduces blind or repetitive tool use. Often combined with chain-of-thought (推理 inside the thought) and with RDD when specs guide 决策s.
工作原理
提示格式为思考 → 行动 → 观察 → 思考 → … → 最终答案。用户给出一个任务；代gent produces a thought (推理 about what to do), then an action (例如 tool call). The environment/tools return an observation, which is appended to the context for the next thought. The loop continues until the agent outputs a final answer. The model decides when to call tools and when to conclude, which reduces arbitrary or repetitive actions. The sequence diagram below summarizes this flow; frameworks like LangChain implement ReAct-style agents with tool registration and message handling.
应用场景
ReAct fits agent workflows where each tool call should be preceded by a clear 推理 step.
- Agents that use tools (search, calculator, API) with explicit 推理
- Reducing arbitrary or repetitive 工具调用 by interleaving thought
- Debuggable agent behavior via visible thought–action–observation traces
外部文档
- ReAct: Synergizing Reasoning and Acting in LLMs (Yao et al.) — Original ReAct paper
- LangChain – ReAct agent — ReAct-style agents in LangChain