ReAct: the framework that unites reasoning and acting in LLMs
In one sentence Yao et al. introduce ReAct, a schema alternating explicit thoughts (Thought) and concrete actions (Act) in LLMs, a schema that will become a reference for LLM-based agents.
Before ReAct, language models could reason or act, but rarely both in a structured way at once. The paper proposes a simple format: the model writes a thought first, then executes an action (like searching Wikipedia), then observes the result, and repeats.
This Thought-Action-Observation loop lets the model correct its reasoning on the fly by consulting external sources such as Wikipedia. On interactive tasks (a simulated text environment and a simulated online shop) it clearly beats earlier methods; on multi-hop questions the best result comes from combining ReAct with classic chain-of-thought.
Companies
Google Research, Princeton University
Tools
ReAct
Tags
Sources