ReAct (Reason + Act)

Mechanism

Under a ReAct prompt, each turn of the model's output follows a fixed structure: first a stretch of reasoning (what I know, what to check next), then an action (call a tool / search a term). The runtime executes it and feeds back an observation, and the model writes the next reasoning step. This repeats until the model produces a final answer.

Difference from chain-of-thought

Chain-of-thought only has the model spell out reasoning, with no external interaction — good for pure logic and maths. ReAct inserts real external actions between reasoning steps, letting the model fetch information it does not have and correct course against facts. The paper's core result: combining the two yields fewer hallucinations and more traceability than either alone.

Common misconceptions

• ReAct is not a framework or library — it is a prompting/interaction structure implementable in a few dozen lines.
• "Thought" text is an intermediate artefact for the process, not the model's "true internal reasoning."

Related terms

Sources

1. Yao et al. — ReAct: Synergizing Reasoning and Acting in Language Models, ICLR 2023 (arXiv:2210.03629)
2. Project page — react-lm.github.io