ReAct Prompting
ReAct is a prompting framework, introduced by Yao et al. in 2022, that interleaves explicit reasoning traces with discrete tool actions so that a language model both thinks aloud and acts step by step against an external environment.
ReAct prompting is a technique for LLM agents, introduced by Shunyu Yao and colleagues in the 2022 paper "ReAct: Synergizing Reasoning and Acting in Language Models" (arXiv:2210.03629). The model is prompted to alternate between Thought, Action, and Observation lines: it writes a short reasoning step, emits a structured action such as a search query or API call, then incorporates the returned observation before deciding what to do next. The original paper evaluated ReAct on knowledge-intensive question answering (HotpotQA), fact verification (FEVER) using a Wikipedia API, and interactive decision-making benchmarks ALFWorld and WebShop. On the interactive benchmarks ReAct outperformed imitation and reinforcement learning baselines by sizable absolute margins while using only a handful of in-context demonstrations. On the QA tasks the framework reduced the hallucination and error-propagation problems typical of pure chain-of-thought reasoning by letting the model check itself against retrieved evidence. ReAct became one of the foundational patterns for agent frameworks built on top of Function Calling (LLM) interfaces, because its explicit Thought-Action-Observation loop maps cleanly onto a tool-calling runtime: each Action is dispatched to a real tool, and each Observation is the structured result fed back into context. It is frequently cited as a mitigation for capability hallucination, since grounding every claimed action in a concrete tool response makes phantom actions visible.