ReAct paper describes interleaving reasoning and acting in language models

Shunyu Yao and co-authors from Princeton and Google Research proposed ReAct, a prompting method that had a language model generate reasoning traces and task-specific actions in an interleaved sequence, rather than producing either a chain of reasoning alone or a plan of actions alone. Reasoning steps helped the model decide what action to take next — such as querying a knowledge base or interacting with a simulated environment — and the results of those actions fed back into further reasoning, letting the model update its plan as it learned new information rather than committing to a fixed sequence upfront.

The paper tested this against baselines on question answering (HotpotQA), fact verification (Fever), and two interactive decision-making benchmarks, ALFWorld (simulated household tasks) and WebShop (simulated online shopping). The authors reported that ReAct reduced hallucination and error propagation compared with prompting for reasoning alone, since actions grounded in an external source could correct a reasoning chain that had gone wrong, and that it produced trajectories that were more interpretable to a human reader than baseline approaches — a person could follow why the model took each step.

ReAct extended the chain-of-thought line of work published earlier that year from a technique for improving answers to a technique for taking actions, and it became one of the most cited templates for what a language-model “agent” was: a loop of think, act, observe, repeat. Product implementations of tool-using agents that followed over the next several years — from early open-source projects to labs’ own agent frameworks — drew directly on this reasoning-and-acting loop, even where they departed from ReAct’s specific prompting format.