“ReAct: Synergizing Reasoning and Acting in Language Models” was posted to arXiv on October 6, 2022 by Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao, with authors from Princeton and Google. It proposed a simple but influential pattern: instead of asking a language model to either reason about a problem or take actions in an environment, let it do both in an interleaved loop. The model writes a short reasoning trace (“thought”), takes an action such as a search query, observes the result, and then reasons again.
The name is a contraction of “Reason” and “Act.” Reasoning helps the model plan and revise what it is doing; acting lets it pull in outside information rather than relying on what is baked into its weights. On the HotpotQA and FEVER question-answering tasks, querying a Wikipedia API inside the loop reduced the hallucination and error-propagation that pure chain-of-thought prompting suffered. On the interactive ALFWorld and WebShop benchmarks, ReAct outperformed imitation and reinforcement-learning baselines by 34 percent and 10 percent respectively, using only a couple of in-context examples.
ReAct mattered because it gave a concrete, promptable shape to the idea of an agent. The thought-action-observation loop it described is essentially the control loop that frameworks like AutoGPT, LangChain agents, and most tool-using assistants adopted, and it remains the default mental model for how a language model drives external tools.