ReAct

Why we included this project

ReAct, the ICLR 2023 paper by Shunyu Yao and colleagues, showed that a language model reasons better when it writes out its thinking and then acts on it, querying Wikipedia or stepping through an environment before settling on an answer. The repo pairs that prompting pattern with reproducible notebooks for question answering and interactive tasks like AlfWorld and WebShop, so you can see the original results instead of taking the idea on faith. Most people meet ReAct today through LangChain's zero-shot agent, which is fine, but the value of this repository is reading the source experiments and understanding the pattern itself. It is a short, readable reference for anyone building agents that need to gather information or call tools before they reply, and a useful grounding for designs built on the current generation of tool-using models.