ysymyth/ReAct

Implements an interleaved prompting approach where language models alternate between generating reasoning traces and executable actions, evaluated across knowledge-intensive QA (HotpotQA, FEVER), interactive simulation (AlfWorld), and e-commerce tasks (WebShop). Provides OpenAI API-based implementations with Jupyter notebooks for each benchmark, while LangChain offers a generalized zero-shot agent variant for custom task integration. Demonstrates that the reasoning-acting loop outperforms chain-of-thought baselines, with performance varying between PaLM and GPT-3 depending on task characteristics.

3,629 stars. No commits in the last 6 months.