Type: url
Original: https://github.com/ysymyth/ReAct
Status: generated
Summary
The authors’ GPT-3 prompting code for the ICLR 2023 ReAct paper, released under the MIT license with one notebook per task for HotpotQA, FEVER, ALFWorld and WebShop and a small Wikipedia environment. Its README reports a four-task comparison of PaLM-540B against GPT-3 (davinci-002) in which the two models come out ahead on two tasks each, summarised as the models being better at different tasks — two of its figures disagree with the paper’s own. It also points readers who want to apply ReAct beyond these four tasks at LangChain’s zero-shot ReAct agent rather than at its own prompts.