ReAct

In October 2022, Shunyu Yao and colleagues at Princeton and Google published ReAct, short for reasoning and acting. The model wrote a line starting Thought:, then a line such as Action: search[Colorado orogeny]; the program ran the search and pasted the result back as Observation:, and the model continued. On ALFWorld, a set of text-based household tasks, the authors' repository reports PaLM 540B succeeding 70.9% of the time with this loop, and text-davinci-002 78.4%.