ReAct: Synergizing Reasoning and Acting in Language Models
Notes on ReAct: Synergizing Reasoning and Acting in Language Models by Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan and Yuan Cao.
ReAct combines two capabilities that are often studied separately: reasoning in language and acting in an external environment.
The model produces a reasoning trace, takes an action such as querying a knowledge base or manipulating an environment, receives an observation, and then reasons again. This gives the model a way to update its plan, handle unexpected results and gather information that was not in its original context.
The authors evaluate the approach on question answering, fact verification and interactive decision-making tasks. On HotpotQA and Fever, interacting with a Wikipedia API helps ground the reasoning and reduce hallucination. On ALFWorld and WebShop, ReAct improves success over the comparison methods while using only one or two in-context examples.
The important idea is the loop itself: reason, act, observe, repeat. The ReAct describes this loop in the context of an agent harness.