The ReAct framework was proposed by a Princeton University team in 2022; its core idea is to have large language models alternately perform thinking and action, incorporating environmental feedback into the next round of reasoning to form a closed loop.