ReAct formalized an interleaved reasoning-and-action loop
In *ReAct: Synergizing Reasoning and Acting in Language Models*, Shunyu Yao and co-authors study language models that interleave reasoning traces with task-specific actions. Their framing treats reasoning as a way to induce, track, and update plans, while actions let the model interact with external sources or environments and gather information.
The paper evaluates the approach on question answering, fact verification, and interactive decision-making tasks. It reports that ReAct overcame issues of hallucination and error propagation on HotpotQA and FEVER by interacting with a simple Wikipedia API. On ALFWorld and WebShop, the authors report absolute success-rate improvements of 34 and 10 percentage points, respectively, over the imitation and reinforcement-learning methods they compare against. These are results in the paper’s specified benchmark settings, not a general guarantee about deployed agents.
The paper’s contribution is a research framing and evaluation of an interleaved loop, not evidence that a loop alone supplies robust tool permissions, verification, or production reliability.