ReAct: Synergizing Reasoning and Acting in Language Models

Chain-of-thought prompting let models reason step by step, but they reasoned from memory, so a wrong fact early in the chain carried through to a confident wrong answer. Separately, researchers had models generate actions, such as moves in a game or searches, but without any reasoning the models lost track of what they were trying to do.

ReAct puts the two together in one loop. The model writes a Thought ("I need to find who the CFO is first"), then an Action (search[Acme Corp]), then receives an Observation from the environment (the search result). The next thought can use that observation, revise the plan or handle a surprise. This repeats until the model calls finish[answer]. The prompt contains only one to six hand-written example trajectories in this format.

The authors tested it on two kinds of tasks. On knowledge questions (HotpotQA) and fact checking (FEVER), with a deliberately simple Wikipedia tool, ReAct beat acting without reasoning and almost eliminated hallucinated facts, though on HotpotQA plain chain of thought still scored slightly higher, and the best results came from combining the two. On interactive tasks, a text-based household game (ALFWorld) and an online shopping site (WebShop), ReAct beat imitation and reinforcement learning methods trained on thousands of examples, by 34 and 10 points of success rate.

Every agent framework since, every "tool calling" API and every coding agent runs a descendant of this loop.

The ReAct loop is the default architecture for agents, so this paper is also the first failure analysis of the agents you will deploy. It is unusually honest about what goes wrong.

Grounding trades one failure for others. The authors hand-labeled why runs failed. Chain-of-thought failures were mostly hallucinated facts (56%). ReAct had none of those. Its failures were reasoning errors, often the model repeating the same thought and action in a loop (47%), and searches that returned nothing useful, after which the model struggled to recover (23%). Expect exactly this in production: agents that loop, and agents derailed by a bad tool result.

The tools decide the ceiling. The Wikipedia tool in the paper was weak on purpose, able only to fetch a page by exact name or find the next sentence containing a string. Nearly a quarter of ReAct's failures came from it. On a deployment, the quality of search, the shape of tool outputs and the error messages a tool returns matter as much as the model.

Design the exits. The paper's best configurations add a fallback: if ReAct has not finished in 7 steps (5 for FEVER), switch to plain reasoning with voting; or if the votes disagree, switch to ReAct. Step budgets, loop detection and fallback paths are not polish. They are part of the method.