What did the Yao et al. 2022 ReAct paper establish about combining reasoning and acting?

**ReAct** is a 2022 paper whose name is short for reasoning and acting. It asked a simple question. If a language model can think step by step, and it can also call tools, what happens when you let it do both, turn by turn? The paper compared three setups. Reasoning only, where the model thinks out loud but never checks the world. Acting only, where the model calls tools but never explains its plan. And the combination, where the model alternates a thought with an action and then reads the result. The combination won. On tasks that need outside facts, like answering trivia by searching Wikipedia, reasoning alone made confident but wrong guesses. Acting alone wandered without a plan. Doing both kept the model grounded in real results while still planning ahead. The lasting gift was a loop shape: Thought, then Action, then Observation, repeated. Nearly every agent framework today is built on that exact rhythm, which is why ReAct is treated as a landmark.

**ReAct**, from Yao et al. in 2022, stands for reasoning and acting. Its core finding is that interleaving reasoning traces with tool actions beats either one alone. The paper ran a clean three-way comparison. Pure chain-of-thought reasons but never touches the world, so it hallucinates facts on knowledge tasks. Pure acting calls tools but has no plan, so it flails on multi-step problems. ReAct alternates a thought with an action, then feeds the observation back. That structure helps two ways. Reasoning steers which action to take next. The observation grounds the next thought in a real result. So errors get corrected instead of compounding. The loop it introduced has a fixed shape: $$\text{Thought} \rightarrow \text{Action} \rightarrow \text{Observation} \rightarrow \text{Thought} \rightarrow \cdots$$ ReAct became the template because that rhythm maps onto a tool using agent. The thought is the model's reasoning, the action is a tool call, and the observation is the tool result appended to context. Modern frameworks reuse this cycle, swapping free text actions for typed function calls.