ReAct: Synergizing Reasoning and Acting in Language Models

ReAct explores the synergy between reasoning and acting by interleaving reasoning traces with actions in an environment.

The paper connects chain-of-thought-style reasoning with action generation, allowing an agent to gather information and revise plans.

Reasoning-only systems can lack external information, while acting-only systems can lack explicit planning. ReAct combines both.

The agent alternates between reasoning traces and environment actions, then incorporates observations into subsequent decisions.

The core architecture is Reason → Act → Observe → Reason, with the language model serving as the controller over actions and observations.

The paper evaluates on knowledge-intensive question answering and interactive decision-making environments rather than relying on one static dataset.

ReAct primarily demonstrates prompting and trajectory construction rather than requiring task-specific weight updates to the underlying language model.

Experiments compare reasoning-only, acting-only, and ReAct-style configurations across several tasks.

Key baselines include chain-of-thought reasoning and action-generation approaches without the same interleaved loop.

The paper reports improved performance and better grounding in tasks where external observations and reasoning both matter.

ReAct established a reusable control pattern for LLM agents: reason, act on the environment, observe the result, and continue reasoning.

Yao et al., ReAct: Synergizing Reasoning and Acting in Language Models. https://arxiv.org/abs/2210.03629