ReAct: Synergizing Reasoning and Acting in Language Models

reactreasoningactingllm-agentschain-of-thought

Abstraction: Interleaved reasoning traces and external actions improve LLM agent reliability

Key points:

ReAct interleaves reasoning traces (chain-of-thought) with task-specific actions in a single LLM prompt, allowing each to reinforce the other
On HotpotQA and Fever, ReAct overcomes hallucination and error propagation prevalent in chain-of-thought-only reasoning by querying Wikipedia mid-reasoning
On interactive decision-making benchmarks ALFWorld and WebShop, outperforms imitation learning and RL methods by 34% and 10% absolute success rate respectively
Requires only 1-2 in-context examples; generates human-interpretable, auditable task-solving trajectories
Reasoning traces help the model induce, track, and update action plans and handle exceptions; actions allow grounding via external knowledge bases or environments

Connections: Princeton · Google · AI Agents · Chain Of Thought · Large Language Models · Reasoning