AIAny
Icon for item

ReAct: Synergizing Reasoning and Acting in Language Models

Interleaves chain-of-thought reasoning with tool-using actions in one LLM loop: the model plans, queries a source like Wikipedia, then revises from results. Cuts hallucination versus reasoning-only prompting and beats trained agents on interactive tasks.

Introduction

Chain-of-thought lets a model reason but leaves it trapped in its own head, free to hallucinate facts it never checked. ReAct's move is almost embarrassingly simple: let the model emit a "thought," then an "action," then read the resulting "observation," and repeat. Reasoning decides what to look up; the lookup grounds the next thought. That single feedback loop is the seed under most of today's agent frameworks.

Key Findings
  • On HotpotQA and Fever, interleaving Wikipedia search actions with reasoning sharply reduces the fact-fabrication that plagues pure chain-of-thought, because the model is checking claims instead of inventing them.
  • On ALFWorld and WebShop interactive tasks, ReAct prompting with one or two in-context examples outperforms imitation-learning and reinforcement-learning agents trained on thousands of episodes, by absolute success-rate margins in the tens of points.
  • The reasoning trace is human-readable, so you can watch where the agent went wrong and intervene, rather than debugging an opaque policy.
Who Should Read It

Great fit if you build LLM agents or tool-using systems and want the conceptual blueprint that ReAct-style scaffolding, LangChain, and function-calling loops all descend from. Look elsewhere if you need a turnkey library or production code; this is a 2022 research paper, and the prompting tricks have since been refined and partly absorbed into native tool-calling APIs. Read it for the idea, not the implementation.

Information

  • Websitearxiv.org
  • OrganizationsGoogle Research, Princeton University
  • AuthorsShunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, Yuan Cao
  • Published date2022/10/06

More Items

Enables RL post-training with million-token prompts under a fixed GPU budget by evaluating shared prompt state without autograd, retaining only minimal model state, and replaying short response branches; instantiated as GRPO and demonstrated on Qwen3.6-27B and GLM-5.2 up to multi-million token execution.

Analyzes adversarial weaknesses of World-Action Models (WAMs) via BadWAM, a framework that crafts visual perturbations to decouple a model’s imagined future from its executed actions. Introduces two attack modes—action-only (disruptive) and imagination-preserving (stealthy)—and shows large drops in closed-loop task success (e.g., 96.5%→43.1%).

Turns fragile, implicit search progress into explicit, persistent, shared state for multi-agent information seeking — externalizes progress as Frontier Task, Evidence Graph, Coverage Map and Failure Memory, and uses pipeline-parallel scheduling plus a middleware harness to avoid repeated failed searches and improve utilization and throughput.