ReAct: Synergizing Reasoning and Acting in Language Models
by Shunyu Yao et al. (ICLR 2023)
The paper that made think-act-observe the default loop for LLM agents.
Overview
ReAct was submitted to arXiv in October 2022 and revised to its ICLR 2023 camera-ready version in March 2023 by Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan and Yuan Cao. It asks what happens when a language model's reasoning traces and its task-specific actions are generated in one interleaved stream rather than separately. The prompting format is few-shot task-solving trajectories built from human-written thoughts, actions, and the environment observations that come back, which is the think/act/observe loop that later became the default control flow in agent frameworks. Reasoning lets the model induce, track and update a plan; actions let it pull facts from an external source instead of hallucinating them. The paper evaluates on four benchmarks. On HotpotQA, ReAct interacting with a simple Wikipedia API reaches 27.4% exact match, and combining it with chain-of-thought self-consistency reaches 35.1%. On FEVER fact verification it reaches 60.9%, and CoT-SC followed by ReAct reaches 64.6%. On ALFWorld it averages a 71% success rate against 37% for the imitation-learning BUTLER baseline, and on WebShop 40.0% against 30.1% for an action-only agent, using only one or two in-context examples. The authors also argue the interleaved trace is far more human-interpretable than reasoning-only or action-only prompting, because you can read what the model believed before each action.
At a Glance
- Topic
- Agentic
- Level
- Advanced
- Format
- Paper
- Cost
- Free
- Duration
- ~60-90 min read for the paper; longer with the appendices and prompts
- Provider
- Shunyu Yao et al. (ICLR 2023)
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓Write few-shot trajectories that interleave thoughts, actions and observations correctly
- ✓See why action grounding reduces hallucination compared with chain-of-thought alone
- ✓Compare ReAct against Act-only, CoT and CoT self-consistency baselines quantitatively
- ✓Understand how ReAct combines with self-consistency for further accuracy gains
- ✓Read the ALFWorld and WebShop setups that became standard agent benchmarks
- ✓Judge when an interpretable reasoning trace matters more than raw accuracy
- ✓Recognise the failure modes the authors report for reasoning-only prompting
Highlights
- •The origin of the think/act/observe loop copied by LangChain, LangGraph and most agent frameworks
- •ICLR 2023 camera-ready, with reported numbers across four distinct benchmark families
- •Gains come from one or two in-context examples, so they are not a fine-tuning artefact
- •Project page at react-lm.github.io links the prompts and code at github.com/ysymyth/ReAct
- •Short enough to read in one sitting and still the clearest statement of the pattern
Who It’s For
Best For
- ✓Engineers who use agent frameworks and want to understand the loop underneath
- ✓Researchers designing new agent prompting or tool-use strategies
- ✓Practitioners debugging agents that hallucinate instead of calling their tools
Prerequisites
- •Familiarity with few-shot prompting and chain-of-thought
- •Comfort reading an ML paper with benchmark tables and ablations
- •Some exposure to LLM tool use or function calling
FAQ
What is ReAct: Synergizing Reasoning and Acting in Language Models?
The 2022 paper that introduced ReAct, the interleaved reasoning-and-acting prompting pattern underneath nearly every modern LLM agent loop. It shows that making a model alternate between free-form thoughts and tool actions beats chain-of-thought alone on question answering and fact verification, and lifts success rates sharply on the ALFWorld and WebShop interactive decision-making benchmarks.
Is ReAct: Synergizing Reasoning and Acting in Language Models free?
ReAct: Synergizing Reasoning and Acting in Language Models is free to access.
What level is ReAct: Synergizing Reasoning and Acting in Language Models for?
ReAct: Synergizing Reasoning and Acting in Language Models is aimed at a advanced audience. Recommended background: Familiarity with few-shot prompting and chain-of-thought, Comfort reading an ML paper with benchmark tables and ablations, Some exposure to LLM tool use or function calling.
How long does ReAct: Synergizing Reasoning and Acting in Language Models take?
Expect roughly ~60-90 min read for the paper; longer with the appendices and prompts. Most learners work through it at their own pace.
What will I learn from ReAct: Synergizing Reasoning and Acting in Language Models?
You'll learn: Write few-shot trajectories that interleave thoughts, actions and observations correctly; See why action grounding reduces hallucination compared with chain-of-thought alone; Compare ReAct against Act-only, CoT and CoT self-consistency baselines quantitatively; Understand how ReAct combines with self-consistency for further accuracy gains; Read the ALFWorld and WebShop setups that became standard agent benchmarks; Judge when an interpretable reasoning trace matters more than raw accuracy; Recognise the failure modes the authors report for reasoning-only prompting.
Topics
Sources
This page was written from 3 sources, 2 on domains other than arxiv.org.