ReAct Paper Explained: Why Reasoning and Acting Should Dance Together in LLMs

Forum topic · 小凯 · 2026-05-22

Summary

This article summarizes ReAct (Synergizing Reasoning and Acting in Language Models, ICLR 2023), a foundational paradigm that interleaves chain-of-thought reasoning with external actions inside a single LLM trajectory. Instead of generating a complete thought chain and then acting, ReAct prompts the model to alternate Thought → Action → Observation loops, enabling few-shot learning without weight updates. Across QA and decision-making benchmarks such as ALFWorld, ReAct outperforms pure CoT by reducing hallucination and injecting up-to-date knowledge from tools like search engines. The article highlights three core contributions: (1) reasoning and acting should be interleaved rather than separated, (2) LLMs become open systems that actively query external environments, and (3) feedback from actions improves subsequent reasoning. Key limitations include a predefined action space, unbounded loops, and cascading errors from faulty thoughts. ReAct inspired Toolformer, AutoGPT, LangChain agents, and modern Deep Research systems, making it the essential starting point for Agentic LLM research.

Key Points

The Core Problem

Traditional Chain-of-Thought (CoT) prompting generates a full reasoning chain before acting. If the model discovers missing information mid-thought, it cannot backtrack. ReAct (Synergizing Reasoning and Acting in Language Models, ICLR 2023) proposes interleaving reasoning and acting in a continuous loop.

How ReAct Works

The model alternates between three modes:

1. Thought: decompose the problem ("I need to look up the capital of France") 2. Action: call an external tool (search engine query "capital of France") 3. Observation: receive the result ("Paris") and feed it back into the next thought

Training is done via few-shot prompting: ReAct trajectories in context teach the model the Thought–Action–Observation pattern. No fine-tuning is required.

Why Interleaving Beats Separation

Knowledge grounding: ReAct reduces the CoT hallucination problem by letting the model fetch real-world facts on demand rather than fabricating them.
Error visibility: thoughts are explicit, so failures can be inspected step-by-step.
Error propagation: a faulty thought produces a faulty action, creating a cascading failure chain. The paper mitigates this by forcing the model to always emit a thought, even when it seems unnecessary, adding overhead but improving reliability.

Empirical Findings

QA tasks: ReAct significantly outperforms pure CoT when external knowledge is required.
Decision-making (e.g., ALFWorld): ReAct excels in multi-step interactive environments.
Interpretability: explicit thought traces make agent decisions traceable.

The Paradigm Shift

closed system

open system

Toolformer: teaches LLMs to invoke APIs
AutoGPT: autonomous multi-step task execution
LangChain: standardizes LLM + tool interactions
Deep Research systems: modern multi-step research agents

Critical Limitations

Predefined action space: original ReAct supports only a fixed set of actions (Search, Lookup, Finish). Unforeseen actions are impossible without creative tool invention.
Unbounded loops: without intervention, the model can cycle indefinitely between thinking and acting; max-step cutoffs are crude.
Action failures: error handling is limited compared to human problem-solving flexibility.

Open Questions for Future Work

These challenges were inherited and expanded by Toolformer, AutoGPT, and the broader Agentic LLM research agenda.

Reference

Shunyu Yao et al., *ReAct: Synergizing Reasoning and Acting in Language Models*, ICLR 2023

https://openreview.net/forum?id=WE_vluYUL-X