← All learnings
📄 Paper

ReAct: Synergizing Reasoning and Acting in Language Models

Yao et al. · December 2024 Completed

The paper that formalized what most practitioners had already intuited. Worth reading to understand the intellectual lineage of modern agentic frameworks.

The Core Insight

Interleaving reasoning traces (think) with actions (act) dramatically reduces hallucination in tool-using agents. The model is forced to make its reasoning visible before taking an action, which creates a natural audit trail and catches errors before they compound.

The Thought → Action → Observation loop is simple to understand but powerful in practice. Every major agentic framework today (LangChain, LangGraph, AutoGPT, etc.) is, at some level, an implementation of this pattern.

Why It Works Better Than Chain-of-Thought Alone

Chain-of-thought prompting encourages the model to reason step by step, but all the reasoning happens in the model’s head (the context). When the model makes a mistake in step 3, it compounds through steps 4 and 5 with no correction mechanism.

ReAct adds external grounding. After each action, the environment provides an observation that anchors the next reasoning step in reality. The model can’t hallucinate what a web search returned — the actual search result is right there in the context.

Limitations Worth Understanding

The paper’s results are on relatively simple benchmarks (HotpotQA, Fever, AlfWorld). Real-world performance depends heavily on:

The “act” step in real systems is much harder than in the paper’s controlled environments. This is why clean tool interfaces matter so much.

Connection to Modern Systems

ReAct is the conceptual foundation of LangChain’s AgentExecutor, LangGraph’s react_agent, and Anthropic’s tool use implementation. Understanding the paper helps you reason about why these systems behave the way they do — and how to fix them when they don’t.

LLMsReasoningTool UseAgents