You know how some AI can reason through problems (like chain-of-thought prompting), and others can take actions (like generating plans)? The ReAct paper shows what happens when you combine both.

The Core Idea

ReAct stands for Reasoning + Acting. Instead of just thinking OR doing, the AI alternates between the two:

  1. Reason: Think about what to do next
  2. Act: Take an action (like querying a knowledge base)
  3. Observe: See what happened
  4. Repeat: Reason about the new information and act again

Why It Matters

This simple pattern solves a huge problem: hallucination. When AI just reasons without checking facts, it makes stuff up. When it can act (like searching Wikipedia), it stays grounded in reality.

The results are impressive:

  • On HotpotQA (question answering): ReAct beats pure reasoning approaches
  • On ALFWorld and WebShop (interactive tasks): It outperforms imitation learning by 34% and 10%
  • All this with just 1-2 examples in the prompt

The Pattern in Practice

Here's what a ReAct trace looks like:

Task: What's the elevation of the mountain where the 2018 Winter Olympics were held?

Thought: I need to find where the 2018 Winter Olympics were held. Action: Search "2018 Winter Olympics location" Observation: The 2018 Winter Olympics were held in Pyeongchang, South Korea.

Thought: Now I need to find mountains near Pyeongchang. Action: Search "Pyeongchang mountains" Observation: The games were held near Mount Taebaek.

Thought: Now I need the elevation of Mount Taebaek. Action: Search "Mount Taebaek elevation" Observation: Mount Taebaek has an elevation of 1,567 meters.

Answer: 1,567 meters

Why This Works

By interleaving reasoning and acting:

  • The AI can check its assumptions
  • It handles exceptions better (what if the first search doesn't work?)
  • It's way more interpretable (you can see the AI's thinking)
  • It stays factual by grounding in external sources

The Takeaway

ReAct shows that the best AI systems aren't just thinkers or just doers - they're both. And that's exactly how modern AI agents work today (like the ones in LangChain).

Read the full paper: ReAct: Synergizing Reasoning and Acting in Language Models (arXiv:2210.03629)

Authors: Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, Yuan Cao

Project site: https://react-lm.github.io/