TL;DR

• LLMs either reason (CoT) or act (tools), but not both effectively.
• ReAct interleaves reasoning steps and actions in a loop.
• Key takeaway: reasoning + environment interaction improves accuracy and reduces hallucination.

Bibliographic Snapshot

Field	Detail
Citation	Yao et al., ICLR 2023
Keywords	agents, reasoning, tool use
Dataset / Benchmarks	HotpotQA, FEVER, ALFWorld
Code / Repo	https://react-lm.github.io

Problem Statement

Chain-of-thought reasoning is static and hallucination-prone, while action-based systems lack reasoning. There is no unified framework combining reasoning and interaction with external environments.

Core Idea

1. Introduce ReAct loop:
   • Thought → Action → Observation
2. Thought:
   • reasoning trace
3. Action:
   • interact with environment (search, lookup)
4. Observation:
   • feedback from environment
5. Iterate until answer

Visual / Diagram Notes

• Figure 1 (page 2):
  • Comparison:
    • CoT (reason only)
    • Act (action only)
    • ReAct (combined)
• Shows reasoning guiding search, and search grounding reasoning

Key Results

• Improves QA and fact verification over CoT
• Reduces hallucination via external grounding
• Strong gains in interactive tasks (ALFWorld +34%)
• Limitation: requires careful prompt design

Personal Analysis

What worked:

• Very intuitive “agent loop”
• Strong improvement in interpretability
• Flexible across tasks

What puzzled you:

• Prompt-based → brittle
• Scaling to long tasks is unclear

Connections & Related Work

• Precursor to modern LLM agents (LangGraph, AutoGPT)
• Complementary to RAG (retrieval as action)
• Related to SELF-RAG (self-reflection vs action loop)

Implementation Sketch

• Prompt template: Thought: Action: Observation:
• Use external tools:
  • Search API
  • DB queries
• Loop until Finish[]

Open Questions / Next Actions

• How to train instead of prompt?
• Combine with memory (MemGPT)?
• Optimize action selection?

Glossary

• Thought: reasoning step
• Action: tool/environment call
• Observation: returned info