ReAct Deep Dive

Series · 1 posts

← Back to Paper Reading

  • ReAct: Interleave Thought and Action, but Do Not Treat a Few-Shot Loop as an Agent Runtime

    Intermediate Agent runtime, safety, and evaluation
    ReAct Deep Dive: Part 1 , Note: Aug 27, 2026 , Paper: 2023 , NLP

    A source-grounded reading of Yao et al., ICLR 2023: language thoughts join the action space, while HotpotQA, FEVER, ALFWorld, and WebShop keep hallucination, search failure, and the abstract's +34% / +10% in separate buckets.

    Understand it in 90 seconds
    Problem
    LLM reasoning (chain-of-thought) and acting (WebGPT, SayCan) were treated as separate lines of work. CoT never touches an environment; act-only methods can query the outside world but lack high-level plans and exception handling.
    Core insight
    Add language thoughts to the action space. A thought does not change the environment and produces no environment observation; it only updates context, then interleaves with environment actions. The changed decision point is no longer “only think” or “only act,” but “in this same trajectory, is the next legal move a sentence to oneself or a touch of the world?”
    Strongest evidence
    ALFWorld best-of-6 ReAct 71% versus Act 45% and BUTLER best-of-8 37%; WebShop SR 40.0 versus IL+RL 28.7. In a human analysis of HotpotQA failures, 56% of CoT failures are hallucinations versus 0% for ReAct (Table 2).
    Main boundary
    On HotpotQA with PaLM-540B, pure ReAct EM is 27.4, below CoT at 29.4. The 35.1 / 64.6 headline cells are ReAct↔CoT-SC switches. The method is few-shot prompting with a Wikipedia API of search, lookup, and finish. It is not a deployable runtime.
    Read the full deep dive

For speaking invitations, internal engineering sessions, or architecture exchange, see the topics and public work I can bring into the conversation.

Speaking & contact