ReAct: Synergizing Reasoning and Acting in Language Models (Google Research Blog)

The Google Research Blog announcement (8 Nov 2022) of the ReAct paper — the accessible, figure-led popularisation of the same work the ICLR 2023 paper reports rigorously. Posted by Shunyu Yao (Student Researcher) and Yuan Cao (Research Scientist), Google Research, Brain Team. It is the wiki’s secondary, non-peer-reviewed register of the ReAct paradigm; the paper is the primary.

TL;DR

• Frames the problem as two pre-existing but separate lines of work: chain-of-thought prompting (reasoning, but ungrounded in the external world — can’t reactively explore or update knowledge) and LLMs for planning/acting in interactive environments (acting, but no abstract reasoning over high-level goals or working memory). ReAct’s pitch: combine them.
• ReAct = Reason + Act. The model generates interleaved verbal reasoning traces and text actions; actions get observation feedback from an external “Env”; reasoning traces don’t change the environment but update the model’s internal context. Lists the useful thought types: decompose goals into plans, inject commonsense, extract from observations, track progress, handle exceptions.
• Prompting setup: frozen PaLM-540B, few-shot in-context. Fine-tuning setup: use prompted PaLM-540B to generate trajectories, filter to successes, fine-tune smaller PaLM-8B/62B (reduces human-annotation need).
• Results (the blog reproduces the headline tables): on HotpotQA/FEVER ReAct beats act-only and is competitive with CoT, with ReAct+CoT best; on ALFWorld ReAct 71 vs Act 45 vs IL 37 (2-shot), on WebShop ReAct 40 vs Act 30.1 vs IL 29.1 (1-shot) — +34% / +10% over baselines trained on ~10⁵ instances.
• Human-in-the-loop correction: a human inspector edits ReAct’s reasoning traces; replacing a hallucinating sentence with inspector hints realigns the agent’s behaviour — “new forms of human–machine collaboration.”
• Concludes ReAct is “simple yet effective,” giving interpretable decision traces, and flags follow-on work: multitask training and coupling ReAct with reward models.

What was actually ingested

The full blog post (9-page PDF capture of the Google Research Blog page), including the four results tables and the figure captions. Content is a faithful subset of the paper; no claims here that the paper does not also make.

Linked entities and concepts

• Promoted to entities this ingest (second-source rule — also author the paper): Shunyu Yao, Yuan Cao.
• Concepts: react-reasoning-acting, ai-agents, agent-harness, foundation-models.
• Publisher: Google Research (Brain team).

Relationships

• supports 2022-10-06-yao-et-al-react-synergizing-reasoning-acting — same work, popular register. Read the paper for the rigorous treatment; this page is the legible summary.