Paper breakdown: ReAct — reasoning and acting in language models

Sample editorial breakdown of Yao et al. 2022, a widely cited pattern for interleaving free-text thoughts with tool actions.

Byline:: AI Agents Info Research
Published:: October 18, 2026
Last updated:: October 26, 2026
Paper:: ReAct: Synergizing Reasoning and Acting in Language Models — Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, Yuan Cao (2022)

Paper Code

Abstract / summary

Sample editorial summary: ReAct prompts a language model to produce traces that mix verbal reasoning ('thoughts') with actions in an environment, then observations, rather than only chain-of-thought or only action. The paper studies question answering and interactive tasks and compares against acting-only and reasoning-only baselines.

Key findings

The paper argues that interleaving thoughts and actions can reduce error propagation compared with acting-only prompts on the tasks they study.
It presents qualitative traces where thoughts are used to track goals and handle exceptions — useful as a design pattern, not as a universal law.
Results are task- and model-specific; treat numbers as historical, not as a 2026 leaderboard.

Methodology

Sample note: the authors evaluate prompting strategies on public QA and interactive environments of that period, using then-current language models. Read the paper's experimental section for splits, baselines, and decoding details rather than relying on this summary.

Limitations

This page is a sample editorial summary, not a peer review or a reproduction.
Prompting recipes drift as models and tool APIs change; do not copy 2022 numbers into a 2026 comparison.
Thought traces are not a window into a mind; they are generated text that may or may not be faithful to the model's internals.

Sample research page. Editorial summary for the /research/[slug] template. Replace this file with a reported breakdown. All bibliographic fields above point at the public paper.

Why this paper is on the syllabus

If you have seen an agent print “Thought:” then “Action:” then “Observation:”, you have seen a descendant of ReAct. The contribution is not a new model. It is a prompting pattern that treats language as both a scratchpad and an interface to tools.

That is why it belongs next to our harness guide and the harness glossary entry.

What to steal (carefully)

Interleave, do not dump a long plan then execute blindly.
Keep observations in the trace so later steps can correct earlier ones.
Measure against acting-only and reasoning-only ablations if you claim the mix matters.

What not to steal

Do not treat the original task scores as a modern benchmark. Tool APIs, context windows, and models have moved. If you need current numbers, use a living eval and cite the date — see reading agent benchmarks.