In short. The ReAct paper studies how language models interleave reasoning traces with task-specific actions that return external feedback.

Definition: ReAct: Synergizing Reasoning and Acting in Language Models introduces a method that interleaves language-model reasoning traces, meaning written intermediate reasoning, with task-specific actions. Actions obtain feedback from an external environment, which the model uses to decide how to continue.

Key facts

The ReAct paper, by Shunyu Yao and colleagues, appears as a conference paper at ICLR 2023.
The authors' project page describes prompts containing human-written reasoning and actions alongside observations returned by the environment.
Google Research's experimental overview describes a main prompting setup using PaLM-540B with its model weights frozen.
As of 17 October 2026, the authors' code repository provides GPT-3 prompting experiments and reports using 500 randomly selected validation examples for HotpotQA and FEVER.

How does the reasoning and action loop work?

The authors' prompting explanation describes a trajectory: a sequence of reasoning, actions and observations that solves a task. A prompt supplies a few completed trajectories as examples. The model then generates steps for a new task in that format.

Google Research's model overview distinguishes the roles within this sequence. Reasoning updates the model's context, while an action produces feedback from the external environment. For question answering, reasoning and actions alternate. For decision-making tasks involving many actions, the prompts contain reasoning more sparsely, and the model decides when to insert it. This allows retrieved information to influence the next action.

What tools do the question-answering experiments use?

The paper's experimental setup defines a Wikipedia interface with these command forms:

search[entity]
lookup[string]
finish[answer]

search[entity] returns the first five sentences of the corresponding Wikipedia page, or suggests similar entities when no matching page exists. lookup[string] returns the next sentence on the page containing the requested string. finish[answer] ends the task with an answer.

HotpotQA tests questions requiring reasoning across multiple Wikipedia passages. FEVER tests claims against Wikipedia evidence, using the labels SUPPORTS, REFUTES and NOT ENOUGH INFO. In these experiments, the model receives the question or claim and retrieves supporting material through the interface.

What do the results establish?

Google Research's results table compares ReAct with chain-of-thought prompting, a baseline that generates intermediate reasoning without external actions. With PaLM-540B, ReAct records 27.4% exact match on HotpotQA against 29.4% for that baseline. On FEVER, ReAct records 60.9% accuracy against 56.3%. The comparison therefore depends on the task.

The same report also evaluates ALFWorld, a text-based game, and WebShop, a webpage navigation task. The authors' project examples include an ALFWorld failure and show how editing model-written reasoning can correct behavior. These are experiments in answering questions and completing tasks; the paper does not evaluate website citation rates.

FAQ

Does ReAct require retraining a model?

The main prompting experiments use a frozen model. Google Research's fine-tuning description also explores training smaller models on successful ReAct trajectories, so the paper studies both approaches.

Where can you inspect the experimental implementation?

The authors' repository contains notebooks for HotpotQA, FEVER, ALFWorld and WebShop, alongside a prompts directory and wikienv.py. Its README identifies the released implementation as GPT-3 prompting code and states the validation sample size, details that matter when comparing its results with the paper's tables.

Sources

Check how a page reads to answer engines. Open the GEO score analyzer