Back to glossary

ReAct Prompting

A prompting technique that combines 'Reasoning' and 'Acting,' allowing an LLM to generate reasoning traces and task-specific actions in an interleaved manner.

AbbreviationReAct
Categoryai

Overview

ReAct (Reason + Act) allows an LLM to interact with the external world (e.g., searching Wikipedia) while maintaining a logical internal monologue. This helps the model stay grounded and reduces hallucinations.

The Cycle

  • Thought: The model reasons about the current state ('I need to find the height of the Eiffel Tower').
  • Action: The model calls a tool ('Search[Eiffel Tower height]').
  • Observation: The model receives the result from the tool ('330 meters').
  • Repeat: The model continues until the goal is reached.

Significance

ReAct is the foundational logic behind most modern AI agents and tool-use implementations.

Related Terms

Chain of ThoughtAgentic Workflow