AutoRefine

Gold definitionUpdated Apr 2, 2026

Definition

AutoRefine is a framework designed for large language model (LLM) agents to extract and maintain 'dual-form Experience Patterns' from execution histories. It captures both procedural logic via specialized subagents and static knowledge as skill patterns, preventing repository degradation through continuous maintenance.

At a glance

Executive summary

AutoRefine is a new framework that helps AI agents learn and remember complex procedures and facts from their experiences, rather than forgetting them. It does this by creating specialized 'experience patterns' and constantly refining them, leading to much better performance on challenging tasks.

TL;DR

AutoRefine helps AI agents learn and remember how to do complex tasks by extracting and maintaining smart 'experience patterns' from their past actions.

Key points

Extracts and maintains 'dual-form Experience Patterns' (subagents for procedural, skill patterns for static knowledge) from agent execution histories.
Solves the problem of LLM agents failing to accumulate knowledge and capture procedural logic, preventing repository degradation.
Used by researchers and engineers developing advanced LLM agents for complex, multi-step reasoning and planning tasks.
Outperforms methods relying on flattened textual knowledge and even manually designed systems for capturing procedural coordination.
Represents a significant trend towards more autonomous, continually learning, and robust LLM agents capable of long-term knowledge management.

Use cases

Developing LLM agents for complex planning tasks like travel itinerary generation, where procedural coordination is key.
Enhancing AI agents in interactive text-based environments such as ALFWorld, requiring sequential decision-making and memory.
Improving LLM agents for scientific discovery tasks in environments like ScienceWorld, demanding both factual recall and experimental procedures.
Creating more capable AI assistants that can learn and adapt over long periods, accumulating skills and knowledge from user interactions.
Building autonomous software engineering agents that can learn to perform complex coding and debugging procedures.

Also known as

Experience Pattern Framework