Reflection-Based Memory For Web navigation Agents

TL;DR

This paper introduces Reflection-Augment Planning (ReAP), a memory-enhanced web navigation system that uses self-reflections to learn from past successes and failures, significantly improving performance on web tasks.

Contribution

The paper presents ReAP, a novel reflection-based memory mechanism for web navigation agents, enabling them to leverage past experiences for better task performance.

Findings

Improves baseline results by 11 points overall.

Achieves 29 points improvement on failed tasks.

Reflections transfer effectively across different web tasks.

Abstract

Web navigation agents have made significant progress, yet current systems operate with no memory of past experiences -- leading to repeated mistakes and an inability to learn from previous interactions. We introduce Reflection-Augment Planning (ReAP), a web navigation system to leverage both successful and failed past experiences using self-reflections. Our method improves baseline results by 11 points overall and 29 points on previously failed tasks. These findings demonstrate that reflections can transfer to different web navigation tasks.