Teaching Agents in their Own Style: Blackbox Teacher Distillation for Web Agents

TL;DR

This paper introduces WebKD, a novel teacher distillation method for training web agents. WebKD enables a teacher model to edit a student's failures in the student's own style, improving performance. This approach outperforms existing baselines on the TimeWarp benchmark.

Why It Matters

WebKD has practical significance as it can improve the capabilities of open-weight web agents, reducing inference and computational costs. This can lead to more efficient and effective web agents in real-world applications. The method's ability to transfer performance to unseen versions and live web environments is also noteworthy.

Key Findings

• WebKD outperforms behavior cloning and outcome-reward RL baselines
• WebKD achieves 5.5-10.7 pp gain on the TimeWarp benchmark
• WebKD's performance transfers to unseen versions and live web environments