We show that domain-general automatic evaluators can significantly improve the performance of agents for web navigation and device control. We experiment with multiple evaluation models that trade off between inference cost, modularity of design, and accuracy. We validate the performance of these models in several popular benchmarks for digital agents, finding between 74.4 and 92.9% agreement with oracle evaluation metrics. Finally, we use these evaluators to improve the performance of existing agents via fine-tuning and inference-time guidance. Without any additional supervision, we improve state-of-the-art performance by 29% on the popular benchmark WebArena, and achieve around 75% relative improvement in device control settings.
@article{arxiv.2404.06474,
title = {Autonomous Evaluation and Refinement of Digital Agents},
author = {Jiayi Pan and Yichi Zhang and Nicholas Tomlin and Yifei Zhou and Sergey Levine and Alane Suhr},
journal= {arXiv preprint arXiv:2404.06474},
year = {2024}
}
Comments
Published at COLM 2024. Code at https://github.com/Berkeley-NLP/Agent-Eval-Refine