Adapting Web Agents with Synthetic Supervision
Zhaoyang Wang, Yiming Liang, Xuchao Zhang, Qianhui Wu, Siwei Han, Anson Bastos, Rujia Wang, Chetan Bansal, Baolin Peng, Jianfeng Gao, Saravan Rajmohan, Huaxiu Yao
2025-11-13
Summary
This paper introduces a new method, SynthAgent, for teaching web agents—computer programs that interact with websites—how to perform tasks on websites they haven't seen before.
What's the problem?
Web agents often struggle when they encounter a new website because it's hard to provide them with enough examples of tasks to learn. Researchers have tried creating fake data to train these agents, but this fake data often contains errors, like tasks that are impossible to complete or instructions that don't quite match what the agent should do, making it difficult for the agent to learn effectively.
What's the solution?
SynthAgent tackles this problem by carefully creating and refining both the tasks and the steps (trajectories) needed to complete them. First, it generates a variety of tasks by systematically exploring the different parts of a website. Then, as it tries to complete these tasks, it checks if the tasks are actually possible and adjusts them if needed. Finally, it cleans up the steps to make sure they are clear and efficient. The agent is then trained on this improved, synthetic data.
Why it matters?
This work is important because it shows that creating high-quality fake data can significantly improve the ability of web agents to adapt to new websites. This means we can build more versatile and helpful AI assistants that can navigate and interact with the web more effectively without needing tons of real-world examples.
Abstract
Web agents struggle to adapt to new websites due to the scarcity of environment specific tasks and demonstrations. Recent works have explored synthetic data generation to address this challenge, however, they suffer from data quality issues where synthesized tasks contain hallucinations that cannot be executed, and collected trajectories are noisy with redundant or misaligned actions. In this paper, we propose SynthAgent, a fully synthetic supervision framework that aims at improving synthetic data quality via dual refinement of both tasks and trajectories. Our approach begins by synthesizing diverse tasks through categorized exploration of web elements, ensuring efficient coverage of the target environment. During trajectory collection, we refine tasks when conflicts with actual observations are detected, mitigating hallucinations while maintaining task consistency. After collection, we conduct trajectory refinement with a global context to mitigate potential noise or misalignments. Finally, we fine-tune open-source web agents on the refined synthetic data to adapt them to the target environment. Experimental results demonstrate that SynthAgent outperforms existing synthetic data methods, validating the importance of high-quality synthetic supervision. The code will be publicly available at https://github.com/aiming-lab/SynthAgent.