SynthAgent - Adapting Web Agents with Synthetic Supervision

See more details in paper and code links below!

📄 Paper: https://arxiv.org/abs/2511.06101

💻 Code: https://github.com/aiming-lab/SynthAgent

Overview

SynthAgent is a framework for adapting web agents to any new environments using synthetic supervision. It efficiently synthesizes diverse user tasks by categorized exploration. Then, it refines these tasks during trajectory collection to mitigate hallucinations. After collection, it refines the trajectories to enhance the data quality. Finally, it utilizes refined data to fine-tune the agent, improving the performance in the target new environment.

Related Resources

Resource	Link
📦 Data	🤗 ChilleD/SynthAgent
🤖 SynthAgent-SFT-Qwen2.5-VL-7B	🤗 ChilleD/SynthAgent-SFT-Qwen2.5-VL-7B
🤖 SynthAgent-SFT-UI-TARS-1.5-7B	🤗 ChilleD/SynthAgent-SFT-UI-TARS-1.5-7B

Citation

If you find our paper or codes useful, please kindly cite:

@article{wang2025adaptingwebagentssynthetic,
      title={Adapting Web Agents with Synthetic Supervision}, 
      author={Zhaoyang Wang and Yiming Liang and Xuchao Zhang and Qianhui Wu and Siwei Han and Anson Bastos and Rujia Wang and Chetan Bansal and Baolin Peng and Jianfeng Gao and Saravan Rajmohan and Huaxiu Yao},
      year={2025},
      eprint={2511.06101},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2511.06101}, 
}