WebHarbor: Docking Real Websites for Evolving GUI Agent Environments
WebHarbor supplies 153 runnable website replicas and 3,518 tasks, including workflows that require signing in.
The Docker-packaged environments reconstruct real services’ interfaces, content and functions under deterministic control. Coding agents expand their capabilities as new tasks arrive, while human contributors check visual fidelity, functionality and task validity. The authors report that a Qwen 4B-level GUI agent trained with reinforcement learning in WebHarbor generalizes to unseen tasks and competes with leading open-source GUI web agents across four benchmarks. Code and data are publicly available.
HF Daily Papers' note
The Docker-packaged environments reconstruct real services’ interfaces, content and functions under deterministic control. Coding agents expand their capabilities as new tasks arrive, while human contributors check visual fidelity, functionality and task validity. The authors report that a Qwen 4B-level GUI agent trained with reinforcement learning in WebHarbor generalizes to unseen tasks and competes with leading open-source GUI web agents across four benchmarks. Code and data are publicly available.
HF Daily Papers' note
score 5