📄 中文摘要
大型视觉语言模型赋予了 GUI 代理强大的界面理解和交互能力。然而，由于在训练过程中对特定领域软件操作数据的曝光不足，这些代理表现出显著的领域偏见，缺乏对特定应用程序操作工作流程和 UI 元素布局的熟悉度，从而限制了其在实际任务中的表现。研究提出了 GUIDE（通过教学视频驱动的专业知识去偏见），这是一个无需训练的即插即用框架，通过自动从网络教程视频中获取特定领域的专业知识，解决了 GUI 代理的领域偏见问题。
指南：通过实时网络视频检索和即插即用注释解决 GUI 代理中的领域偏见
📄 English Summary
GUIDE: Resolving Domain Bias in GUI Agents through Real-Time Web Video Retrieval and Plug-and-Play Annotation
Large vision-language models have equipped GUI agents with robust capabilities for interface understanding and interaction. However, these agents exhibit significant domain bias due to insufficient exposure to domain-specific software operation data during training, which limits their familiarity with the specific operation workflows and UI element layouts of particular applications, thereby constraining their real-world task performance. The study presents GUIDE (GUI Unbiasing via Instructional-Video Driven Expertise), a training-free, plug-and-play framework that autonomously acquires domain-specific expertise from web tutorial videos, effectively resolving the domain bias in GUI agents.