VeriGUI: Verifiable Long-Chain GUI Dataset - Paper Overview

The provided text introduces VeriGUI, a new dataset designed to improve the development and evaluation of autonomous Graphical User Interface (GUI) agents. Unlike previous datasets that focus on short, simple tasks, VeriGUI emphasizes long-chain complexity, breaking down tasks into numerous interdependent subtasks that can involve hundreds of steps across various applications. A key innovation is subtask-level verifiability, which allows for detailed assessment of progress at each stage, rather than just the final outcome. The dataset includes human-annotated task trajectories for both web and desktop environments, and initial experiments reveal that current GUI agents still struggle significantly with these complex, multi-step tasks, highlighting the need for more robust planning and decision-making capabilities.