arXivZixiang Chen
JarvisGUI: Towards Cross-Device GUI Agents with Dynamic Task Composition
Papers81
Most GUI benchmarks evaluate agents inside isolated, static environments, overlooking how workflows actually cross devices. JarvisGUI establishes a unified evaluation framework across Android, Windows, and Ubuntu, formalizing tasks as typed input-output transformations to compose multi-device workflows dynamically. Experiments in multi-OS virtual environments reveal that leading open-source agents struggle with state-transfer awareness and cross-platform contextual reasoning, exposing capability gaps that single-device evaluations hide.
Why it's worth reading
It breaks the single-device benchmark norm by testing how GUI agents handle cross-platform state transfers across Android, Windows, and Ubuntu in dynamic workflows.
Tags
GUI AgentBenchmarkCross-DeviceMulti-OSTask CompositionAgent Evaluation