A new preprint on arXiv argues that modern mobile GUI agents are limited because they can see the current screen and perform local actions, yet lack persistent, app-specific navigation knowledge. The authors describe a task-agnostic exploration framework that aims to fill this gap by learning where functions live, which pages contain them, and which screens are equivalent.
The framing distinction matters: not all app graphs are maps. The paper suggests that effective GUI navigation requires a structured understanding of an app's layout, not just reactive screen interpretation. By separating exploration from task-specific objectives, the framework could make agents more general and adaptable across different mobile apps.
As a single-source arXiv announcement, the abstract gives limited details on implementation and evaluation. Still, the core idea points toward a useful shift: giving mobile agents a durable sense of app geography rather than treating each screen as a fresh decision point.