Mobile GUI agents normally do everything by tapping and swiping, so a task that starts far from its target screen accumulates many navigation steps. This work adds deeplinks, which are Android entry points that open a specific screen directly, as an additional action. The authors mine candidate deeplinks with static analysis, execute them on real devices to see which work, and write natural-language descriptions of the screens they land on. That produces a verified deeplink catalog for the agent to use.
They then train GUI-Hopper, an agent that decides between invoking a deeplink and using ordinary GUI actions, with fallback when a deeplink is wrong or unavailable. On MobileWorld it completes more tasks with fewer steps than a GUI-only agent, across three model scales, and the advantage remains even when deeplinks are not available at inference. On MobileWorld-Real it also improves task success on commercial apps on real devices.
Why this matters: navigation is a major source of error in long-horizon mobile tasks. Treating deeplinks as first-class agent actions gives the agent a structured view of where it can go, reducing trajectory length and error opportunities without giving up the fine-grained control of GUI actions.