Jev-Mobile uses a VLM only to define local goals and exact input values. A typed decision model then selects multiple actions from candidates generated from the latest Android accessibility tree, returning control when finished or blocked.
On the full AndroidWorld suite, it achieved 79% task success, versus 78% for SeeAct-V and 84% for a step-wise VLM baseline. Among successful trajectories, it reduced mean execution time by 32.7% and mean model API cost by 73.4% relative to the step-wise baseline.
The result suggests that expensive multimodal reasoning need not run at every interaction step. Structured interfaces can support a cheaper execution loop while preserving much of the stronger baseline's task performance.