Jev-Mobile: Jev as an Executor for Mobile GUI Agents
Jev-Mobile cuts Android GUI-agent API cost by 73.4% against a step-wise VLM baseline while staying near its task success rate.
The paper moves the VLM out of every interaction step: it sets local goals, while Jev repeatedly chooses executable actions from the Android accessibility tree. On AndroidWorld, it reports 79% task success, versus 84% for the step-wise VLM baseline and 78% for SeeAct-V. For successful runs, mean end-to-end execution time falls 32.7% relative to the step-wise VLM setup. ArXiv · AI/CL/LG's note
The paper moves the VLM out of every interaction step: it sets local goals, while Jev repeatedly chooses executable actions from the Android accessibility tree. On AndroidWorld, it reports 79% task success, versus 84% for the step-wise VLM baseline and 78% for SeeAct-V. For successful runs, mean end-to-end execution time falls 32.7% relative to the step-wise VLM setup. ArXiv · AI/CL/LG's note
score 5