Mobile devices can be hard to use, especially for older adults and people with visual impairments. Autonomous agents powered by vision-language models, AI that understands both images and text, could help, but current approaches require expensive datasets or slow trial-and-error learning.
MAPLE is a multi-agent framework that works without expensive retraining (gradient-free). It lets agents share real-time insights through a “playbook” of learned interface rules. Agents don't blindly repeat mistakes; they reference prior successful actions during each step.
On standard tests, MAPLE achieved 75.9% success on AndroidWorld and 60.25% on MobileWorld, outperforming earlier methods. This means more reliable, direct mobile assistance for users who delegate complex tasks to digital assistants.