How MobileForge Trains GUI Agents Without Any Labels — and Nails 77.6% Accuracy

Mobile GUI agents powered by multimodal LLMs have shown promise in understanding and executing actions on smartphone screens, but adapting them to new or updated apps remains prohibitively costly due to the need for human-written tasks, demonstrations, and reward labels. A team of researchers from multiple institutions (including lead author Guangyi Liu) addresses this with MobileForge, an annotation-free adaptation framework that eliminates manual supervision entirely.

MobileForge consists of two core components: MobileGym, which grounds task generation and rollout evaluation in real mobile app interaction, and HiFPO (Hierarchical Feedback-Guided Policy Optimization), a novel algorithm that converts trajectory outcomes, step-level process feedback, and corrective hints into step-level GRPO updates. Using purely automatically generated data, MobileForge adapts the Qwen3-VL-8B model to achieve 67.2% Pass@3 on the AndroidWorld benchmark—nearly matching the closed-data GUI-Owl-1.5-8B baseline at 69.0%. Moreover, the MobileForge-adapted ForgeOwl-8B reaches 77.6% Pass@3 on AndroidWorld and 41.0% success on the out-of-domain MobileWorld GUI-only split, establishing the strongest open-data mobile GUI agent reported to date. Code, data, and trained models are scheduled for open release.