ForgeOwl-8B-200tasks
ForgeOwl-8B-200tasks is the GUI-Owl-1.5-8B-Instruct policy adapted with MobileForge on 200 automatically generated target-app tasks. MobileForge uses the policy's own rollouts, hierarchical critic feedback, corrective hints, and hint-contextualized step-level GRPO. No human-written adaptation tasks, demonstrations, or reward labels are used.
This repository is part of an anonymous ICLR submission artifact. Author and paper-identifying metadata will be added after review.
Anonymous project page: https://mobileforge-anonymous.github.io/
Evaluation
On AndroidWorld (116 tasks), this checkpoint obtains 75/116 (64.7%) Pass@1, 85/116 (73.3%) Pass@2, and 86/116 (74.1%) Pass@3.
Reproduction code and raw evaluation archives are available from the anonymous code repository and benchmark-results dataset.
Usage and limitations
Use the loading and prompting interface of mPLUG/GUI-Owl-1.5-8B-Instruct. The model can make incorrect or unsafe GUI actions; run it only in isolated test environments and inspect actions before using it with personal data.