Reproduce “MobileForge: Annotation-Free Adaptation for Mobile GUI Agents with Hierarchical Feedback-Guided Policy Optimization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Guangyi Liu, Pengxiang Zhao, Gao Wu, Yiwen Yin, Mading Li, Liang Liu, Congxiao Liu, Zhang Qi and 4 more
MobileForge introduces an annotation-free system for adapting mobile GUI agents, employing a unified interaction substrate called MobileGym and a hierarchical feedback-guided policy optimization method named HiFPO. The system enables both generalist and GUI-specialized MLLMs to enhance their performance on in-domain tasks and generalize effectively to out-of-domain scenarios, achieving up to 77.6% Pass@3 on AndroidWorld for an adapted GUI-Owl-1.5-8B.
Key insights
- Providing fine-grained, hierarchical feedback, including corrective hints from previous attempts, is essential for agents to learn effectively from self-collected experience in mobile GUI environments.
- A unified substrate like MobileGym, which grounds task generation and evaluation in real app interaction, ensures tasks are functionally relevant and feedback is precise.
- Adapting Group-Relative Policy Optimization with hint-contextualized, step-level rewards and filtering for 'reasonable' actions provides superior policy refinement compared to traditional supervised fine-tuning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.