Reproduce “OpenMobile: Building Open Mobile Agents with Task and Trajectory Synthesis” 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
Kanzhi Cheng, Zehao Li, Zheng Ma, Nuo Chen, Jialin Cao, Qiushi Sun, Zichen Ding, Fangzhi Xu and 6 more
OpenMobile introduces an open-source framework and dataset for synthesizing task instructions and agent trajectories, decoupling exploration from instruction generation and incorporating error-recovery signals. Models trained on this data achieved up to 64.7% Pass@1 on the AndroidWorld benchmark, substantially outperforming existing open-data baselines and approaching closed-data industrial system performance.
Key insights
- Decoupling task instruction generation from environment exploration, using a global environment memory, yields more complex and sound instructions that lead to improved agent performance.
- Systematically incorporating error-recovery signals into training data via expert intervention during learner execution significantly enhances an agent's robustness and ability to handle deviations at test time.
- Broad and diverse functionality coverage in the synthesized data, rather than overfitting specific benchmark tasks, is a critical driver for achieving high mobile agent performance and generalization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.