Reproduce “Online Dense Reinforcement Learning with Planning-Grounding Agent for UI Navigation on Live IOS Apps” 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
Peitong Duan, Khashayar Gatmiry, Jeffrey Nichols, Samy Bengio, Emmanuel Abbe
The authors train an 8B vision-language planner through live interaction on iOS, using feedback on intermediate progress and task completion, while a separate grounder turns its decisions into screen actions. The planner reaches 32.8% task success on AndroidWorld, versus 4.7% for its supervised fine-tuning baseline. This shows transfer from iOS training to AndroidWorld, not broad generalization; evaluation uses Gemini-3.5-Flash grounding.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.