Online Dense Reinforcement Learning with Planning-Grounding Agent for UI Navigation on Live iOS Apps

Researchers developed an online reinforcement learning framework with a Planning agent and UI Grounding model to train UI navigation agents on live iOS apps, achieving strong performance and generalization to Android. Their method uses an LLM-as-a-Judge for detailed reward feedback and introduces GiGPPO for improved value estimation, outperforming existing baselines and Gemini models.