Understand this story
On November 2, 2026, Apple Machine Learning Research proposed an online reinforcement learning framework that trains UI navigation agents directly on real iOS applications. The framework splits the agent into a Planning agent and a UI Grounding model, and uses an LLM-as-a-Judge to provide step-level dense rewards and task-completion terminal rewards, optimizing with PPO plus step-level advantage estimation. According to its disclosure, when the 8B agent generalizes out-of-distribution to the Android environment, it improves nearly 7-fold over the SFT baseline, and on real iOS applications it surpasses Gemini-3.5-Flash and matches Gemini-3.1-Pro. There are currently no follow-up developments.
AI-generated from public reports
Developments
1 confirmed developments
Apple proposes an online reinforcement learning UI navigation framework for real-time iOS apps
Source:Apple Machine Learning Research
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Nov 2
Apple Machine Learning ResearchFirst-party
Apple proposes an online dense reinforcement learning framework: a planning-grounding agent navigating real iOS applications
Apple proposes an online reinforcement learning framework that directly trains a UI navigation agent on real iOS applications, splitting the agent into a Planning agent and a UI Grounding model, and uses an LLM-as-a-Judge to provide step-level dense rewards and task-completion terminal rewards, optimizing with PPO plus step-level advantage estimation. Its 8B agent achieves nearly a 7x improvement over the SFT baseline when generalizing out-of-distribution to the Android environment, and on real iOS applications surpasses Gemini-3.5-Flash and matches Gemini-3.1-Pro.