Found first: a primary source the press has not covered yet.
Researchers at Apple trained an 8-billion-parameter agent to navigate iOS applications using online reinforcement learning run directly on live iOS devices, eliminating the need for pre-recorded trajectories. On out-of-distribution Android tasks (AndroidWorld), the agent reaches 31.9% task success against the supervised fine-tuning baseline's 4.7%, a nearly sevenfold gain. The paper, submitted October 29, 2026, is available at arxiv.org/abs/2610.39264.
What the source says
Authors Peitong Duan, Khashayar Gatmiry, Jeffrey Nichols, Samy Bengio, and Emmanuel Abbe at Apple built a two-component system: a Planning Agent that decomposes tasks and a UI Grounding Model that resolves screen elements. Training runs GiGPPO, a neighbor-conditioned PPO variant the paper describes as improving critic stability through step-level LLM-as-Judge rewards, orchestrated across hundreds of concurrent virtual iOS devices. On UINavBench, which tests on real iOS applications, the RL-trained 8B model scores 32.4%, matching Gemini-3.1-Pro's 32.4% and surpassing Gemini-3.5-Flash's 28.4%. On iOSWorld, GiGPPO reaches 9.8% against Gemini-3.1-Pro's 39.5%, so parity does not extend across all iOS benchmarks. The AndroidWorld result (31.9% vs. SFT baseline 4.7%) comes from a model trained on iOS, with no Android-specific training.
Why it matters
Most phone-UI agent research trains on static, recorded trajectories and never encounters the failures a live device produces. Training in the live environment closes that gap. An 8B model matching a frontier system on a real-device benchmark at a fraction of the parameter count matters for anyone trying to run phone agents outside a cloud budget. The Android generalization result is the more surprising finding: the representations learned on iOS transfer substantially to a different OS without any Android-specific fine-tuning, which suggests the agent is learning something closer to general UI interaction than platform-specific pattern matching.