Apple trains compact SI agent to operate apps on live iPhones
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
A small agent can learn to recover from errors directly inside real apps and transfer that skill to another mobile operating system without a huge cloud model.
Apple researchers split the agent into a planning module and a module grounding actions in interface elements. Reinforcement learning took place in live iOS apps across hundreds of concurrent virtual devices rather than on recorded trajectories. A judge model assessed both intermediate steps and final outcomes, while GiGPPO used visually similar states from parallel attempts to stabilize training. The 8-billion-parameter agent outperformed variants trained on prepared examples and those rewarded only for the final result. When transferred to Android, its generalization was almost seven times the SFT baseline; on real iOS apps it tied Gemini-3.1-Pro and outperformed Gemini-3.5-Flash. This is a research result on UINavBench, iOSWorld and AndroidWorld, not a consumer feature announcement.
Sources 1