An 8B agent trained on live iOS apps tied Gemini-3.1-Pro at using real apps

Apple has described an AI agent that taps and swipes through real iOS apps, and a small version of it matched a much bigger rival. The evidence is a research paper from Apple Machine Learning Research, with Peitong Duan, Seyed Khashaiar Gatmiry and colleagues as authors.

The problem

Big vision-language models, which read a screen and decide what to press, are good at using apps. But they are costly to run, so they are hard to use every day.

Smaller models are cheaper. Yet when they learn from recorded examples, they tend to memorise the exact path. Then one unexpected pop-up can throw them off, and they struggle to recover.

What Apple did

The team trained the agent while it worked on live iOS apps, not on stored recordings. To make that possible, they built infrastructure that runs hundreds of virtual iOS devices at the same time.

The agent has two parts. A planning agent decides what to do next. A separate UI grounding model finds the right spot on the screen to act.

An AI judge scores the work. It gives small points for progress at each step, plus a final score for finishing the task. Imagine a teacher who marks each line of a maths answer, not only the final result.

The team also added a method called GiGPPO. It helps the training stay steady by comparing notes from screens that look alike across the many parallel runs.

The results

Tests used UINavBench, iOSWorld and AndroidWorld. The step-by-step reward agents beat both recorded-example training and training that only scored the final outcome. They also learned faster.

The 8B agents (8 billion parameters, a measure of model size) improved nearly sevenfold over the recorded-example baseline when moved to Android, a system they had not trained on. On real iOS apps, they beat Gemini-3.5-Flash and tied Gemini-3.1-Pro.

The source gives no figures for cost or speed, no timeline for any product, and no tests on tasks involving risky or irreversible actions. It is a research result, not a release.