- AndroidWorld success rose from 4.7% to 32.8%, nearly 7x, for the PPO step-336 planner.
- The comparison uses a separate Gemini-3.5-Flash grounding model on AndroidWorld.
- The same checkpoint performed unevenly across three benchmarks; this is research, not an iPhone feature announcement.
Apple researchers report a nearly sevenfold jump in AndroidWorld task success, from 4.7% for their supervised-training baseline to 32.8% for an eight-billion-parameter planner trained with reinforcement learning. Their October 29 preprint describes a research result, not a new iPhone feature.1
The training rewards useful intermediate steps as well as completed tasks, rather than just copying examples or rewarding the final outcome. The nearly-sevenfold figure measures a success-rate ratio, not faster execution.1
There is a catch to the compact-model story. On AndroidWorld, a separate Gemini-3.5-Flash model translated plans into screen actions. Earlier supervised training already included Android examples; only the reinforcement-learning phase used iOS apps exclusively.1
The same PPO checkpoint tied Gemini-3.1-Pro at 32.4% on UINavBench, which uses real iOS apps. It scored 2.3% on synthetic-app benchmark iOSWorld, below the 4.5% baseline. Gemini-3.1-Pro scored 60.8% on AndroidWorld in the authors' setup. These are the authors' benchmark results, not an across-the-board win.1
Bloomboff's take
Seven times better sounds like a victory lap. A roughly one-in-three success score suggests keeping a hand on the wheel.
