An October 9 preprint reports resettable app training and a separate physical-phone test.
The big change
- What changed: The team built 35 commercial-style simulated apps with checkable state.
- Why it matters: Controlled tasks let researchers compare screen actions and experimental app-native tools; they do not prove live-app utility.
- What to watch: One physical-phone benchmark tests transfer, not everyday reliability.
The measured boundary
MobileGym++ checks 215 simulated tasks. On SPA-Bench L3 physical phones, the 27B model scored 57.4% after hybrid supervised training and 59.6% after reinforcement learning: a 2.2 percentage-point gain. BIG CHANGE did not run either test.
