MobileGym: A Verifiable and Highly Parallel Simulation Platform for Mobile GUI Agent Research
TL;DR AI
2 min readKey summary
Researchers introduced MobileGym, a browser-hosted mobile simulation platform for mobile GUI agents.
It represents app state as structured JSON, making states controllable, forkable, and judgeable with deterministic evaluation.
MobileGym can run hundreds of parallel instances on one server and ships with MobileGym-Bench, a 416-task benchmark across 28 apps.
Training Qwen3-VL-4B-Instruct with GRPO improved benchmark results and transferred much of the gain to real-device tests.
The platform offers a scalable way to train and test mobile agents without relying as heavily on slow, brittle device-based evaluation.
