RESEARCH · RESEARCH · #2088
MAI-Gym: open platform and MAI-Gym-9B for scalable GUI agent training and evaluation
MAI-Gym is an open all-in-one platform that unifies GUI agent evaluation, data collection, and online RL training across mobile, desktop and web, supporting ten benchmarks and scaling to 20,000 parallel environment instances while enabling automated app synthesis and execution-validated task verifiers. Using this platform the authors train MAI-Gym-9B (9B parameters), report 78.4% on MobileWorld and 76.2% on OSWorld-Verified, show transfer from a mocked train-ticketing app to a physical phone (task success 28.8% → 48.1%), and release the framework, synthesis skills, mocked apps, 2,000 verified tasks and the model checkpoint (arXiv:2610.20939v1).
KEY POINTS
- MAI-Gym is an open all-in-one platform that unifies GUI agent evaluation, data collection, and online RL training across mobile, desktop and web, supporting ten benchmarks and scaling to 20,000 parallel environment instances while enabling automated app synthesis and execution-validated task verifiers.
- Using this platform the authors train MAI-Gym-9B (9B parameters), report 78.4% on MobileWorld and 76.2% on OSWorld-Verified, show transfer from a mocked train-ticketing app to a physical phone (task success 28.8% → 48.1%), and release the framework, synthesis skills, mocked apps, 2,000 verified tasks and the model checkpoint (arXiv:2610.20939v1).
- This release provides an open, scalable infrastructure, large synthetic task set and a released 9B checkpoint that advance reproducible RL training and state-of-the-art results for GUI agents under 10B, including demonstrated real-device transfer.
WHY IT MATTERS
This release provides an open, scalable infrastructure, large synthetic task set and a released 9B checkpoint that advance reproducible RL training and state-of-the-art results for GUI agents under 10B, including demonstrated real-device transfer.