AI
MAI-Gym links GUI-agent training and evaluation across devices
MAI-Gym is an open platform for evaluating and training GUI agents across mobile, desktop and web environments. Its shared rollout system supports ten benchmarks and up to 20,000 parallel environments, while synthesized applications provide resettable tasks and execution-validated checks. The authors report that a 9B agent reached leading results among open sub-10B agents and that training in a mocked ticketing app improved success on its real counterpart. The results come from a preprint and remain to be independently assessed.
Editorial summary · Publisher article
What matters
- The platform combines evaluation, data collection and online reinforcement learning in one rollout pipeline.
- The reported MAI-Gym-9B scores were 78.4% on MobileWorld and 76.2% on OSWorld-Verified.
- Training on a mocked ticketing app reportedly raised physical-phone success from 28.8% to 48.1%.
Context & caveats
- The supplied evidence does not provide task-level error analysis, comparisons beyond the cited open-agent category, or independent replication.
Why this made the edition
The paper presents an open training and evaluation platform, concrete benchmark results, a transfer result to a physical phone, and released artifacts. The evidence is abstract-level and author-reported, but it contains useful technical and reproducibility detail.
Source material