Tongyi-MAI researchers introduced MAI-Gym, an open platform for evaluating GUI agents, collecting data and training them with online reinforcement learning across mobile, computer and web environments. The platform uses a shared rollout pipeline so agents behave consistently during evaluation and training, supports 10 established benchmarks and can run 20,000 environment instances in parallel. The team says high interaction costs and a shortage of realistic tasks have made it difficult to scale reinforcement learning for GUI agents.

MAI-Gym can generate diverse applications and tasks, including resettable simulated apps and linked tasks spanning multiple apps, with code-based verifiers that check task execution. Its release includes 49 simulated apps, 2,000 verified tasks, reusable synthesis skills and model weights. The researchers trained MAI-Gym-9B, a unified 9B-parameter agent for mobile and computer use, and report scores of 78.4% on MobileWorld and 76.2% on OSWorld-Verified, which they describe as state of the art among open GUI agents with fewer than 10B parameters.

In a simulated train-ticketing app, reinforcement learning improved task success on a physical phone from 28.8% to 48.1% when applied to the real-world counterpart. The team is releasing the framework, applications, tasks, synthesis skills and model checkpoint to support reproducible research and further work on GUI agent training. The paper was published on October 14 and submitted to Hugging Face Papers on October 16.

Tongyi-MAI researchers introduced MAI-Gym, an open platform for evaluating GUI agents, collecting data and training them with online reinforcement learning across mobile, computer and web environments. The platform uses a shared rollout pipeline so agents behave consistently during evaluation and training, supports 10 established benchmarks and can run 20,000 environment instances in parallel. The team says high interaction costs and a shortage of realistic tasks have made it difficult to scale reinforcement learning for GUI agents.

MAI-Gym can generate diverse applications and tasks, including resettable simulated apps and linked tasks spanning multiple apps, with code-based verifiers that check task execution. Its release includes 49 simulated apps, 2,000 verified tasks, reusable synthesis skills and model weights. The researchers trained MAI-Gym-9B, a unified 9B-parameter agent for mobile and computer use, and report scores of 78.4% on MobileWorld and 76.2% on OSWorld-Verified, which they describe as state of the art among open GUI agents with fewer than 10B parameters.

In a simulated train-ticketing app, reinforcement learning improved task success on a physical phone from 28.8% to 48.1% when applied to the real-world counterpart. The team is releasing the framework, applications, tasks, synthesis skills and model checkpoint to support reproducible research and further work on GUI agent training. The paper was published on October 14 and submitted to Hugging Face Papers on October 16.