MAI-Gym: An Open All-in-One Platform for GUI Agent Evaluation and RL Training across Mobile and Computer Environments

MAI-Gym. An Open All-in-One Platform for GUI Agent Evaluation and RL Training across Mobile and Computer Environments. By Panrong Tong and fourteen others, from Alibaba Group. Published on the fourteenth of October. Its code repository is public and real: ten source files.

Online reinforcement learning with verifiable rewards has driven rapid progress in mathematics, coding and tool use. Carrying that success over to agents that operate graphical interfaces has proved harder, for two reasons. Interactive environments are expensive to run, and realistic training tasks are scarce. Online training with verifiable rewards leans on two things: somewhere for the agent to act, and tasks whose outcome can be checked. For an agent working on a phone or a computer screen, the first is costly and the second is in short supply.

MAI-Gym is an open, all-in-one platform that brings GUI agent evaluation, data collection and online reinforcement learning training together, across mobile, computer and web environments. The unifying piece is a shared rollout pipeline. Because evaluation and training run through the same pipeline, the agent behaves consistently in both. What an agent is trained on and what it is measured on pass through one path, and data collection sits on that path too.

On the infrastructure side, the platform supports ten established benchmarks and scales to twenty thousand parallel environment instances. Beyond scale, it tackles the shortage of tasks by synthesizing them automatically: diverse GUI applications and challenging tasks, including mocked apps that can be reset, and coherent worlds that span several apps, each with task verifiers that are validated by execution. A mocked app that resets gives a clean start for every attempt. A world spanning several apps gives tasks that cross from one application to another. And a verifier checked by actually running it is what lets a reward be trusted.

Using the platform, the authors train MAI-Gym-9B, a single nine billion parameter agent for both mobile and computer use. Among open GUI agents under ten billion parameters, it reaches state-of-the-art performance: seventy-eight point four percent on MobileWorld, and seventy-six point two percent on OSWorld-Verified.

A second result, which the authors flag as important, is about transfer. Reinforcement learning inside a mocked train-ticketing app carries over to the real app it imitates. On a physical phone, task success rises from twenty-eight point eight percent to forty-eight point one percent.

That mock-to-real step is what ties the two halves of the problem together. Synthetic apps answer the cost of environments and the scarcity of tasks, and the phone result shows training there can count in the real world. The release covers the framework, the synthesis skills, the mocked apps, two thousand verified tasks and the model checkpoint.