MAI-Gym
library modelYour notes
Online RL with verifiable rewards has carried maths, coding and tool use a long way, and GUI agents have been left behind for a dull reason: interactive environments are expensive and realistic training tasks are scarce. MAI-Gym is Alibaba's attempt to remove both constraints at once — one platform unifying evaluation, data collection and online RL training across mobile, computer and web, from the MAI-UI team at Alibaba Token Hub (not Microsoft, despite the MAI- prefix this index also uses for MAI-Thinking-1 and the MAI-Code line).
The design choice that matters is a shared rollout pipeline, so an agent behaves identically whether it is being evaluated or trained — which is what makes an RL gain legible as a real gain rather than an artefact of two different harnesses. It covers ten established benchmarks and scales to 20,000 parallel environment instances. On top of the infrastructure it synthesises its own subjects: resettable mocked apps, coherent cross-app worlds, and tasks whose verifiers are execution-validated rather than written by hand.
The sim-to-real result is the one to read. RL inside a mocked train-ticketing app transferred to the real application, lifting task success on a physical phone from 28.8% to 48.1%. That is the claim the whole synthetic-environment approach stands on, and it is reported on one app rather than a suite, so treat it as a strong existence proof rather than a measured transfer rate. MAI-Gym-9B, the unified mobile-and-computer agent trained with it, reports 78.4% on MobileWorld and 76.2% on OSWorld-Verified, which Alibaba puts as state of the art among open GUI agents under 10B.
Released as a framework plus synthesis skills, the mocked apps, 2,000 verified tasks and the checkpoint. The paper is CC BY-NC-SA 4.0 and the repository declares no licence at filing, so the terms on the code are unsettled even though the weights are downloadable. The title page also notes MAI-Gym was used to train Qwen-UI-Agent, which is not yet tracked here.
Outputs 2
MAI-Gym
libraryThe platform itself: shared rollout pipeline across evaluation and online RL, ten benchmarks, 20,000 parallel environment instances, plus automated synthesis of mocked apps and execution-validated task verifiers. Ships 2,000 verified tasks.
MAI-Gym-9B
modelUnified 9B GUI agent for mobile and computer use, trained with MAI-Gym. Reported state of the art among open GUI agents under 10B.
Benchmark Scores
| Benchmark | Score | Mode |
|---|---|---|
| MobileWorld | 78.4 | — |
| OSWorld-Verified | 76.2 | — |