Back home

MAI-Gym

x 0.68
Visit source

Intro

MAI-Gym is an open, all-in-one platform for GUI-agent evaluation, data collection and online reinforcement learning (RL) across mobile, computer and web environments, released by the MAI-UI team at Alibaba (Tongyi / Alibaba Token Hub). The paper "MAI-Gym: An Open All-in-One Platform for GUI Agent Evaluation and RL Training across Mobile and Computer Environments" (arXiv:2610.20939) appeared in mid-October 2026 and was picked up by Chinese AI accounts on X on October 31, 2026.

Alongside the framework, the team trained and released MAI-Gym-9B (from Qwen3.5-9B), a single agent for both phone and desktop use that reaches 78.4% on MobileWorld and 76.2% on OSWorld-Verified — state of the art among open GUI agents under 10B parameters. MAI-Gym was also used to train Alibaba's Qwen-UI-Agent.

Explain

Key facts

  • Developer: MAI-UI team, Alibaba Group (Tongyi Lab / Alibaba Token Hub); corresponding authors include Steven Hoi
  • Paper: arXiv:2610.20939 (submitted Oct 14, 2026)
  • Scope: one shared rollout pipeline for evaluation and training, so agents behave identically in both
  • Benchmarks: 10 established benchmarks supported (MobileWorld, AndroidWorld, OSWorld-Verified, OSWorld 2.0, MacAgentBench, MyPCBench, MobileGym, …)
  • Scale: up to 20,000 parallel environment instances
  • Synthesis: automatically generates GUI apps and tasks — 49 resettable mocked apps, coherent cross-app worlds, execution-validated task verifiers
  • Released: framework, synthesis skills, mocked apps, 2,000 verified tasks, and the MAI-Gym-9B checkpoint

MAI-Gym-9B training recipe

  1. SFT warm-up on mixed mobile + desktop GUI/CLI data.
  2. Joint RL (GRPO) on 5,000 synthesized tasks from MobileWorld and OSWorld environments in a single run — instead of training per-domain experts and merging weights.
  3. Code-verified rewards, policy-adaptive task sampling and windowed history; ~96 hours on 512 GPUs.

Results

  • 78.4% MobileWorld, 76.2% OSWorld-Verified (SOTA among open GUI agents < 10B)
  • RL in a mocked train-ticketing app transferred to the real app, raising success on a physical phone from 28.8% to 48.1%
  • Training on entirely novel synthesized apps improved held-out success from 25.8% to 45.2%

Why it matters

GUI-agent research has suffered from fragmented evaluation harnesses and expensive interactive environments. MAI-Gym offers a reproducible, open stack that unifies them — useful for anyone building phone-use or computer-use agents.

Usage

Getting started

  1. Read the paper (arXiv:2610.20939) and the project page tongyi-mai.github.io/MAI-Gym.
  2. Clone the open-source framework from the Tongyi-MAI GitHub organisation and set up the environment backends you need (Android emulators, desktop VMs, web).
  3. Run evaluation on supported benchmarks with your own agent, or download the MAI-Gym-9B checkpoint as a baseline.
  4. For training, use mai-gym train with the provided configuration, and the app/task synthesis skills to create new mocked apps and verifiers.

Good fits

  • Benchmarking phone-use / computer-use agents under one consistent harness
  • Online RL for GUI agents with verifiable rewards
  • Generating synthetic but resettable app environments for safe training

Tips

  • Full-scale RL needs substantial compute (the paper used 512 GPUs); start with evaluation or small-scale runs.
  • Validate on real devices before deployment.