AI & ML interests

UI-MOPD

Recent Activity

Organization Card

UI-MOPD: Multi-platform On-Policy Distillation for Continual GUI Agent Learning

We build cross-platform GUI agents that can operate both desktop and mobile interfaces through a unified training framework.

Research

UI-MOPD introduces a two-stage training pipeline:

  • Stage 1: Supervised Fine-Tuning (SFT) on platform-specific teacher models
  • Stage 2: Reinforcement Learning distillation (DAPO) with multi-teacher on-policy guidance

Our student model (8B) learns from multiple 32B teacher models to achieve strong cross-platform GUI interaction capabilities. All models are based on Qwen3-VL-Thinking (thinking/reasoning variants).

Models

ModelSizeDescription
Qwen3-VL-32B-Thinking-Desktop-Teacher32BDesktop platform teacher (thinking)
Qwen3-VL-32B-Thinking-Mobile-Teacher32BMobile platform teacher (thinking)
Qwen3-VL-8B-Thinking-Desktop-SFT8BDesktop SFT checkpoint (thinking)
Qwen3-VL-8B-Thinking-Mobile-SFT8BMobile SFT checkpoint (thinking)
Qwen3-VL-8B-Thinking-UI-MOPD-Student8BFinal cross-platform student (thinking)

Datasets

DatasetDescription
Uni-GUI-OpenCUAPost-processed desktop trajectories from OpenCUA (~832 episodes, ~14K steps)
Uni-GUI-Desktop-1Large-scale desktop GUI trajectories (~2.7K episodes, ~36K steps)
Uni-GUI-Desktop-2OSWorld desktop trajectories (~1.2K episodes, ~14.8K steps)
Uni-GUI-MobileMobile GUI trajectories (~871 episodes, ~14K steps)
Uni-GUI-OpenMobileOpen-source Android app trajectories (~2.6K episodes, ~25.9K steps)
AndroidControl*Static mobile GUI evaluation subset (4,260 step records, 781 episodes)

Evaluation Results

BenchmarkDescription
OSWorld-Eval-ResultsDesktop evaluation on OSWorld (359 tasks, 35.1% success rate)
MobileWorld-Eval-ResultsMobile evaluation on MobileWorld (117 tasks, 10.3% success rate)

Links