| --- |
| title: README |
| emoji: '🖥️' |
| colorFrom: blue |
| colorTo: purple |
| sdk: static |
| pinned: false |
| --- |
| |
| <p align="center" style="font-size: 1.4em; font-weight: bold;"> |
| 🖥️📱 UI-MOPD |
| </p> |
|
|
| <p align="center"> |
| <b>Multi-platform On-Policy Distillation for Continual GUI Agent Learning</b> |
| </p> |
|
|
| <p> |
| We build cross-platform GUI agents that can operate both desktop and mobile interfaces through a unified training framework. |
| </p> |
|
|
| <h3>🔬 Research</h3> |
|
|
| <p> |
| UI-MOPD introduces a two-stage training pipeline: |
| </p> |
| <ul> |
| <li><b>Stage 1</b>: Supervised Fine-Tuning (SFT) on platform-specific teacher models</li> |
| <li><b>Stage 2</b>: Reinforcement Learning distillation (DAPO) with multi-teacher on-policy guidance</li> |
| </ul> |
| <p> |
| Our student model (8B) learns from multiple 32B teacher models to achieve strong cross-platform GUI interaction capabilities. All models are based on <b>Qwen3-VL-Thinking</b> (thinking/reasoning variants). |
| </p> |
|
|
| <h3>🤖 Models</h3> |
|
|
| <table> |
| <tr><th>Model</th><th>Size</th><th>Description</th></tr> |
| <tr><td><a href="https://huggingface.co/UI-MOPD/Qwen3-VL-32B-Thinking-Desktop-Teacher">Qwen3-VL-32B-Thinking-Desktop-Teacher</a></td><td>32B</td><td>Desktop platform teacher (thinking)</td></tr> |
| <tr><td><a href="https://huggingface.co/UI-MOPD/Qwen3-VL-32B-Thinking-Mobile-Teacher">Qwen3-VL-32B-Thinking-Mobile-Teacher</a></td><td>32B</td><td>Mobile platform teacher (thinking)</td></tr> |
| <tr><td><a href="https://huggingface.co/UI-MOPD/Qwen3-VL-8B-Thinking-Desktop-SFT">Qwen3-VL-8B-Thinking-Desktop-SFT</a></td><td>8B</td><td>Desktop SFT checkpoint (thinking)</td></tr> |
| <tr><td><a href="https://huggingface.co/UI-MOPD/Qwen3-VL-8B-Thinking-Mobile-SFT">Qwen3-VL-8B-Thinking-Mobile-SFT</a></td><td>8B</td><td>Mobile SFT checkpoint (thinking)</td></tr> |
| <tr><td><a href="https://huggingface.co/UI-MOPD/Qwen3-VL-8B-Thinking-UI-MOPD-Student">Qwen3-VL-8B-Thinking-UI-MOPD-Student</a></td><td>8B</td><td>Final cross-platform student (thinking)</td></tr> |
| </table> |
|
|
| <h3>📊 Datasets</h3> |
|
|
| <table> |
| <tr><th>Dataset</th><th>Description</th></tr> |
| <tr><td><a href="https://huggingface.co/datasets/UI-MOPD/Uni-GUI-OpenCUA">Uni-GUI-OpenCUA</a></td><td>Post-processed desktop trajectories from OpenCUA (~832 episodes, ~14K steps)</td></tr> |
| <tr><td><a href="https://huggingface.co/datasets/UI-MOPD/Uni-GUI-Desktop-1">Uni-GUI-Desktop-1</a></td><td>Large-scale desktop GUI trajectories (~2.7K episodes, ~36K steps)</td></tr> |
| <tr><td><a href="https://huggingface.co/datasets/UI-MOPD/Uni-GUI-Desktop-2">Uni-GUI-Desktop-2</a></td><td>OSWorld desktop trajectories (~1.2K episodes, ~14.8K steps)</td></tr> |
| <tr><td><a href="https://huggingface.co/datasets/UI-MOPD/Uni-GUI-Mobile">Uni-GUI-Mobile</a></td><td>Mobile GUI trajectories (~871 episodes, ~14K steps)</td></tr> |
| <tr><td><a href="https://huggingface.co/datasets/UI-MOPD/Uni-GUI-OpenMobile">Uni-GUI-OpenMobile</a></td><td>Open-source Android app trajectories (~2.6K episodes, ~25.9K steps)</td></tr> |
| <tr><td><a href="https://huggingface.co/datasets/UI-MOPD/AndroidControl-Star">AndroidControl*</a></td><td>Static mobile GUI evaluation subset (4,260 step records, 781 episodes)</td></tr> |
| </table> |
| |
| <h3>📈 Evaluation Results</h3> |
| |
| <table> |
| <tr><th>Benchmark</th><th>Description</th></tr> |
| <tr><td><a href="https://huggingface.co/datasets/UI-MOPD/OSWorld-Eval-Results">OSWorld-Eval-Results</a></td><td>Desktop evaluation on OSWorld (359 tasks, 35.1% success rate)</td></tr> |
| <tr><td><a href="https://huggingface.co/datasets/UI-MOPD/MobileWorld-Eval-Results">MobileWorld-Eval-Results</a></td><td>Mobile evaluation on MobileWorld (117 tasks, 10.3% success rate)</td></tr> |
| </table> |
| |
| <h3>🔗 Links</h3> |
| |
| <ul> |
| <li>🌐 <a href="https://elispctre.github.io/UI-MOPD">Project Page</a></li> |
| </ul> |
| |
