File size: 19,179 Bytes
3f4631a |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 |
---
license: apache-2.0
datasets:
- GUIAgent/Magic-RICH
language:
- en
base_model:
- Qwen/Qwen2-VL-7B-Instruct
---
## News
* [2025-07-20] 📄📄📄 We have released the **technical report** of MagicGUI! Check it out [here](https://arxiv.org/abs/2508.03700).
* [2025-07-20] 🚀🚀🚀 We have open-sourced **MagicGUI**, an on-device GUI agent capable of operating Chinese & English apps and equipped with RFT-enhanced reasoning abilities.
## Overview
MagicGUI is an open-source GUI agent model developed by Honor, built on Qwen2-VL with 7 billion parameters. It demonstrates outstanding capabilities in visual grounding, screen question answering, and action sequence planning and execution. MagicGUI enables multimodal perception, understanding, and automated execution of user tasks on mobile devices.
**Data Collection Framework**: Propose a scalable and modular framework for GUI data collection that efficiently gathers high-quality data on mobile devices.
**Powerful Perception and Grounding Capabilities**: Enhance the perception and grounding abilities on mobile device screens by integrating large-scale knowledge through tasks such as element referring, element grounding, and screen captioning.
**Unified Action Space**: Develop a comprehensive and unified action space for various mobile platforms, encompassing fundamental operations like Tap, Text Input, and Scroll, while also supporting more complex actions such as Wait, Drag, and Takeover.
**Planning-Oriented Reasoning**: Implement a planning-oriented reasoning mechanism to improve the stability of task execution and enhance the accuracy of action decisions in dynamic environments.
**Two-Stage Training Paradigm**: Strengthen core perception, localization, and navigation capabilities through Continued Pre-training (CPT), while enhancing model robustness and generalization via Reinforcement Fine-tuning (RFT).
## Framework
The overall training framework of our MagicGUI contains two stages:
**Stage I**: Continue Pre-training (CPT), which involves training a
foundational model on a large and diverse dataset followed by an annealing phase using a balanced and high-quality
dataset.
**Stage II**: Reinforcement Fine-tuning (RFT), aimed at further enhancing the
model’s robustness and generalization capabilities.
## Quick Start
### Install dependencies
```bash
git clone https://github.com/MagicAgent-GUI
cd MagicGUI
conda create -n gui_agent python=3.11
conda activate gui_agent
pip install -r requirements.txt
```
### Download the model
Download [MagicGUI-RFT](https://huggingface.co/GUIAgent/MagicGUI_RFT) and [MagicGUI-CPT](https://huggingface.co/GUIAgent/MagicGUI_CPT).
#### Huggingface Inference
```python
import torch
from utils.model import Qwen2VLChat
# 1. Load the model and tokenizer
model_path = "./models/RFT" # model path
model = Qwen2VLChat.from_pretrained(model_path, min_pixels=4*28*28, max_pixels=768*28*28)
model = model.to("cuda:0")
# 2. Build the input
instruction = """你是一个训练有素的手机智能体，能够帮助用户进行单步导航任务。已知当前智能手机的截图<image>，和用户指令"查看会员信息"请输出正确的函数调用以实现用户指令。除了函数调用之外，你不能输出任何其他内容。你可以调用以下函数来控制智能手机：- UI基础操作：1. tap(x: float,y: float) 该函数用于在智能手机屏幕上点击特定点。坐标 x 和 y 表示待点击控件的中心位置。2. scroll(x: float,y: float,direction: str) 该函数用于从起始坐标 (x,y) 开始在智能手机屏幕上滑动操作，方向为手指滑动的方向。坐标 x 和 y 表示屏幕上待滑动控件的中心位置。方向可以是 "up"、"down"、"left" 或 "right"。3. text(x: float,y: float,text_input: str) 该函数用于在智能手机屏幕上输入指定的text。坐标 x 和 y 表示待点击控件的中心位置。- 手机按键操作：4. navigate_back() 该函数用于返回智能手机的上一个屏幕。5. navigate_home() 该函数用于返回手机的home screen或关闭当前应用。- 其他操作：6. long_press(x: float,y: float) 该函数用于在智能手机屏幕上的特定点执行长按操作。坐标 x 和 y 表示待点击控件的中心位置。7. wait() 该函数表示在当前页面等候。8. enter() 该函数表示按下enter键。9. take_over(text_input: str) 该函数用于提示用户接管智能手机，其中 text_input 是提示用户接管手机的原因。如果原因不确定，请填写“请您接管当前界面”。10. drag(x1: float,y1: float,x2: float,y2: float) 该函数执行一个对起始和终点敏感的拖动操作，表示手指从点1拖到点2。常见的场景包括滑块拖动、滚动选择器拖动和图片裁剪。11. screen_shot() 该函数用于截图。12. long_screen_shot() 该函数执行长截图。13. call_api(api_name: str,params: str) 调用指定的API并传入给定的参数。api_name是API的名称。params包含API所需的输入参数。例如，call_api(Amazon, open)意味着打开亚马逊APP。如果你发现当前指令无法在当前页面上执行，你需要输出no_answer。如果你发现当前指令已完成，你需要输出action_completed。"""
image_path = "./assets/test_action.png"
# 3. Build the message format
messages = [{"type": "image", "value":f"{image_path}",
{"type": "text", "value":f"{instruction}"]
# 4. Inference
response = model.generate(
message = messages,
)
print(response)
```
Expected output:
```JSON
{"tap(700,964)"}
```
### Action Space
At each step, the agent outputs is a single JSON object that contains:
- One (and only one) primitive action, chosen from the list below;
- Optional modifiers (`duration`, `thought`) and/or a task-level flag (`STATUS`).
Note that all keywords are **case-sensitive**, and we use **compact JSON** (i.e., no extra whitespace), which affects the tokenizer’s behavior.
<table>
<thead>
<tr>
<th>Action</th>
<th>Description</th>
<th>Conditions for R<sub>acc</sub> = +2</th>
<th>Example</th>
</tr>
</thead>
<tbody>
<tr>
<td><b>Tap</b></td>
<td>Click at coordinate (x, y)</td>
<td>dist([x, y], [x<sub>c</sub>, y<sub>c</sub>]) ≤ 14%</td>
<td><code>tap(x,y)</code></td>
</tr>
<tr>
<td><b>Scroll</b></td>
<td>Scroll at coordinate (x, y) with<br>direction up / down / left / right</td>
<td>dist([x, y], [x<sub>c</sub>, y<sub>c</sub>]) ≤ 14%<br>and direction = gt[direction]</td>
<td><code>scroll(x,y,direction)</code></td>
</tr>
<tr>
<td><b>Text Input</b></td>
<td>Type <i>text</i> at coordinate (x, y)</td>
<td>dist([x, y], [x<sub>c</sub>, y<sub>c</sub>]) ≤ 14%<br>and F1(text, gt[text]) > 0.5</td>
<td><code>text(x,y,text_input)</code></td>
</tr>
<tr>
<td><b>Navigation Back</b></td>
<td>Adb command to go back to the previous page</td>
<td>–</td>
<td><code>navigate_back()</code></td>
</tr>
<tr>
<td><b>Navigation Home</b></td>
<td>Adb command to go to the home screen of the mobile</td>
<td>–</td>
<td><code>navigate_home()</code></td>
</tr>
<tr>
<td><b>Long Press</b></td>
<td>Long press at coordinate (x, y)</td>
<td>dist([x, y], [x<sub>c</sub>, y<sub>c</sub>]) ≤ 14%</td>
<td><code>long_press(x,y)</code></td>
</tr>
<tr>
<td><b>Finish</b></td>
<td>Indicate that navigation task has been completed</td>
<td>–</td>
<td><code>finish()</code></td>
</tr>
<tr>w
<td><b>Wait</b></td>
<td>Wait for several seconds</td>
<td>–</td>
<td><code>wait()</code></td>
</tr>
<tr>
<td><b>Enter</b></td>
<td>Adb command to press enter</td>
<td>–</td>
<td><code>enter()</code></td>
</tr>
<tr>
<td><b>Takeover</b></td>
<td>Request user takeover</td>
<td>–</td>
<td><code>take_over(message)</code></td>
</tr>
<tr>
<td><b>Drag</b></td>
<td>Drag from coordinate (x₁, y₁) to (x₂, y₂)</td>
<td>
dist([x₁, y₁], [x<sub>1c</sub>, y<sub>1c</sub>]) ≤ 7.5%<br>
and dist([x₂, y₂], [x<sub>2c</sub>, y<sub>2c</sub>]) ≤ 7.5%
</td>
<td><code>drag(x1,y1,x2,y2)</code></td>
</tr>
<tr>
<td><b>Call API</b></td>
<td>Adb command to <i>open</i> or <i>kill</i> app</td>
<td>app = gt[app]<br>and open/kill = gt[operation]</td>
<td><code>call_api(api_name,operation)</code></td>
</tr>
<tr>
<td><b>Screenshot</b></td>
<td>Adb command to take a screenshot</td>
<td>–</td>
<td><code>screen_shot()</code></td>
</tr>
<tr>
<td><b>Long Screenshot</b></td>
<td>Adb command to take a long screenshot</td>
<td>–</td>
<td><code>long_screen_shot()</code></td>
</tr>
</tbody>
</table>
## Evaluation
### 1.Data preparation
Please download the four compressed files from the [Magic-RICH dataset](https://huggingface.co/datasets/GUIAgent/Magic-RICH) and extract them into the .datasets/ directory.
- `assets/`
- `datasets/`
- `Routine`
- `Instruction`
- `Complex`
- `Handing_Exception`
- `utils/`
For the preparation of other open-source datasets, please refer to [Other datasets preparation](datasets/eval_data_process/readme.md).
### 2. Param
We use run_eval.py for evaluation.
- `--data`: Name of a eval dataset
- `--model`: Path to the model
- `--work-dir (str, default to '.')`: Directory to save evaluation results
- `--mode (str, default: 'all', choices: ['all', 'infer'])`: If set to "all", the script performs both inference and evaluation; if set to "infer", it performs inference only.
- `--eval_model_path (str, default: 'None')`:'Path to eval model (required if mode is 'all' and data is 'ScreenQA-short')'
### 3. Run
```python
# Referring Benchmark
python run_eval.py --data ScreenQA-short --model MagicGUI_Path --mode all --eval_model_path Eval_Model_Path
python run_eval.py --data ScreenSpot_v2_mobile --model MagicGUI_Path --mode all
python run_eval.py --data Os-Atlas-mobile --model MagicGUI_Path --mode all
# Magic-RICH dataset
python run_eval.py --data Routine --model MagicGUI_Path --mode all
python run_eval.py --data Complex --model MagicGUI_Path --mode all
python run_eval.py --data Instruction --model MagicGUI_Path --mode all
python run_eval.py --data Handling_Exception --model MagicGUI_Path --mode all
# Open-source AndroidControl and GUI-Odyssey
python run_eval.py --data AC-Low --model MagicGUI_Path --mode all
python run_eval.py --data AC-High --model MagicGUI_Path --mode all
python run_eval.py --data GUI-Odyssey --model MagicGUI_Path --mode all
```
## Performance Evaluation
### Performance comparison on the Referring Benchmark
<table>
<thead>
<tr>
<th rowspan="1">Agent Models</th>
<th colspan="1">ScreenQA-short</th>
<th colspan="1">ScreenSpot v2 mobile</th>
<th colspan="1">Os-Atlas-mobile</th>
</tr>
</thead>
<tbody>
<!-- Closed-source Models -->
<tr><td colspan="4"><em>Closed-source Models</em></td></tr>
<tr>
<td>GPT-4o (Hurst et al., 2024)</td>
<td>90.3</td><td>10.6</td><td>4.6</td>
</tr>
<tr>
<td>Gemini 2.0 (Pichai et al., 2024)</td>
<td>90.4</td><td>10.6</td><td>5.8</td>
</tr>
<!-- Open-source Models -->
<tr><td colspan="4"><em>Open-source Models</em></td></tr>
<tr>
<td>InternVL-2-8B (Chen et al., 2024)</td>
<td>88.4</td><td>4.2</td><td>2.4</td>
</tr>
<tr>
<td>Qwen2-VL-7B (Wang et al., 2024)</td>
<td>92.6</td><td>70.7</td><td>27.2</td>
</tr>
<tr>
<td>Qwen2.5-VL-7B (Bai et al., 2025)</td>
<td>92.1</td><td>56.1</td><td>26.6</td>
</tr>
<tr>
<td>UI-TARS-7B (Qin et al., 2025)</td>
<td><b>95.4</b></td><td>88.6</td><td>82.5</td>
</tr>
<tr>
<td>UI-TARS-1.5-7B (Seed, 2025)</td>
<td>93.0</td><td>85.8</td><td>79.3</td>
</tr>
<!-- MagicGUI -->
<tr style="background-color:#e8eafc;">
<td>MagicGUI-CPT</td>
<td>94.6</td><td><b>90.2</b></td><td><b>95.2</b></td>
</tr>
</tbody>
</table>
### Performance comparison on the Magic-RICH dataset
<table>
<thead>
<tr>
<th rowspan="2">Agent Models</th>
<th colspan="3">Routine</th>
<th colspan="3">Instruction</th>
<th colspan="3">Complex</th>
<th rowspan="2">Handing Exception</th>
</tr>
<tr>
<th>Type</th><th>Grd</th><th>SR</th>
<th>Type</th><th>Grd</th><th>SR</th>
<th>Type</th><th>Grd</th><th>SR</th>
</tr>
</thead>
<tbody>
<!-- Closed-source Models -->
<tr><td colspan="11"><em>Closed-source Models</em></td></tr>
<tr>
<td>GPT-4o (Hurst et al., 2024)</td>
<td>49.3</td><td>16.7</td><td>4.6</td>
<td>56.6</td><td>13.5</td><td>19.8</td>
<td>49.0</td><td>14.6</td><td>7.4</td>
<td>85.1</td>
</tr>
<tr>
<td>Gemini 2.0 (Pichai et al., 2024)</td>
<td>89.2</td><td>49.4</td><td>34.7</td>
<td>84.1</td><td>54.2</td><td>51.4</td>
<td>83.3</td><td>50.3</td><td>42.0</td>
<td>73.7</td>
</tr>
<!-- Open-source Models -->
<tr><td colspan="11"><em>Open-source Models</em></td></tr>
<tr>
<td>InternVL-2-8B (Chen et al., 2024)</td>
<td>30.1</td><td>2.8</td><td>1.3</td>
<td>37.1</td><td>4.0</td><td>15.8</td>
<td>17.1</td><td>6.0</td><td>1.3</td>
<td>70.8</td>
</tr>
<tr>
<td>Qwen2-VL-7B (Wang et al., 2024)</td>
<td>71.7</td><td>41.0</td><td>28.1</td>
<td>73.6</td><td>43.9</td><td>41.5</td>
<td>65.6</td><td>28.7</td><td>21.2</td>
<td>68.3</td>
</tr>
<tr>
<td>Qwen2.5-VL-7B (Bai et al., 2025)</td>
<td>94.3</td><td>92.6</td><td>76.3</td>
<td>89.3</td><td><u>95.7</u></td><td>83.6</td>
<td>86.6</td><td>69.6</td><td>60.0</td>
<td>67.0</td>
</tr>
<tr>
<td>UI-TARS-7B (Qin et al., 2025)</td>
<td>83.5</td><td>84.9</td><td>73.3</td>
<td>76.6</td><td>85.6</td><td>69.8</td>
<td>91.4</td><td>69.1</td><td>67.0</td>
<td>3.6</td>
</tr>
<tr>
<td>UI-TARS-1.5-7B (Seed, 2025)</td>
<td>85.6</td><td>96.2</td><td>81.5</td>
<td>78.6</td><td>92.1</td><td>72.2</td>
<td><b>94.7</b></td><td>74.3</td><td>71.1</td>
<td>1.0</td>
</tr>
<tr>
<td>MiMo-VL-7B-SFT (Xiaomi, 2025)</td>
<td>93.0</td><td>77.9</td><td>65.3</td>
<td>89.7</td><td>85.7</td><td>75.4</td>
<td>89.1</td><td>80.1</td><td>71.0</td>
<td>57.0</td>
</tr>
<tr>
<td>AgentCPM-GUI (Zhang et al., 2025)</td>
<td>84.3</td><td>92.2</td><td>75.1</td>
<td>70.4</td><td>80.7</td><td>56.0</td>
<td>72.3</td><td>54.6</td><td>39.4</td>
<td>2.4</td>
</tr>
<!-- MagicGUI -->
<tr style="background-color:#e8eafc;">
<td>MagicGUI-CPT</td>
<td><b>98.5</b></td><td><b>98.5</b></td><td><b>97.2</b></td>
<td><b>95.5</b></td><td><b>96.3</b></td><td><b>92.9</b></td>
<td>88.5</td><td><b>82.3</b></td><td><b>72.9</b></td>
<td><b>93.2</b></td>
</tr>
<tr style="background-color:#e8eafc;">
<td>MagicGUI-RFT</td>
<td><b>99.7</b></td><td>97.5</td><td><b>97.5</b></td>
<td><b>97.2</b></td><td>95.6</td><td><b>94.0</b></td>
<td>92.1</td><td>80.4</td><td><b>74.1</b></td>
<td>92.1</td>
</tr>
</tbody>
</table>
### Performance comparison on open-source AndroidControl and GUI-Odyssey datasets.
<table>
<thead>
<tr>
<th rowspan="2">Agent Models</th>
<th colspan="2">AC-Low</th>
<th colspan="2">AC-High</th>
<th colspan="2">GUI-Odyssey</th>
</tr>
<tr>
<th>Type</th><th>SR</th>
<th>Type</th><th>SR</th>
<th>Type</th><th>SR</th>
</tr>
</thead>
<tbody>
<!-- Closed-source Models -->
<tr><td colspan="7"><em>Closed-source Models</em></td></tr>
<tr>
<td>GPT-4o (Hurst et al., 2024)</td>
<td>-</td><td>19.5</td>
<td>-</td><td>20.8</td>
<td>-</td><td>20.4</td>
</tr>
<tr>
<td>Gemini 2.0 (Pichai et al., 2024)</td>
<td>-</td><td>28.5</td>
<td>-</td><td>60.2</td>
<td>-</td><td>3.3</td>
</tr>
<tr>
<td>Claude 2.0 (Anthropic, 2024)</td>
<td>-</td><td>28.5</td>
<td>-</td><td>12.5</td>
<td>60.9</td><td>-</td>
</tr>
<!-- Open-source Models -->
<tr><td colspan="7"><em>Open-source Models</em></td></tr>
<tr>
<td>Qwen2-VL-7B (Wang et al., 2024)</td>
<td>55.7</td><td>36.2</td>
<td>45.8</td><td>21.2</td>
<td>58.6</td><td>13.3</td>
</tr>
<tr>
<td>Qwen2.5-VL-7B (Bai et al., 2025)</td>
<td>94.1</td><td>85.0</td>
<td>75.1</td><td>62.9</td>
<td>59.5</td><td>46.3</td>
</tr>
<tr>
<td>Aguvis-7B (Xu et al., 2024)</td>
<td>93.9</td><td>89.4</td>
<td>65.6</td><td>54.2</td>
<td>26.7</td><td>13.5</td>
</tr>
<tr>
<td>OS-Atlas-7B (Wu et al., 2024)</td>
<td>73.0</td><td>67.3</td>
<td>70.4</td><td>56.5</td>
<td>91.8*</td><td>76.8*</td>
</tr>
<tr>
<td>UI-TARS-7B (Qin et al., 2025)</td>
<td>95.2</td><td>91.8</td>
<td>81.6</td><td>74.4</td>
<td>86.1</td><td>67.9</td>
</tr>
<tr>
<td>AgentCPM-GUI (Zhang et al., 2025)</td>
<td>94.4</td><td>90.2</td>
<td>77.7</td><td>69.2</td>
<td><b>90.9</b></td><td><b>75.0</b></td>
</tr>
<!-- MagicGUI -->
<tr style="background-color:#e8eafc;">
<td>MagicGUI-CPT</td>
<td>94.5</td><td>86.7</td>
<td>84.6</td><td>73.1</td>
<td><b>90.4</b></td><td>73.5</td>
</tr>
<tr style="background-color:#e8eafc;">
<td>MagicGUI-RFT</td>
<td><b>97.2</b></td><td><b>93.5</b></td>
<td><b>84.7</b></td><td><b>76.3</b></td>
<td>89.7</td><td><b>74.3</b></td>
</tr>
</tbody>
</table>
## License
* This project is licensed under the [Apache-2.0](./LICENSE) license. The model weights are fully open for academic research, and commercial use licenses can be applied for by contacting magicgui@honor.com. This project uses the pre-trained Qwen2VL-7B-Instruct for initialization, which is also licensed under the Apache- 2.0 License.
## Citation
If **MagicGUI** is useful for your research, please cite:
```bibtex
@misc{tang2025magicguifoundationalmobilegui,
title={MagicGUI: A Foundational Mobile GUI Agent with Scalable Data Pipeline and Reinforcement Fine-tuning},
author={Liujian Tang and Shaokang Dong and Yijia Huang and Minqi Xiang and Hongtao Ruan and Bin Wang and Shuo Li and Zhiheng Xi and Zhihui Cao and Hailiang Pang and Heng Kong and He Yang and Mingxu Chai and Zhilin Gao and Xingyu Liu and Yingnan Fu and Jiaming Liu and Xuanjing Huang and Yu-Gang Jiang and Tao Gui and Qi Zhang and Kang Wang and Yunke Zhang and Yuran Wang},
year={2025},
eprint={2508.03700},
archivePrefix={arXiv},
primaryClass={cs.HC},
url={https://arxiv.org/abs/2508.03700},
}
``` |