AgentCPM-GUI: Building Mobile-Use Agents with Reinforcement Fine-Tuning
Paper
•
2506.01391
•
Published
A real-world Chinese Android GUI benchmark designed to evaluate GUI agent models on two complementary capabilities:
See AgentCPM-GUI for more details.
| Aspect | Grounding | Agent |
|---|---|---|
| Objective | GUI widgets grounding / OCR text | Follow natural-language instructions to operate an app |
| Data | 2 × JSONL files (cap.jsonl, ocr.jsonl) + screenshots |
Per-episode JSON + step-level screenshots |
| Actions | n/a | Tap, long-tap, text-input, scroll, etc. (result_action_type, see here) |
CAGUI
├─ CAGUI_agent
│ └─ domestic/
│ └─ <episode_id>/
│ ├─ <episode_id>.json # episode file
│ ├─ <episode_id>_0.jpeg # step-0 screenshot
│ ├─ <episode_id>_1.jpeg
│ └─ ...
└─ CAGUI_grounding
├─ code/
│ ├─ cap.jsonl # function to point & text to point
│ └─ ocr.jsonl # bbox to text
└─ images/
├─ cap/
│ ├─ 0.jpeg
│ └─ ...
└─ ocr/
├─ 0.jpeg
└─ ...
CAGUI_agent/domestic/<episode_id>/<episode_id>.json)
Each file is a list of steps:
| Field | Type | Description |
|---|---|---|
episode_id |
str | Unique id |
episode_length |
int | Total steps |
step_id |
int | Step index (0-based) |
instruction |
str | user goal |
image_path |
str | Relative path to screenshot |
image_width / image_height |
int | Raw resolution |
ui_positions |
str (JSON list) | Normalised [[y, x, h, w], …] UI element boxes |
result_action_type |
int | Action code, see here |
result_action_text |
str | Text typed (if any) |
result_touch_yx / result_lift_yx |
str | Normalised touch coords, [-1,-1] if no touch |
duration |
float | null | Action time (s) |
cap.jsonl, ocr.jsonl)
One JSON object per line:
| Field | Example | Description |
|---|---|---|
task |
"bbox2function" / "bbox2text" |
Sub-task type |
image |
"grounding_eval/dataset/images/0.jpeg" |
Screenshot path |
id |
0 |
Unique int id |
abs_position |
"<x1, y1, x2, y2>" |
Pixel-level bbox |
rel_position |
"<x1, y1, x2, y2>" |
Normalised bbox |
text |
"UI元素是一个菜单按钮…" |
Target description / OCR string |
CAGUI is released under the CC-BY-NC 4.0 license for non-commercial research. Screenshots originate from publicly available Chinese apps and are used under fair-use for research purposes only. Remove them if local regulations require.
This benchmark is used for evaluating AgentCPM-GUI. If our model and benchmark are useful for your research, please cite:
@article{zhang2025agentcpmgui,
title={Agent{CPM}-{GUI}: Building Mobile-Use Agents with Reinforcement Fine-Tuning},
author={Zhong Zhang and Yaxi Lu and Yikun Fu and Yupeng Huo and Shenzhi Yang and Yesai Wu and Han Si and Xin Cong and Haotian Chen and Yankai Lin and Jie Xie and Wei Zhou and Wang Xu and Yuanheng Zhang and Zhou Su and Zhongwu Zhai and Xiaoming Liu and Yudong Mei and Jianming Xu and Hongyan Tian and Chongyi Wang and Chi Chen and Yuan Yao and Zhiyuan Liu and Maosong Sun},
year={2025},
journal={arXiv preprint arXiv:2506.01391},
}