UI-R1: Enhancing Efficient Action Prediction of GUI Agents by Reinforcement Learning

🔥 Overview

We propose UI-R1, the first framework to explore how rule-based RL can enhance the reasoning capabilities of multimodal large language models (MLLMs) for GUI action prediction tasks.

Experimental results demonstrate that our proposed UI-R1-3B achieves significant improvements over the base model (i.e. Qwen2.5-VL-3B) on both in-domain (ID) and out-of-domain (OOD) tasks, with average accuracy gains of 22.1% on ScreenSpot, 6.0% on ScreenSpot-Pro, and 12.7% on AndroidControl. Furthermore, UI-R1-3B delivers competitive performance compared to larger models (e.g., OS-Atlas-7B) trained via supervised fine-tuning (SFT) on 76K samples.

Grounding Leaderboard: UI-I2E-Bench

Model	ScreenSpot	UI-I2E-Bench Avg	ScreenSpot-Pro	Average
UI-TARS-1.5-7B	88.1	73.2	42.2	67.8
Uground-V1-72B	89.7	76.3	34.3	66.8
UI-TARS-72B	88.4	73.7	38.1	66.7
UI-R1-E-3B	89.2	69.1	33.5	63.9
Uground-V1-7B	87.1	70.3	31.1	62.8
InfiGUI-R1	87.5	69.7	29.6	62.3
UI-TARS-7B	89.5	61.4	35.7	62.2
Qwen2.5-VL-72B	87.1	51.4	43.6	60.7
UI-I2E-VLM-7B	82.5	69.5	23.6	58.5
UI-TARS-2B	82.3	62	27.7	57.3
Qwen2.5-VL-7B	84.7	53.8	29	55.8
OmniParser-V2	72	54.8	39.6	55.5
Uground-V1-2B	78.8	57.4	26.6	54.3
OS-Atlas-7B	82.5	58.6	18.9	53.3
UI-R1-3B	83.3	58.5	17.8	53.2
UGround-7B	74.1	54.2	16.5	48.3
UI-I2E-VLM-4B	70.4	53.4	12.2	45.3
OmniParser	73.9	53.1	8.3	45.1
ShowUI-2B	76.8	41.5	7.7	42
Qwen2.5-VL-3B	55.5	41.7	23.9	41.3
Aguvis-7B	84.4	53.2	22.9	40.4
OS-Atlas-4B	70.1	44.3	3.7	39.4
Qwen2-VL-7B	42.6	48.7	1.6	31
Seeclick	55.8	26.4	1.1	27.8
InternVL2-4B	4.2	0.9	0.3	1.8

🔥Insight 1 : Fast Grounding

Thinking is not needed for GUI grounding.

Inspired by concurrent works studying efficient LRM, we realize efficient reasoning by RFT training. UI-R1-3B-E's training consists of two steps:

1. DAST (Difficulty-Adaptive Slow-Thinking): Add difficulty-adaptive length reward to make reasoning from slow to fast.
2. Nothinking: Not output reasoning process.

Note: UI-R1-3B (v2) and UI-R1-3B-E both train on larger dataset (2K grounding data in GUI-R1-3K) compared to UI-R1-3B (v1).

Benchmark 1: ScreenSpotV2

ScreenSpotV2	inference mode	Mobile-T	Mobile-I	Desktop-T	Desktop-I	Web-T	Web-I	Avg↑ / Len↓
OS-ATLAS-7B	w/o thinking	95.2	75.8	90.7	63.6	90.6	77.3	84.1 /
UI-TARS-7B	w/o thinking	95.2	79.1	90.7	68.6	90.6	78.3	84.7 /
UI-R1-3B (v1)	w/ thinking	96.2	84.3	92.3	63.6	89.2	75.4	85.4 / 67
GUI-R1-3B	w/ thinking	97.6	78.2	94.3	64.3	91.0	72.4	85.0 / 80
UI-R1-3B (v2)	w/ thinking	97.6	79.6	92.3	67.9	88.9	77.8	85.8 / 60
UI-R1-E-3B	w/o thinking	98.2	83.9	94.8	75.0	93.2	83.7	89.5 / 28

Benchmark 2: ScreenSpot-Pro

ScreenSpot-Pro	inference mode	Average Length↓	Average Accuracy↑
UGround-7B	w/o thinking	-	16.5
OS-ATLAS-7B	w/o thinking	-	18.9
UI-R1-3B (v1)	w/ thinking	102	17.8
GUI-R1-3B	w/ thinking	114	26.6
UI-R1-3B (v2)	w/ thinking	129	29.8
UI-R1-E-3B	w/o thinking	28	33.5

Analysis

1. Our UI-R1-3B-E achieves SOTA with least answer tokens in 3B/7B Open-source methods, demonstrating GUI grounding needs no reasoning.

Todo

• Performance on 7B may be opposite.
• Performance on Planning may be opposite. The author predicts that Fast Grounding, Slow Planning.
• The checkpoints of UI-R1-3B-E will be released soon.
• The updated paper will come soon.
• The efficient training code will come soon. (in src/script/train_e.sh)

Setup

conda create -n ui-r1 python=3.10
conda activate ui-r1
bash setup.sh

Data

Our training mobile data is a subset from AndroidControl and ScreenSpot.

You can also prepare your training or inference data like:

images/:
	image1.png
	image2.png

test.json:
[
	{
	"img_filename": "image1.png",
        "bbox": [
            825,
            72,
            1673,
            149
        ],
        "instruction": "search bar"
     },
     {
	"img_filename": "image2.png",
        "bbox": [
            123,
            732,
            334,
            812
        ],
        "instruction": "check weather"
     }
]

where bbox : [x1,y1,x2,y2] is the coordinate of the left top and the right bottom of the ground truth bbox

Inference

We provide an example here

cd evaluation/
bash test.sh

Please fill the MODEL_PATH, IMG_PATH, TEST_JSON with your real checkpoint path and data path.

Training

cd src/script/
bash train.sh
# efficient training
bash train_e.sh

🗞️ News

• 2025-11-08: Our paper was accepted by AAAI-2026.
• 2025-05-14: We update the paper with UI-R1-E-3B.
• 2025-05-12: We release the checkpoints of the UI-R1-E-3B model.
• 2025-05-12: We fix the bug of scales when batch_size > 1.
• 2025-05-11: We release the efficient training code of the UI-R1-E-3B model.
• 2025-04-02: We release the datasets of the UI-R1-3B (v1) model.
• 2025-03-30: We release the checkpoints of the UI-R1-3B (v1) model.
• 2025-03-30: We release the UI-R1 repository.
• 2025-03-27: We release our paper.

⭐️ Citation

If you find this project useful, welcome to cite us.

@article{lu2025ui,
  title={UI-R1: Enhancing Action Prediction of GUI Agents by Reinforcement Learning},
  author={Lu, Zhengxi and Chai, Yuxiang and Guo, Yaxuan and Yin, Xi and Liu, Liang and Wang, Hao and Xiong, Guanjing and Li, Hongsheng},
  journal={arXiv preprint arXiv:2503.21620},
  year={2025}
}

🤝 Acknowledgements

We sincerely thank projects R1-V, Open-R1, and Open-r1-multimodal, VLM-R1 for providing their open-source resources.