Yuqi-Zhou/GUI-G1-3B-0.1K

This repository contains the model presented in GUI-G1: Understanding r1-zero-like training for visual grounding in gui agents.

Project page: https://github.com/Yuqi-Zhou/GUI-G1

Training Details:

• Training Dataset: We used the UI-R1-3B-Train to train our Qwen2.5-VL-3B-Instruct, which contains 101 samples with grounding annotations.
• Other: During training, the vision encoder was frozen. We used a learning rate of 1e-6, a sampling temperature of 0.9, and generated 8 outputs per prompt. Training was conducted on 4 L20 (48G) GPUs with 4 samples per GPU, beta set to 0, and a gradient accumulation step of 1.

Benchmark 1: ScreenSpotV2

ScreenSpotV2	inference mode	Mobile-T	Mobile-I	Desktop-T	Desktop-I	Web-T	Web-I	Avg
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
GUI-R1-3B	w/ thinking	97.6	78.2	94.3	64.3	91.0	72.4	85.0
UI-R1-3B (v2)	w/ thinking	97.6	79.6	92.3	67.9	88.9	77.8	85.8
UI-R1-E-3B	w/o thinking	98.2	83.9	94.8	75.0	93.2	83.7	89.5
GUI-G1-3B-0.1K	w/o thinking	98.3	93.36	92.8	80.0	88.5	79.3	89.8

Benchmark 2: ScreenSpot-Pro

ScreenSpot-Pro	inference mode	Average Accuracy↑
UGround-7B	w/o thinking	16.5
OS-ATLAS-7B	w/o thinking	18.9
UI-R1-3B (v1)	w/ thinking	17.8
GUI-R1-3B	w/ thinking	26.6
UI-R1-3B (v2)	w/ thinking	29.8
UI-R1-E-3B	w/o thinking	33.5
GUI-G1-3B-0.1K	w/o thinking	43.9

Evaluation Code for GUI Grounding

Here we show a code snippet to show you how to use the chat model with transformers and qwen_vl_utils:

from transformers import Qwen2_5_VLForConditionalGeneration, AutoTokenizer, AutoProcessor
from qwen_vl_utils import process_vision_info
# default: Load the model on the available device(s)
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "Yuqi-Zhou/GUI-G1-3B-0.1K", torch_dtype="auto", device_map="auto"
)
# We recommend enabling flash_attention_2 for better acceleration and memory saving, especially in multi-image and video scenarios.
# model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
#     "Yuqi-Zhou/GUI-G1-3B-0.1K",
#     torch_dtype=torch.bfloat16,
#     attn_implementation="flash_attention_2",
#     device_map="auto",
# )

# default processer
processor = AutoProcessor.from_pretrained("Yuqi-Zhou/GUI-G1-3B-0.1K")

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg",
            },
            {"type": "text", "text": "Grounding instruction is:{Question}. Help to locate and output its bbox coordinates using JSON format."},
        ],
    }
]
# Preparation for inference
text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
)
inputs = inputs.to("cuda")
# Inference: Generation of the output
generated_ids = model.generate(**inputs, max_new_tokens=128, use_cache=True)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text)

Citation

If you find our work helpful, feel free to give us a cite.

@article{zhou2025gui,
  title={GUI-G1: Understanding r1-zero-like training for visual grounding in gui agents},
  author={Zhou, Yuqi and Dai, Sunhao and Wang, Shuai and Zhou, Kaiwen and Jia, Qinglin and Xu, Jun},
  journal={arXiv preprint arXiv:2505.15810},
  year={2025}
}