Graphical User Interface (GUI) agents demonstrate promising potential in assisting human-computer interaction, automating human user's navigation on digital devices. An ideal GUI agent is expected to achieve high accuracy, low latency, and compatibility for different GUI platforms. Recent vision-based approaches have shown promise by leveraging advanced Vision Language Models (VLMs). While they generally meet the requirements of compatibility and low latency, these vision-based GUI agents tend to have low accuracy due to their limitations in element grounding. To address this issue, we propose SpiritSight, a vision-based, end-to-end GUI agent that excels in GUI navigation tasks across various GUI platforms. First, we create a multi-level, large-scale, high-quality GUI dataset called GUI-Lasagne using scalable methods, empowering SpiritSight with robust GUI understanding and grounding capabilities. Second, we introduce the Universal Block Parsing (UBP) method to resolve the ambiguity problem inherited from the dynamic resolution strategy, further enhancing SpiritSight's ability to ground GUI objects. Through these efforts, SpiritSight agent outperforms other advanced methods on diverse GUI benchmarks, demonstrating its superior capability and compatibility in GUI navigation tasks. Models are available at https://huggingface.co/SenseLLM/SpiritSight-Agent-8B.
训练集不变，base 模型换成Qwen2.5-VL
| Agent | Odyssey | AMEX | AndroidCtrl | GUIAct | Multimodal-Mind2Web | ||||
|---|---|---|---|---|---|---|---|---|---|
| High | Low | Multi | Single | Cross-Task | Cross-Website | Cross-Domain | |||
| SeeClick-9.6B | 53.9% | - | 59.1% | 75.0% | - | 72.3% | 25.5% | 16.4% | 20.8% |
| Qwen2-VL-7B | 60.2% | - | 69.7% | 82.6% | - | 82.3% | - | - | - |
| OS-Atlas-4B | 56.4% | - | 67.5% | 80.6% | - | 81.1% | - | - | - |
| OS-Atlas-7B | 62.0% | - | 71.2% | 85.2% | - | 82.7% | - | - | - |
| Aguvis-7B | - | - | 61.5% | 80.5% | - | - | 60.4% | 54.6% | 56.6% |
| Aguvis-72B | - | - | 66.4% | 84.4% | - | - | 64.0% | 56.5% | 58.2% |
| UI-TARS-2B | 83.4% | - | 68.9% | 89.3% | - | - | 56.3% | 50.8% | 52.3% |
| UI-TARS-7B | 87.0% | - | 72.5% | 90.8% | - | - | 67.1% | 61.7% | 60.5% |
| UI-TARS-72B | 87.0% | - | 74.7% | 91.3% | - | - | 68.6% | 63.5% | 62.1% |
| SpiritSight2-3B | 78.6% | 84.3% | 74.4% | 89.3% | 55.2% | 79.1% | 59.5% | 45.2% | 48.6% |
| SpiritSight2-7B | tbd | tbd | tbd | tbd | tbd | tbd | tbd | tbd | tbd |
| SpiritSight2-72B | tbd | tbd | tbd | tbd | tbd | tbd | tbd | tbd | tbd |
| Agent | Odyssey | AMEX | AndroidCtrl | GUIAct | ||
|---|---|---|---|---|---|---|
| High | Low | Multi | Single | |||
| GPT-4o | 20.4% | - | 21.2% | 28.4% | - | 41.8% |
| Previous SOTA* | 74.3% | 70.7% | 64.8% | 80.0% | 45.4% | 74.9% |
| SpiritSight-2B | 72.3% | 74.5% | 64.9% | 86.3% | 45.5% | 76.0% |
| SpiritSight-8B | 75.8% | 80.7% | 68.1% | 87.6% | 49.3% | 78.2% |
| SpiritSight-26B | 75.9% | 82.1% | 69.6% | 88.2% | 52.6% | 78.3% |
* The results are as of Nov. 2024.
| Agent | Model Size | ScreenSpot | Multimodal-Mind2Web | ||||
|---|---|---|---|---|---|---|---|
| Web | Mobile | Desktop | Cross-Task | Cross-Website | Cross-Domain | ||
| GPT4V | - | 5.0% | 7.5% | 4.6% | - | - | - |
| Qwen-VL | 9.6B | 3.0% | 7.2% | 5.4% | - | - | - |
| Fuyu | 8B | 19.2% | 21.2% | 18.3% | 15.6% | 12.2% | 11.7% |
| CogAgent | 18B | 49.5% | 45.5% | 47.1% | 62.3%* | 54.0%* | 59.4%* |
| SeeClick | 9.6B | 44.1% | 65.0% | 51.1% | 25.5% | 16.4% | 20.8% |
| SpiritSight-2B | 2B | 63.6% | 62.5% | 61.8% | 44.9% | 37.8% | 36.9% |
| SpiritSight-8B | 8B | 68.3% | 68.4% | 62.9% | 52.7% | 44.0% | 44.4% |
| SpiritSight-26B | 26B | 74.4% | 73.6% | 69.9% | 54.7% | 48.1% | 49.2% |
* indicate select from candidate elements
@misc{huang2025spiritsightagentadvancedgui,
title={SpiritSight Agent: Advanced GUI Agent with One Look},
author={Zhiyuan Huang and Ziming Cheng and Junting Pan and Zhaohui Hou and Mingjie Zhan},
year={2025},
eprint={2503.03196},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2503.03196},
}