lidingm/SpatialEvo-160K

---
license: apache-2.0
task_categories:
- visual-question-answering
language:
- en
size_categories:
- 100K<n<1M
---
# SpatialEvo: Self-Evolving Spatial Intelligence via Deterministic Geometric Environments

<a href="https://arxiv.org/abs/2604.14144" target="_blank">
<img alt="Paper" src="https://img.shields.io/badge/arXiv-SpatialEvo-red?logo=arxiv" height="20" />
</a>
<a href="https://github.com/ZJU-REAL/SpatialEvo" target="_blank">
<img alt="Code" src="https://img.shields.io/badge/Code-SpatialEvo-white?logo=github" height="20" />
</a>
<a href="https://huggingface.co/lidingm/SpatialEvo-3B" target="_blank">
<img alt="Model" src="https://img.shields.io/badge/%F0%9F%A4%97%20_Model-SpatialEvo_3B-ffc107?color=ffc107&logoColor=white" height="20" />
</a>
<a href="https://huggingface.co/lidingm/SpatialEvo-7B" target="_blank">
<img alt="Model" src="https://img.shields.io/badge/%F0%9F%A4%97%20_Model-SpatialEvo_7B-ffc107?color=ffc107&logoColor=white" height="20" />
</a>

## SpatialEvo-160K

### Dataset Description

**SpatialEvo-160K** is an offline spatial reasoning QA dataset generated by the **Deterministic Geometric Environment (DGE)** from [SpatialEvo: Self-Evolving Spatial Intelligence via Deterministic Geometric Environments](https://arxiv.org/abs/2604.14144). This dataset is not used in the SpatialEvo training pipeline reported in the paper; it is released separately to demonstrate the data generation capability and correctness of the DGE. All QA pairs are programmatically derived from 3D scene assets with exact ground truth, containing zero annotation noise.

The dataset covers 16 spatial reasoning task categories across scene-level, single-image, and image-pair settings, constructed from **ScanNet**, **ScanNet++**, and **ARKitScenes** as data sources.

### Validation

To verify the quality of DGE-generated data, we fine-tune Qwen2.5-VL-7B-Instruct on SpatialEvo-160K via supervised fine-tuning using the same training hyperparameters as reported in the paper, and evaluate on multiple spatial reasoning benchmarks. Results are shown below:

| Benchmark | Qwen2.5-VL-7B (Base) | SpatialEvo-160K SFT |
|-----------|----------------------|----------------------|
| VSI-Bench | 31.08% | **51.67%** |
| EmbSpatial | 63.57% | **67.23%** |
| SpatialViz | 26.95% | **29.07%** |
| ViewSpatial | 36.43% | **44.63%** |
| V-STAR | 78.53% | **85.86%** |
| **AVG** | 47.31% | **55.69%** |

These results confirm that the DGE produces high-quality, noise-free spatial QA data that leads to consistent performance gains across diverse spatial reasoning benchmarks.

### Data Sources

SpatialEvo-160K is constructed from the following 3D scene datasets. Images must be downloaded from their respective official sources:

- **ScanNet**: [http://www.scan-net.org](http://www.scan-net.org)
- **ScanNet++**: [https://kaldir.vc.in.tum.de/scannetpp](https://kaldir.vc.in.tum.de/scannetpp)
- **ARKitScenes**: [https://github.com/apple/ARKitScenes](https://github.com/apple/ARKitScenes)

After downloading, please follow the dataset processing pipelines provided in our [GitHub repository](https://github.com/ZJU-REAL/SpatialEvo) to convert the raw data into the unified scene format required by the DGE. The full QA generation pipeline is also available in the repository.

### Citation

If you find SpatialEvo-160K useful, please consider citing our work:

```bibtex
@misc{li2026spatialevoselfevolvingspatialintelligence,
title={SpatialEvo: Self-Evolving Spatial Intelligence via Deterministic Geometric Environments},
author={Dinging Li and Yingxiu Zhao and Xinrui Cheng and Kangheng Lin and Hongbo Peng and Hongxing Li and Zixuan Wang and Yuhong Dai and Haodong Li and Jia Wang and Yukang Shi and Liang Zhao and Jianjian Sun and Zheng Ge and Xiangyu Zhang and Weiming Lu and Jun Xiao and Yueting Zhuang and Yongliang Shen},
year={2026},
eprint={2604.14144},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2604.14144},
}
```

### Related Resources

- 📄 [Paper](https://arxiv.org/abs/2604.14144)
- 💻 [GitHub Repository](https://github.com/ZJU-REAL/SpatialEvo)
- 🤗 [SpatialEvo-3B](https://huggingface.co/lidingm/SpatialEvo-3B)
- 🤗 [SpatialEvo-7B](https://huggingface.co/lidingm/SpatialEvo-7B)