Zeren Chen^1,2*, Xiaoya Lu^2,3*, Zhijie Zheng^1,2, Pengrui Li¹, Lehan He^1,4, Yijin Zhou^2,3,4,
Jing Shao², Bohan Zhuang^5†, Lu Sheng^1†

¹School of Software, Beihang University, ²Shanghai AI Laboratory, ³Shanghai Jiao Tong University,
⁴Shanghai Innovation Institute, ⁵ZIP Lab, Zhejiang University

^*Equal Contribution    ^†Corresponding Author

📖 arXiv | 🤗 HF Paper | 🌐 Homepage

• [2025-12-14] 📝 We release the code of GCA.
• [2025-12-1] 🔥 We release the paper of GCA.

The code requires python>=3.11 and torch>=2.5.1. Please follow the instructions here to install the dependencies and third party repositories.

We evaluate GCA on several challenging spatial reasoning benchmarks, including MMSI-Bench, MindCube, OmniSpatial, SPBench and CVBench. Please follow the instructions here to prepare these evaluation datasets.

For detailed configuration (JSON/Env Vars/CLI) and VLM deployment instructions, please refer to the Usage Documentation.

Run the GCA on supported benchmarks (MMSI, MindCube, CVBench, etc.):

python -m entrypoints.agent --benchmark mmsi --concurrency 16

If you find our work and this codebase helpful, please consider starring this repo 🌟 and cite:

@article{chen2025geometrically,
      title={Geometrically-Constrained Agent for Spatial Reasoning}, 
      author={Zeren, Chen and Xiaoya, Lu and Zhijie, Zheng and Pengrui, Li and Lehan, He and Yijin, Zhou and Jing, Shao and Bohan, Zhuang and Lu, Sheng},
      journal={arXiv preprint arXiv:2511.22659},
      year={2025}
}