README.md

---
license: mit
task_categories:
- image-text-to-text
tags:
- multimodal
- benchmark
- spatial-reasoning
- indoor-scenes
---

This repository contains the dataset for the paper [SpaCE-10: A Comprehensive Benchmark for Multimodal Large Language Models in Compositional Spatial Intelligence](https://huggingface.co/papers/2506.07966).

<div align="center">
<h1><img src="https://raw.githubusercontent.com/Cuzyoung/SpaCE-10/main/assets/space-10-logo.png" width="8%"> SpaCE-10: A Comprehensive Benchmark for Multimodal Large Language Models in Compositional Spatial Intelligence</h1>
</div>

**GitHub Repository:** [https://github.com/Cuzyoung/SpaCE-10](https://github.com/Cuzyoung/SpaCE-10)

---
# 🧠 What is SpaCE-10?

**SpaCE-10** is a **compositional spatial intelligence benchmark** for evaluating **Multimodal Large Language Models (MLLMs)** in indoor environments. Our contribution as follows:

- 🧬 We define an **Atomic Capability Pool**, proposing 10 **atomic spatial capabilities.**
- 🔗 Based on the composition of different atomic capabilities, we design **8 compositional QA types**.
- 📈 SpaCE-10 benchmark contains 5,000+ QA pairs.
- 🏠 All QA pairs come from 811 indoor scenes (ScanNet++, ScanNet, 3RScan, ARKitScene)
- 🌍 SpaCE-10 spans both 2D and 3D MLLM evaluations and can be seamlessly adapted to MLLMs that accept 3D scan input.

<div align="center">
<br><br>
<img src="https://raw.githubusercontent.com/Cuzyoung/SpaCE-10/main/assets/space-10-teaser.png" width="100%">
<br><br>
</div>

---
# 🔥🔥🔥 News
- [2025/07/12] Adjust some QAs of Space-10 and update RemyxAI models' performance to leader board. 
- [2025/06/11] Scans for 3D MLLMs and our manually collected 3D snapshots will be coming soon.
- [2025/06/10] Evaluation code is released at followings.
- [2025/06/09] We have released the benchmark for 2D MLLMs at [Hugging Face](https://huggingface.co/datasets/Cusyoung/SpaCE-10).
- [2025/06/09] The paper of SpaCE-10 is released at [Arxiv](https://arxiv.org/abs/2506.07966v1)!
---

# Performance Leader Board - Single-Choice
🎉 LLaVA-OneVision-72B achieves the Rank 1 in all tested models.

🎉 GPT-4o achieves the best score in tested Close-Source models.

A large gap still exists between human and models in compositional spatial intelligence.

<div align="center">
<img src="https://raw.githubusercontent.com/Cuzyoung/SpaCE-10/main/assets/Perfomance_Leader_Board.png" width="100%">
<br>
</div>

# Single-Choice vs. Double-Choice
<div align="center">
<img src="https://raw.githubusercontent.com/Cuzyoung/SpaCE-10/main/assets/single-double.png" width="100%">
<br>
</div>

# Capability Score Ranking - Single-Choice
<div align="center">
<img src="https://raw.githubusercontent.com/Cuzyoung/SpaCE-10/main/assets/Capability_Score_Matrix.png" width="100%">
<br>
</div>

# Environment
The evaluation of SpaCE-10 is based on lmms-eval. Thus, we follow the environment settings of lmms-eval.
```bash
git clone https://github.com/Cuzyoung/SpaCE-10.git
cd SpaCE-10
uv venv dev --python=3.10
source dev/bin/activate
uv pip install -e .
```

# Evaluation
Take InternVL2.5-8B as an example:
```bash
cd lmms-eval/run_bash
bash internvl2.5-8b.sh
```
Notably, each time we test a new model, the corresponding environment of this model needs to be installed.

---
# Sample Usage

You can load the dataset using the Hugging Face `datasets` library:

```python
from datasets import load_dataset

dataset = load_dataset("Cusyoung/SpaCE-10")

# To explore the dataset splits:
print(dataset)

# Example of accessing a split (assuming a 'train' split exists):
# train_split = dataset["train"]
# print(train_split[0])
```

---
# Citation
If you use this dataset, please cite the original paper:

```bibtex
@article{gong2025space10,
  title={SpaCE-10: A Comprehensive Benchmark for Multimodal Large Language Models in Compositional Spatial Intelligence},
  author={Ziyang Gong, Wenhao Li, Oliver Ma, Songyuan Li, Jiayi Ji, Xue Yang, Gen Luo, Junchi Yan, Rongrong Ji},
  journal={arXiv preprint arXiv:2506.07966},
  year={2025}
}
```