---
task_categories:
- question-answering
language:
- en
tags:
- code
pretty_name: PFPdatasets
size_categories:
- 100K<n<1M
license: apache-2.0
---


<p align="center">
  <h1 align="center"><strong>Paper Folding Puzzles: A Benchmark for Evaluating Spatial Reasoning in Multimodal Large Language Models</strong></h1>
</p>

<p align="center">
    🌐 <a href=""><b>Homepage</b></a>&nbsp&nbsp | &nbsp&nbsp
    💻 <a href="https://github.com/hznuer/PFP_bench"><b>GitHub</b></a>&nbsp&nbsp | &nbsp&nbsp
    🤗 <a href="https://huggingface.co/datasets/hznuer/PFP_datasets"><b>Hugging Face</b></a>&nbsp&nbsp 
</p>

# 👋 Introduction
Recent advancements in multimodal large language models (MLLMs) have shown remarkable progress in various reasoning tasks. However, spatial reasoning, particularly in paper folding scenarios, remains a significant challenge due to limitations in understanding geometric transformations and spatial relationships. To address this gap, we present Paper Folding Puzzles (PFP), a comprehensive benchmark designed to evaluate and enhance spatial reasoning capabilities in MLLMs. Our benchmark systematically covers five distinct task types, from basic single-step transformations to complex 3D spatial visualization, providing a rigorous framework for assessing spatial intelligence in AI systems.

# 📌 Highlights

- **We introduce Paper Folding Puzzles (PFP), a multi-dimensional benchmark for spatial reasoning.** It systematically covers five key task types—Single-Step, Inverse, Multi-Step, 3D-Folding, and 2D-Unfolding—addressing different aspects of spatial intelligence.

- **Comprehensive scale with 153,000 carefully curated samples.** The dataset includes 150,000 training samples and 3,000 test samples, ensuring robust evaluation across all task categories.

- **Structured difficulty levels within complex tasks.** The 3D-Folding and 2D-Unfolding categories include easy and hard sub-levels, enabling granular assessment of model capabilities.

- **Standardized format for easy integration.** The dataset uses parquet format with consistent JSON structure, facilitating seamless integration with existing MLLM frameworks.

### Dataset Structure

The structure of Paper Folding Puzzles is shown as follows:

``` 
PFP_dataset/
├── train/
│   ├── Single-Step.parquet
│   ├── Inverse.parquet
│   ├── Multi-Step.parquet
│   ├── 3D-Folding/
│   │   ├── _2DTo3D_N.parquet
│   │   └── _2DTo3D_Y.parquet
│   └── 2D-Unfolding/
│       ├── _3DTo2D_N.parquet
│       └── _3DTo2D_Y.parquet
└── test/
    ├── Single-Step.parquet
    ├── Inverse.parquet
    ├── Multi-Step.parquet
    ├── 3D-Folding.parquet
    └── 2D-Unfolding.parquet
```


### Data Instances
For each instance in the dataset, the following fields are provided:
``` json
{
    "image": "circle_001.png",
    "answer": "D"
}
``` 

### Data Fields
- `image`: a string containing the relative path to the paper folding puzzle image (e.g., "circle_001.png")
- `answer`: a string indicating the correct answer option (A, B, C, or D)

# 🚀 Quick Start

## Loading the Dataset
``` python
from datasets import load_dataset

# Load the entire dataset
dataset = load_dataset("hznuer/PFP_datasets")

# Or load specific splits
train_dataset = load_dataset("hznuer/PFP_datasets", split="train")
test_dataset = load_dataset("hznuer/PFP_datasets", split="test")

# Load specific task types
single_step_data = load_dataset("hznuer/PFP_datasets", "Single-Step")
``` 

## Basic Usage Example
``` python
# Example of processing the dataset
dataset = load_dataset("hznuer/PFP_datasets", split="train")

for sample in dataset:
    image_path = sample["image"]
    correct_answer = sample["answer"]
    # Process your paper folding puzzle here
``` 

# ✒️ Citation

If you find Paper Folding Puzzles helpful, please consider giving this repo a :star: and citing:

``` latex
@inproceedings{zhou2026paperfolding,
  title={Paper Folding Puzzles: A Benchmark for Evaluating Spatial Reasoning in Multimodal Large Language Models},
  author={Zhou, Dibin and Xu, Yantao and Huang, Zongming and Yan, Zengwei and Liu, Wenhao and Miao, Yongwei and Ren, Jianfeng and Liu, Fuchang},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2026}
}
``` 

# 👥 Authors

**Dibin Zhou**, **Yantao Xu**, **Zongming Huang**, **Zengwei Yan**, **Wenhao Liu**, **Yongwei Miao**, **Jianfeng Ren**, **Fuchang Liu**

**Affiliation**: School of Information Science and Technology, Hangzhou Normal University & The Digital Port Technologies Lab, School of Computer Science, University of Nottingham Ningbo China

# 📞 Contact

For questions or issues regarding this dataset:
- Open an issue on the [GitHub repository](https://github.com/hznuer/PFP_bench)
- Contact the authors through the paper correspondence

---

**Paper Folding Puzzles: Advancing spatial reasoning evaluation for multimodal AI systems** 🧠