FigureBench is a benchmark dataset for generating scientific illustrations, released in 2026 by the Text Intelligence Lab at Westlake University. Related research papers include... AutoFigure: Generating and Refining Publication-Ready Scientific IllustrationsIt aims to solve the task of automatically generating high-quality scientific illustrations from long scientific texts, and provides a challenging and diverse testing platform for research on automatic scientific illustration generation. The dataset contains 3,300 samples, consisting of a development set (Dev) and a test set (Test). The development set uses dialogue format annotation for model training and development, while the test set uses a simple pairing format of plain text and images for model evaluation only.

Dataset composition

Development set (Dev): Contains 3,000 samples, including

• 2,900 research papers
• 20 technical blog posts
• 40 review articles
• 40 textbooks Test set (Test): Contains 300 samples, including
• 200 research papers
• 20 technical blog posts
• 40 review articles
• 40 textbooks

Citation

@inproceedings{
zhu2026autofigure,
title={AutoFigure: Generating and Refining Publication-Ready Scientific Illustrations},
author={Minjun Zhu and Zhen Lin and Yixuan Weng and Panzhong Lu and Qiujie Xie and Yifan Wei and Sifan Liu and Qiyao Sun and Yue Zhang},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=5N3z9JQJKq}
}