MathGen

Introduction

Modern generative models can solve increasingly difficult mathematical problems in text, but many real use cases require answers to be expressed visually through diagrams, plots, geometric constructions, and symbolic layouts. MathGen asks a direct question: does mathematical competence persist when the answer must be rendered as an image?

We introduce MathGen, a benchmark of 420 curated problems across seven core mathematical domains. The benchmark contains 350 Clean-Scene problems and 70 paired Open-Scene problems. Each task is evaluated with a Script-as-a-Judge protocol: problem-specific executable checks verify numerical, geometric, structural, and logical constraints deterministically.

Clean-Scene Leaderboard

Main results on the 350-problem Clean-Scene set. Accuracy is reported for each domain with 50 problems per domain. Closed-source models lead, but even the strongest systems remain far from reliable mathematical rendering.

BibTeX

@misc{liu2026mathgen,
  title={MathGen: Revealing the Illusion of Mathematical Competence through Text-to-Image Generation},
  author={Liu, Ruiyao and Shen, Hui and Zhang, Ping and Hsieh, Yunta and Zhang, Yifan and Xu, Jing and Han, Qi and Li, Junchen and Lu, Jiawei and Ma, Jianing and Mo, Jiaqi and Chen, Sicheng and Zhang, Zhen and Wan, Zhongwei and Xiong, Jing and Wang, Xin and Liu, Ziyuan and Cao, Hangrui and Wong, Ngai},
  year={2026},
  eprint={2603.27959},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2603.27959}
}