# MixTeX: Data-Efficient LaTeX OCR via Synthetic Pretraining and Limited Fine-Tuning

arXiv: https://arxiv.org/abs/2406.17148 · HTML: https://mlsanity.com/paper/2406.17148/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Builds on 7 of its 15 references that are in the map.
- First posted on arXiv 24 Jun 2024.

## Builds on (most cited first, 7 of 7)

- [Swin Transformer: Hierarchical Vision Transformer using Shifted Windows](https://mlsanity.com/paper/2103.14030/) (2021)
- [Evaluation: from precision, recall and F-measure to ROC, informedness, markedness and correlation](https://mlsanity.com/paper/2010.16061/) (2020)
- [OCR-free Document Understanding Transformer](https://mlsanity.com/paper/2111.15664/) (2021)
- [Nougat: Neural Optical Understanding for Academic Documents](https://mlsanity.com/paper/2308.13418/) (2023)
- [Synthetic Data for Deep Learning](https://mlsanity.com/paper/1909.11512/) (2019)
- [TableBank: A Benchmark Dataset for Table Detection and Recognition](https://mlsanity.com/paper/1903.01949/) (2019)
- [Image-to-Markup Generation with Coarse-to-Fine Attention](https://mlsanity.com/paper/1609.04938/) (2016)

## Built on by (most cited first, 1 of 1)

- [Beyond Human Annotation: Recent Advances in Data Generation Methods for Document Intelligence](https://mlsanity.com/paper/2601.12318/) (2026)

## Abstract (arXiv)

LaTeX OCR converts scientific document images into editable LaTeX code. Existing systems rely on large paired datasets, which are costly to collect and limited for low-resource languages. This paper presents MIXTEX, a data-efficient system using synthetic pretraining without real LaTeX sources. Unlike Nougat that depends on arXiv datasets, we generate training data by randomly pairing grammatical Wikipedia text with LaTeX formulas, requiring only syntactic correctness. This eliminates dependency on real document collections, enables scalable data generation (120M tokens), and supports low-resource languages. Following synthetic pretraining, adaptation requires only 400 real samples. Evaluation on a 977-sample benchmark with printed and handwritten English and Chinese shows that this two-stage strategy outperforms methods trained on large real datasets while requiring less human effort and computation. Data, code, and models are publicly available.