UniRec-0.1B: Unified Text and Formula Recognition with 0.1B Parameters

Jul 16, 2026

Upload by Anda

Text and formulas are central information units in many documents, and reliable recognition is essential for building robust, generalizable parsing systems. Vision-language models can unify text and formula understanding but are often too large and computationally expensive for broad deployment. UniRec-0.1B introduces a unified 0.1B-parameter model with multi-level outputs, supported by a 40M dataset and training designs addressing structural variability and cross-modal semantic entanglement. Experiments show superior accuracy and 2–9× faster inference on multi-domain Chinese and English benchmarks.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages24

Reading Time1 h 0 min

Introduction

Document parsing and unified recognition

Motivation: efficiency limits of large VLMs

UniRec-0.1B approach

arXiv :2512 .21095v2 [ cs .CV] 11 Jul 2026 UniRec-0.1B: Unified Text and Formula Recognition with 0.1B Parameters Yongkun Du1, Zhineng Chen1,†, Yazhen Xie1, Weikang Bai1, Hao Feng2, Wei Shi2, Yuchen Su1, Can Huang2, Yu-Gang Jiang1,† 1Fudan University, 2 ByteDance Abstract Text and formulas constitute the core informational components of many documents. Accurately and efficiently recognizing both is crucial for developing robust and generalizable document parsing systems. Recently, vision-language models (VLMs) have achieved impressive unified recognition of text and formulas. However, they are large-sized and computationally demanding, restricting their usage in many applications. In this paper, we propose UniRec-0.1B, a unified recognition model with only 0.1B parameters. It is capable of performing text and formula recognition at multiple levels, including characters, words, lines, paragraphs, and documents. To implement this task, we first establish UniRec40M, a large-scale dataset comprises 40 million text, formula and mixed samples, enabling the training of a powerful yet lightweight model. Secondly, we identify two challenges when building such a lightweight but unified expert model. They are: structural variability across levels and semantic entanglement between textual and formulaic content. To tackle these, we introduce a hierarchical supervision training that explicitly guides structural comprehension, and a semantic-decoupled tokenizer that separates text and formula representations. Finally, we develop a comprehensive evaluation benchmark covering Chinese and English documents from multiple domains and with multiple levels. Experimental results on this and public benchmarks demonstrate that UniRec-0.1B outperforms both general-purpose VLMs and leading document parsing expert models, while achieving 2-9× speedup, validating its effectiveness and efficiency. Correspondence: [zhinchen@fudan.edu.cn](zhinchen@fudan.edu.cn) , [ygj@fudan.edu.cn](ygj@fudan.edu.cn) Codebase and Dataset: [https://github.com/Topdu/OpenOCR](https://github.com/Topdu/OpenOCR) 1 Introduction Document parsing [81] serves as a key component in real-world applications like document understanding, digital education, and information retrieval, etc. As the core tasks of document parsing, text and formula recognition have traditionally been addressed as distinct tasks, each with extensive research [3, 17–19, 22, 37, 45, 55–57, 67, 77, 80] and significant progress over the past decades. Recent advances in vision-language models (VLMs) [2, 10, 26, 29, 62, 74, 75, 87] have enabled end-to-end document parsing [5, 7, 9, 11, 15, 35, 38, 42, 47, 49, 50, 54, 71, 72], where a single large model unifies text, formula, table, and chart understanding within one framework. While conceptually elegant, such unified paradigms typically rely on billion-scale parameters, resulting in substantial computational cost and inference latency. On the other hand, a closer examination of real-world document distributions reveals a critical †Corresponding authors. Quantity Proportion Text Formula Mixed Table Text (1-Edit)% Formula (1-Edit)% Table (TEDS) Inference Speed Performence 100 90 80 70 60 50 Peak | | | | | | | | | | | | | | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | | | | | | | | | | | | | | | | | | | | | | | | | | | 1.88 | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | 0.78 | | | | | | | | | | | <br> 0.09 | 0.37<br>.15 | 0.49 | | | | | | | | | | | 3 2 1 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 Inference Speed (second) Number of Parameters (B) (a) (b) 1-Edit 1.0 0.9 0.9 0.8 0.8 0.7 | | | | | --- | --- | --- | | * | | | | | | | | | | | | | | | 0 1 2 3 Number of Parameters (B) (c) * UniRec-0 .1B Dolphin-0 .3B

FAQs

What problem does UniRec-0.1B address?

It targets unified recognition of both text and mathematical formulas in documents, aiming for strong accuracy while reducing model size and computational cost compared with large vision-language models.

How does UniRec-0.1B enable multi-level text and formula recognition?

The method supports recognition across characters, words, lines, paragraphs, and documents, using a large-scale dataset (UniRec40M) and training strategies tailored for structure and cross-modal relations.

Which challenges are highlighted for building a lightweight unified expert model?

The document identifies structural variability across levels and semantic entanglement between textual and formulaic content as two key challenges.

Jul 16, 2026

Upload by Anda