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

TL;DR

The paper tackles the efficiency bottleneck of unified text-and-formula recognition in document parsing by introducing UniRec-0.1B, a lightweight 0.1B-parameter model trained on UniRec40M. It proposes two novel techniques—Hierarchical Supervision Training (HST) to capture multi-level layout structure and Semantics-Decoupled Tokenizer (SDT) to remove cross-modality semantic entanglement—enabling accurate recognition at character, word, line, and paragraph levels. Extensive evaluation on UniRec-Bench and OmniDocBench shows UniRec-0.1B achieving competitive or superior accuracy while offering 2–9× speedups compared to large VLM-based or multi-stage systems. The work provides a practical, scalable solution for real-world document parsing and highlights future directions for enhancing fine-grained and cross-domain performance.

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 their mix 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 hierarchies 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 a 2-9$\times$ speedup, validating its effectiveness and efficiency. Codebase and Dataset: https://github.com/Topdu/OpenOCR.

Figures (11)

• Figure 1: Top: Text and formulas take up 97.43% of page regions in quantity in OmniDocBench ouyang2025omnidocbench (the first chart) and consume 87.90% of inference time in one A800 40G GPU by using raw MinerU2.5 niu2025mineru2 (the second chart). When replacing the corresponding MinerU2.5 modules by UniRec-0.1B, the parsing speed becomes 4× faster (the third chart). Bottom: Comparison with advanced document parsing models in our constructed UniRec-Bench. UniRec-0.1B outperforms or on par with existing models in accuracy while only has 3%-33% of their parameters.
• Figure 2: The construction and composition of UniRec40M. Detailed descriptions are provided in Sec. 3.
• Figure 3: The architecture overview of UniRec-0.1B. 'Ġ' represents a space. Detailed descriptions are provided in Sec. 4.
• Figure 4: Recognition result visualizations. Red characters indicate recognition errors, _ indicate missed characters, and * indicate recognition results that are meaningless hallucination. Green indicates correct recognition. More cases are presented in the supplementary.
• Figure 5: An example containing multi-line Chinese plain text mixed with inline formulas. UniRec-0.1B, PaddleOCR-VL, MinerU2.5, and MonkeyOCR-pro correctly recognize the content. Dolphin-1.5 misclassifies the inline formulas as plain text, resulting in incorrect predictions. The remaining three methods extract the content but mistakenly render the inline formulas in display mode. Red characters marks recognition errors, while green text indicates correct recognition.