PaperlayerView exact revision on arXiv

Source-linked AI summary

MinerU2.5: A Decoupled Vision-Language Model for Efficient High-Resolution Document Parsing

Junbo Niu, Zheng Liu, Zhuangcheng Gu, Bin Wang, Linke Ouyang, Zhiyuan Zhao, Tao Chu, Tianyao He, Fan Wu, Qintong Zhang, Zhenjiang Jin, Guang Liang, Rui Zhang, Wenzheng Zhang, Yuan Qu, Zhifei Ren, Yuefeng Sun, Yuanhong Zheng, Dongsheng Ma, Zirui Tang, Boyu Niu, Ziyang Miao, Hejun Dong, Siyi Qian, Junyuan Zhang, Jingzhou Chen, Fangdong Wang, Xiaomeng Zhao, Liqun Wei, Wei Li, Shasha Wang, Ruiliang Xu, Yuanyuan Cao, Lu Chen, Qianqian Wu, Huaiyu Gu, Lindong Lu, Keming Wang, Dechen Lin, Guanlin Shen, Xuanhe Zhou, Linfeng Zhang, Yuhang Zang, Xiaoyi Dong, Jiaqi Wang, Bo Zhang, Lei Bai, Pei Chu, Weijia Li, Jiang Wu, Lijun Wu, Zhenxiang Li, Guangyu Wang, Zhongying Tu, Chao Xu, Kai Chen, Yu Qiao, Bowen Zhou, Dahua Lin, Wentao Zhang, Conghui He

arXiv:2509.22186v2cs.CVcs.CL

TL;DR

Document images combine high resolution, dense content, and complex layouts, challenging models to capture fine detail efficiently. MinerU2.5 separates global layout analysis from local recognition in a two-stage approach and achieves state-of-the-art performance across public benchmarks.

  • Problem

    Document images’ high resolution, dense content, and complex layouts challenge models to process native-resolution detail while remaining parameter-efficient and robust.

    Source: §1 · p. 4

  • Method

    MinerU2.5 uses a coarse-to-fine two-stage strategy, analyzing global layout on downsampled images before recognizing content in guided crops from the original high-resolution input.

    Source: §1 · p. 5

  • Results

    Across public benchmarks, MinerU2.5 achieves state-of-the-art performance, outperforming MonkeyOCR-pro-3B [17] by 1.82 points on OmniDocBench overall.

    Source: §1 · p. 5, §5.1.2 · p. 21

  • Takeaways & Limitations

    MinerU2.5 is positioned as a tool for converting unstructured document collections into structured data for pretraining-corpus curation.

    Source: §6 · p. 25

    The paper notes that modular systems can have interdependencies that make their use, maintenance, and updates cumbersome and less efficient.

    Source: §2.1 · p. 6

Abstract

from arXiv · show

We introduce MinerU2.5, a 1.2B-parameter document parsing vision-language model that achieves state-of-the-art recognition accuracy while maintaining exceptional computational efficiency. Our approach employs a coarse-to-fine, two-stage parsing strategy that decouples global layout analysis from local content recognition. In the first stage, the model performs efficient layout analysis on downsampled images to identify structural elements, circumventing the computational overhead of processing high-resolution inputs. In the second stage, guided by the global layout, it performs targeted content recognition on native-resolution crops extracted from the original image, preserving fine-grained details in dense text, complex formulas, and tables. To support this strategy, we developed a comprehensive data engine that generates diverse, large-scale training corpora for both pretraining and fine-tuning. Ultimately, MinerU2.5 demonstrates strong document parsing ability, achieving state-of-the-art performance on multiple benchmarks, surpassing both general-purpose and domain-specific models across various recognition tasks, while maintaining significantly lower computational overhead.

1 Introduction

MinerU2.5 uses a lightweight, coarse-to-fine architecture to separate global layout analysis from local, native-resolution recognition, supported by a data engine for complex documents. The introduction reports state-of-the-art performance across public benchmarks and highlights improvements in layout, formula, and table parsing.

1 Introduction

  • Document parsing must handle high-resolution, dense, structurally complex pages while balancing recognition, layout analysis, model efficiency, and robust inference.The introduction contrasts modular pipelines, which can propagate errors, with VLM approaches, which face long-document hallucination and high-resolution efficiency bottlenecks compounded by token redundancy.Source: §1 · pp. 4–5
  • Its two-stage approach analyzes downsampled pages globally, then recognizes layout-guided crops from the original high-resolution image.The paper says this decoupling reduces computational cost by an order of magnitude, improves interpretability, mitigates hallucination, and allows independent optimization of the stages; the model has 1.2B parameters.Source: §1 · p. 5
  • A closed-loop data engine generates large-scale, high-quality document corpora to address limited diversity, sample imbalance, and inconsistent annotations.The corpora support robustness across varied layouts, document types, and complex elements.Source: §1 · p. 5
  • MinerU2.5 achieves state-of-the-art performance across a wide range of public benchmarks.The introduction also reports a qualitative leap over the previous MinerU2 version in practical application and user experience.Source: §1 · p. 5
  • Its granular layout analysis preserves headers, footers, and page numbers and standardizes labels for lists, references, and code blocks.Source: §1 · p. 5
  • It parses complex, lengthy formulas and recognizes equations that mix Chinese and English.Source: §1 · p. 5
  • It robustly parses rotated, borderless, and partially bordered tables.Source: §1 · p. 5

2 Related Work

Related work spans modular OCR pipelines, general-purpose vision-language models, and domain-specific end-to-end or multi-stage parsers. Each paradigm offers advantages but faces limitations in robustness, efficiency, or deployment complexity.

2.1 Traditional Pipelines

  • Traditional OCR pipelines [8] [24] [46] sequence layout, recognition, and reading-order modules, enabling specialized refinement but risking propagated errors and cumbersome interdependencies.Marker combines Surya OCR with layout analysis and reading-order prediction, while MinerU [46] orchestrates specialized models. Such pipelines can struggle with complex layouts, including multi-column text and cross-page structures.Source: §2.1 · pp. 5–6

2.2 General-Purpose Vision Language Models

  • General-purpose VLMs [1] [3] [7] [63] show OCR promise, but cost, scale, and dense-text hallucinations limit deployment.Gemini2.5 Pro [7] surpasses traditional pipeline models like MinerU [46] in text parsing and approaches UniMERNet [45] in formula recognition; Qwen2.5-VL-72B [3] achieves the best results among open-source models. Proprietary models can be expensive and slow, while open-source models require massive parameter scales for optimal performance.Source: §2.2 · p. 6

2.3 Domain-Specific Vision Language Models

  • Domain-specific end-to-end models [4] [6] [15] [23] [35] [52] unify parsing but face scalability, hallucination, and high-resolution efficiency limits, while multi-stage VLMs [11] [17] decouple layout and recognition.GOT [52] pioneered unified recognition across text, formulas, tables, and charts; end-to-end methods also face layout-accuracy challenges and O(N^2) native-resolution complexity. Dolphin [11] uses fixed-resolution crops that can distort extreme-aspect-ratio regions, whereas MonkeyOCR [17] uses native-resolution recognition in its second stage but requires multiple specialized models; a unified model with native-resolution parsing is presented as a promising direction.Source: §2.3 · p. 6

3 MinerU2.5

MinerU2.5 combines a compact vision-language architecture with two-stage parsing: global layout analysis guides native-resolution recognition of local regions. Its training recipe progresses from modality alignment through document-parsing pre-training and targeted fine-tuning.

3.1 Model Architecture

  • MinerU2.5 uses a 675M-parameter NaViT vision encoder, a Qwen2-Instruct 0.5B decoder, and pixel-unshuffle patch merging.The encoder supports dynamic resolutions and aspect ratios; M-RoPE replaces 1D-RoPE for cropped images, while patch merging aggregates adjacent 2 × 2 vision tokens.Source: §3.1 · p. 7

3.2 Two-Stage Parsing Strategy

  • The two-stage strategy separates global layout analysis from local recognition to reduce redundancy while preserving layout context and fine details.Stage I resizes pages to 1036 × 1036 pixels; Stage II recognizes native-resolution crops capped at 2048 × 28 × 28 pixels.Source: §3.2 · pp. 7–8

3.3 Training Recipe

  • Training begins from Qwen2-VL-2B-Instruct vision and Qwen2-Instruct-0.5B language-model initializations, then proceeds through three stages.The three-stage procedure comprises modality alignment, document-parsing pre-training, and document-parsing fine-tuning.Source: §3.3 · p. 8

3.3.1 Stage 0-Modality Alignment

  • Stage 0 aligns visual features with the language model, then tunes all parameters on visual-instruction and OCR data; skipping it raises losses and reduces performance.Language-image alignment trains only the patch-merger MLP with the vision encoder and language model frozen; visual instruction tuning unfreezes all parameters.Source: §3.3.1 · pp. 8–9

3.3.2 Stage 1-Document Parsing Pre-training

  • Stage 1 trains all parameters on layout detection and text, formula, and table recognition, producing a baseline and a hard-sample miner.It runs for 2 epochs of 6.9M samples each: 2.3M layout, 2.4M text, 1.1M formula, and 1.1M table samples per epoch.Source: §3.3.2 · p. 9

3.3.3 Stage 2-Document Parsing Fine-tuning

  • Stage 2 preserves diverse pre-training examples and adds manually annotated difficult cases, improving challenging scenarios while retaining established parsing abilities.Fine-tuning runs for 3 epochs of 630K samples each, including 43K layout, 300K text, 147K formula, and 140K table samples per epoch.Source: §3.3.3 · p. 10

3.3.4 Data Augmentation Strategies

  • Targeted augmentations in Stages 1 and 2 simulate document interference, with spatial transformations excluded from layout-analysis samples.Augmentation parameters and probabilities are tailored by element type to balance performance and robustness.Source: §3.3.4 · p. 10

3.4 Model Deployment

  • 2.12 pages/s end-to-end throughput and 2337.25 tokens/s generation speed; MinerU2.5 achieves 4× MonkeyOCR-Pro-3B and 7× dots.ocr page throughput.Without deployment optimizations, it reaches 0.95 pages/s and 1045.14 tokens/s, already surpassing the compared models under default configurations.Source: Table 3 · p. 11, §3.4 · p. 11

4 Data Engine

MinerU2.5’s Data Engine builds training data through curation, annotation refinement, and hard-case mining, while also reformulating layout, formula, and table recognition tasks with specialized representations and evaluation methods.

4.1.1 Data Curation

  • The pipeline balances a diverse Chinese-English corpus across layouts, document types, element classes, and languages to counter the raw pool’s long-tail distribution.It uses page-level image clustering, metadata-based stratified sampling, element detection, and language filtering.Source: §4.1.1 · p. 12

4.1.2 Pre-training Dataset Preparation

  • Pre-training annotations begin with the MinerU2 pipeline and are refined by specialized text, formula, and table models into image-annotation pairs for four core tasks.Qwen2.5-VL-72B-Instruct verifies text, UniMERNet supplies higher-fidelity formulas, and an in-house parser regenerates table structures.Source: §4.1.2 · pp. 12–13

4.1.3 Fine-tuning Dataset Construction

  • Fine-tuning combines human-reviewed hard cases selected by IMIC with a smaller random sample of regular examples to target complex real-world parsing.An AI-assisted workflow pre-annotates selected cases, using foundation models such as Gemini-2.5-Pro for complex tables, before expert review and correction.Source: §4.1.3 · p. 13

4.2 Task Reformulation and Enhancement

  • The task reformulation introduces more robust standards, task paradigms, metrics, and representations for layout analysis, formula recognition, and table recognition.Source: §4.2 · p. 13

4.2.1 Layout Analysis

  • The layout tagging system uses hierarchical categories to cover often-omitted headers, footers, and page numbers while distinguishing semantic text types and figure subcategories.It separately tags code, algorithms, references, and lists, and distinguishes images, charts, and chemical structures and their captions.Source: §4.2.1 · p. 13
  • Layout analysis predicts each element’s position, class, rotation angle, and reading order in one pass, addressing rotated elements and reducing pipeline coupling.Source: §4.2.1 · pp. 13–14
  • PageIoU evaluates page-level spatial coverage to address mismatches between fixed-threshold IoU scores and visual judgments of document layouts.It scores the visually superior line-by-line prediction 0.97 versus 0.78 for the poorer coarse prediction.Source: §4.2.1 · pp. 14–15

4.2.2 Formula Recognition

  • The ADR framework recognizes compound formulas by decomposing them into ordered atomic lines, recognizing each as LaTeX, and recombining outputs using layout positions.MinerU2.5 performs both layout analysis and recognition, turning a difficult whole-formula task into simpler component tasks while preserving structure.Source: §4.2.2 · pp. 15–16

5 Evaluation

MinerU2.5 is evaluated on full-document and element-specific parsing benchmarks, with the paper reporting leading results across most metrics. The evaluation covers layout analysis, table recognition, and formula recognition alongside full-document parsing.

5.1.2 Evaluation Results

  • On Ocean-OCR, MinerU2.5 achieves English edit distance of 0.033 and F1-score of 0.945, and Chinese F1-score of 0.965 and Precision of 0.966.Its English BLEU and METEOR scores are 0.909 and 0.950; its Chinese BLEU and METEOR scores are 0.817 and 0.887.Source: Table 7 · p. 20, §5.1.2 · p. 21
  • MinerU2.5 scores 75.2 overall on olmOCR-bench, 1.6 points above dots.ocr’s 73.6.It scores 76.6 on arXiv Math, 4.4 points above Qwen2.5-VL-72B [3], 54.6 on Old Scans Math, and 83.5 on Long Tiny Text, above MonkeyOCR-pro-1.2B [17] at 81.7. AR and OSM formula scores use CDM ExpRate, which compares rendered outputs for exact matches.Source: Table 8 · p. 20, §5.1.1 · p. 21, §5.1.2 · p. 22

5.1.1 Evaluation Details and Metrics

  • The updated OmniDocBench evaluation contains 1,355 pages and uses hybrid matching for its overall score.Notes and Newspapers are evaluated at 200 DPI rather than 72 DPI, and 374 pages were added to balance Chinese-English content and expand formula coverage. For olmOCR-bench, AR and OSM formula scores use CDM ExpRate, comparing rendered outputs and assigning 1 for exact matches and 0 otherwise.Source: §5.1.1 · p. 21

5.1.2 Evaluation Results

  • MinerU2.5 scores 90.67 overall on OmniDocBench, beating MonkeyOCR-pro-3B [17] by 1.82 and dots.ocr by 2.26 points.It also records a text edit distance of 0.047, a formula CDM score of 88.46, TEDS of 88.22, TEDS-S of 92.38, and reading-order edit distance of 0.044. It is best or second-best in 6 of 9 document-type categories.Source: §5.1.2 · p. 21

5.2.1 Layout Analysis

  • MinerU2.5 achieves the top Full Page F1-score@PageIoU across all three layout benchmarks and leading F1-scores for most individual element types.The zero-shot comparison maps labels to five broad categories and uses PageIoU, which measures spatial overlap without category labels for Full Page scoring.Source: Table 9 · p. 22, §5.2.1 · pp. 22–23

5.2.2 Table Recognition

  • MinerU2.5 achieves state-of-the-art table recognition on most benchmarks and competitive results on the rest.On PubTabNet, despite using 20% of the training set, it places second in TEDS and third in TEDS-S; on the diverse in-house benchmark of approximately 500 tables, it slightly outperforms Gemini-2.5 Pro. The in-house set varies row and column counts, merged cells, content, and line styles.Source: Table 10 · p. 23, §5.2.2 · pp. 23–24

5.2.3 Formula Recognition

  • MinerU2.5 achieves the best results on four of seven formula-recognition datasets and a second-best result on one.Its CDM scores are 96.4 on SCE and 90.6 on LaTeX-80MM. On the in-house Chinese set, it places second with 90.6, and it leads on both Fuzzy Math and Complex.Source: Table 11 · p. 24, §5.2.3 · pp. 24–25

6 Conclusion · Appendix

MinerU2.5 is a 1.2B-parameter vision-language model that uses a decoupled, coarse-to-fine strategy for state-of-the-art efficient document parsing. The paper presents it as a tool for preparing structured data for pretraining and improving RAG over complex documents.

6 Conclusion

  • MinerU2.5 is a 1.2B-parameter model that achieves state-of-the-art efficient document parsing with a decoupled, coarse-to-fine strategy.It separates global layout analysis from local recognition to address the trade-off between performance and cost.Source: §6 · p. 25
  • By rapidly converting unstructured document collections into clean, structured data, MinerU2.5 can support the curation of high-quality pretraining corpora.Source: §6 · p. 25
  • By preserving the semantic integrity of tables, formulas, and layouts, MinerU2.5 is positioned to improve the quality and reliability of RAG systems.Source: §6 · p. 25
  • The paper frames MinerU2.5 as a foundational tool for the LLM era, enabling access to knowledge contained in complex documents for next-generation AI applications.Source: §6 · p. 25

A Qualitative examples · A.1 Overview

The qualitative examples illustrate MinerU2.5’s document, table, and formula parsing across varied page types, and compare its outputs with earlier and state-of-the-art models.

A Qualitative examples

  • Document-parsing examples span academic literature, books, textbooks, reports, slides, exams, notes, newspapers, and magazines.Figures 8 to 10 present these PDF-page examples.Source: A Qualitative examples · p. 30
  • Table-recognition examples cover photographed, colorful, formula-containing, empty-cell, handwritten, large, rotated, no-line, three-line, and full-line tables.Figures 11 and 12 show these table types.Source: A Qualitative examples · p. 30
  • Formula-recognition examples include formulas with backgrounds, Chinese text, matrices, conditions, handwriting, blur, multiple columns, or degradation.Figures 13 and 14 include nested conditions as well.Source: A Qualitative examples · p. 30
  • Compared with MinerU2-VLM and MinerU2-pipeline, MinerU2.5 improves parsing of rotated and merged-cell tables, Chinese formulas, and multiline complex formulas.Figures 19 and 20 also illustrate finer layout-detection bounding boxes and better performance on watermark pages than the previous version.Source: A Qualitative examples · p. 30
  • MinerU2.5 performs strongly on PDF pages with complex elements and is relatively better than existing state-of-the-art models.Source: A Qualitative examples · p. 30
  • MinerU2.5 parses complex-table pages better than other models, which show table-structure, content-loss, and table-splitting errors.Examples include full-page, dense, colorful, tightly arranged, irregularly merged, and line-free tables.Source: A Qualitative examples · p. 30
  • MinerU2.5 correctly parses nested conditional expressions and complex or nested matrices where other state-of-the-art models may produce incorrect outputs.These examples are shown in Figures 27 to 29.Source: A Qualitative examples · p. 30
  • MinerU2.5 shows strong results on complex-layout pages with alternating text and images, few-frame tables, and watermarks compared with other models.These examples appear in Figures 30 to 32.Source: A Qualitative examples · p. 30

A.1.1 Among PDF types

Figures show document layouts alongside rendered markdown outputs across several PDF types, from academic and financial documents to notes and periodicals.

  • Figure 8 shows layouts and rendered markdown outputs for academic literature, books, and textbooks.Source: Figure 8 · p. 31
  • Figure 9 shows layouts and rendered markdown outputs for research reports, financial reports, slides, and exam papers.Source: Figure 9 · p. 32
  • Figure 10 shows layouts and rendered markdown outputs for notes, newspapers, and magazines.Source: Figure 10 · p. 33

A.1.2 Among Table types · A.1.3 Among Formula types · A.2 Compare to Previous Versions

The supplied passages identify rendered outputs for various table types and formula types.

A.1.2 Among Table types

A.1.3 Among Formula types

A.2.1 Table · A.2.2 Formula

MinerU2.5 is reported to perform better than its previous version on rotated and merged-cell tables, as well as Chinese, multi-line, and complex formulas.

A.2.1 Table · A.2.2 Formula

  • MinerU2.5 performs better than its previous version across the reported table and formula cases.The comparisons cover rotated and merged-cell tables, plus formulas with Chinese text and multi-line or complex layouts.Source: Figure 15 · p. 38, Figure 16 · p. 39, Figure 17 · p. 40

A.2.1 Table

A.2.2 Formula

  • MinerU2.5 performs better than its previous version on formulas with Chinese text.Source: Figure 17 · p. 40
  • The formula comparisons report improvements for Chinese formulas and multi-line, complex formulas.Source: Figure 17 · p. 40, Figure 18 · p. 41
  • MinerU2.5 performs better than its previous version on multi-line and complex formulas.Source: Figure 18 · p. 41

A.2.3 Layout&OCR · A.3 Compare with Others

MinerU2.5 is shown with finer layout detection than its previous version and fewer detection omissions on a watermark page.

A.2.3 Layout&OCR

  • MinerU2.5 achieves finer layout detection than its previous version.Figure 19 presents the comparison.Source: Figure 19 · p. 42
  • MinerU2.5 has fewer detection omissions than its previous version on a watermark page.Figure 20 presents the comparison.Source: Figure 20 · p. 43

A.3.1 Table … B Prompt Details

The appendices show comparisons across table, formula, and layout/OCR examples, then specify the prompts, input constraints, and output formats used in MinerU2.5’s two-stage inference. The examples illustrate varied document structures, while the prompt details define how layout and content are represented.

A.3.1 Table

A.3.2 Formula

A.3.3 Layout&OCR

B Prompt Details

  • The prompt appendix details the prompts used during MinerU2.5’s two-stage inference and their corresponding output formats.Source: B Prompt Details · p. 56

B.1 Layout Detection

  • Layout detection resizes images to 1036 × 1036 and outputs elements in sequence with relative box coordinates, categories, and rotation directions.Source: §B.1 · p. 56

B.2 Text Recognition

  • Text recognition outputs raw OCR results from native-resolution images, limiting image tokens to 4–2048 and scaling images if that limit is exceeded.Source: §B.2 · p. 56

B.3 Formula Recognition

  • Formula recognition extracts formulas as LaTeX from native-resolution images, with image tokens limited to 4–2048 and scaling applied if necessary.Source: §B.3 · p. 56

B.4 Table Recognition

  • Table recognition outputs tables in OTSL format for data processing, using native-resolution images subject to the same 4–2048 image-token limit and scaling rule.Source: §B.4 · p. 57

Related papers

Loading 2509.22186v2…