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
TL;DR
Document parsing requires efficient handling of high-resolution, dense, and structurally complex pages. MinerU2.5 addresses this with a decoupled coarse-to-fine pipeline that combines downsampled global layout analysis with native-resolution local recognition, and it achieves state-of-the-art performance while using a lightweight design.
Problem
High-resolution, dense, and structurally complex documents challenge native-resolution processing, parameter efficiency, and robust parsing.
Method
MinerU2.5 decouples global layout analysis from local content recognition through a two-stage coarse-to-fine inference strategy.
Results
90.67 overall score on OmniDocBench, outperforming MonkeyOCR-pro-3B by 1.82 points and dots.ocr by 2.26 points.
Takeaways & Limitations
MinerU2.5 serves as a lightweight foundation for converting unstructured document collections into structured data while preserving table, formula, and layout semantics.
Takeaways & Limitations
Traditional modular pipelines have interdependencies that make usage, maintenance, and updates cumbersome and less efficient.
Abstract
from arXiv · showhide
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
Document parsing must handle high-resolution, dense, structurally complex pages while remaining efficient and robust. MinerU2.5 addresses this with decoupled coarse-to-fine inference and reports broad recognition improvements.
- High-resolution, dense documents require native-resolution processing, parameter efficiency, and robustness to complex layouts.
- MinerU2.5 separates global layout analysis from local content recognition in a two-stage coarse-to-fine architecture.It analyzes downsampled images globally, then recognizes text, formulas, and tables in native-resolution crops.
- The decoupled strategy reduces computational cost by an order of magnitude while improving interpretability, hallucination control, and independent stage optimization.The cost reduction primarily avoids the O(N2) visual-token complexity of end-to-end native-resolution approaches.
- MinerU2.5 achieves state-of-the-art performance across public benchmarks and improves practical application and user experience over MinerU2.
- Its layout analysis preserves headers, footers, and page numbers while standardizing labels for lists, references, and code blocks.
- The system delivers high-quality complex and mixed-language formula parsing and robust recognition of rotated, borderless, and partially bordered tables.
2 Related Work
Document parsing research spans modular pipelines, end-to-end VLMs, and newer multi-stage systems. These approaches trade specialized optimization, semantic consistency, resolution flexibility, efficiency, and maintenance complexity in different ways.
- Traditional Pipelines: Traditional pipelines sequentially specialize layout detection, text recognition, and reading order, but their interdependencies make usage, maintenance, and updates cumbersome.
- General-Purpose VLMs: General-purpose and open-source VLMs provide an alternative document-understanding paradigm with strong OCR and formula-recognition results.
- End-to-End Approaches: End-to-end domain-specific models unify text, formula, table, and chart recognition, eliminating cascaded processing stages.
- Multi-Stage Approaches: Multi-stage VLM methods combine page-level layout analysis with regional recognition, but fixed-resolution crop processing can distort extreme-aspect-ratio subregions.
3 MinerU2.5
MinerU2.5 combines a native-resolution vision-language architecture with a two-stage parsing pipeline that first analyzes global layout and then recognizes cropped regions. Its staged alignment, document parsing training, and targeted fine-tuning support strong parsing performance across challenging documents.
- Model Architecture: MinerU2.5 uses a Qwen2-VL-inspired architecture with a vision encoder, patch merger, and language model.The language model uses M-RoPE, the vision encoder supports dynamic resolutions, and pixel-unshuffle balances efficiency and performance.
- Parsing Strategy: High-resolution VLM parsing suffers from redundant tokens in blank regions, while crop-based and native-resolution approaches respectively lose context or incur O(N2) complexity.
- Two-Stage Inference: MinerU2.5 first performs global layout analysis on a 1036 × 1036 thumbnail, then parses layout-guided native-resolution crops capped at 2048 × 28 × 28 pixels.The two stages balance global visibility, fine-grained detail preservation, and computational efficiency.
- Training: Two-stage modality alignment training freezes the vision encoder and language model during image-caption alignment, then unfreezes all parameters for visual instruction tuning.
- Training: VQA-based modality alignment improves layout analysis and content recognition, whereas omitting it increases losses and reduces overall performance.
- Training: Document parsing pre-training and targeted fine-tuning build layout and recognition abilities while improving challenging scenarios and retaining established capabilities.The data engine mixes model-labeled and public data, samples diverse examples, and supports hard-sample mining and data iteration.
4 Data Engine
MinerU2.5’s data engine combines curated, balanced document data, refined automated annotations, and expert-selected hard cases with redesigned parsing tasks, layout standards, and recognition methods.
- Data Curation: The Data Engine builds a diverse, balanced Chinese-English corpus through curation across layout, document type, element distribution, and language.Page-level clustering, stratified metadata sampling, preliminary detection, and language filtering support this coverage.
- Pre-training Data Preparation: Specialized models refine automated text, formula, and table annotations into image-annotation pairs covering layout, text, formula, and table recognition.The resulting pre-training dataset uses targeted refinement for each content type.
- Layout Analysis: A hierarchical tagging system expands layout coverage to headers, footers, page numbers, fine-grained figure types, captions, code, algorithms, references, and lists.The system addresses inconsistent element definitions, granularity, and scope while preserving semantic distinctions.
- Layout Analysis: The enhanced layout paradigm treats analysis as a multi-task problem to handle rotated elements and streamline document analysis.It responds to limitations of object-detection formulations that ignore rotation and defer reading order prediction.
- Layout Analysis: PageIoU measures page-level spatial coverage, aligning evaluation more closely with visual quality than fixed-threshold IoU-based recall.In the illustrated cases, PageIoU scores the poorer prediction 0.78 and the superior prediction 0.97.
- Table Recognition: OTSL supports table recognition after geometric normalization by representing table structure with fewer structural tokens and shorter sequences than HTML.The four-stage pipeline detects the table box and rotation, canonicalizes the image, predicts OTSL, and converts it to HTML.
- Fine-tuning Dataset Construction: IMIC identifies difficult unlabeled samples for targeted human annotation, directing limited curation effort toward cases with the greatest expected improvement value.The strategy supports iterative dataset expansion and continuous model improvement.
5 Evaluation
MinerU2.5 is evaluated across diverse full-document, layout, table, and formula recognition benchmarks, achieving state-of-the-art or highly competitive results across most reported tasks. Its strongest results include leading overall scores, dense OCR accuracy, layout analysis, tables, and formulas.
- MinerU2.5 achieves state-of-the-art results across most evaluated benchmarks and metrics.
- 90.67 overall on OmniDocBench, exceeding MonkeyOCR-pro-3B by 1.82 points and dots.ocr by 2.26 points.
- 0.033 edit distance and 0.945 F1-score on dense English Ocean-OCR documents, with BLEU 0.909 and METEOR 0.950.
- 75.2 overall on olmOCR-bench, surpassing dots.ocr’s 73.6 by 1.6 points.
- Layout analysis: MinerU2.5 achieves the top Full Page F1-score@PageIoU across all layout-analysis benchmarks and leads on most individual element types.
- Table recognition: Using only 20% of PubTabNet’s training set, MinerU2.5 ranks second in TEDS and third in TEDS-S, while achieving the best FinTabNet result.
- Formula recognition: Across seven formula datasets, MinerU2.5 is best on four and second-best on one, including CDM scores of 96.4 on SCE and 90.6 on LaTeX-80MM.
6 Conclusion
MinerU2.5 is presented as a lightweight, efficient document-parsing model whose decoupled coarse-to-fine strategy delivers state-of-the-art performance. The authors position it as useful for converting complex documents into structured data for downstream AI systems.
- MinerU2.5 is a 1.2B-parameter vision-language model using a decoupled coarse-to-fine strategy for efficient document parsing.
- The model is positioned as a tool for converting unstructured document collections into clean, structured data and preserving tables, formulas, and layouts for RAG systems.
A Qualitative examples
Qualitative examples cover document, table, and formula parsing across varied page types and difficult visual conditions. They show improvements over earlier MinerU versions and competitive behavior on complex pages compared with other models.
- Document parsing: Examples span document parsing across academic literature, books, textbooks, reports, slides, exams, notes, newspapers, and magazines.
- Table recognition: Table examples include photographs, colorful backgrounds, formulas, empty cells, handwriting, rotation, large tables, and multiple line styles.
- Formula recognition: Formula examples cover Chinese text, matrices, nested conditions, handwriting, blur, multiple columns, and degradation.
- Improvements: MinerU2.5 improves on previous MinerU versions for rotated tables, merged cells, Chinese and complex formulas, finer bounding boxes, and watermark pages.
- Complex pages: On complex table pages, other models encounter structure, content-loss, and table-splitting errors, while MinerU2.5 produces better parsing outputs.
- Complex pages: MinerU2.5 correctly parses nested conditional expressions and complex matrices where other state-of-the-art models may produce incorrect outputs.
A.1.1 Among PDF types
The qualitative document examples cover academic, educational, professional, presentation, examination, personal, newspaper, and magazine pages. Figures 8–10 provide rendered layout and markdown outputs for these document types.
- Figure 8 presents layout and rendered markdown outputs for academic literature, books, and textbooks.
- Figure 9 presents layout and rendered markdown outputs for research reports, financial reports, slides, and exam papers.
- Figure 10 presents layout and rendered markdown outputs for notes, newspapers, and magazines.
A.1.2 Among Table types
The rendered examples cover multiple table and formula types, including borderless, three-line, and empty-cell tables alongside varied formulas.
- Rendered outputs illustrate diverse table types, including borderless tables, three-line tables, and tables with empty cells.
- Rendered outputs also illustrate various formula types across the examples.
A.2.1 Table
Compared with the previous version, MinerU2.5 performs better on rotated tables and tables with merged cells, as well as Chinese and complex formulas.
- Table: MinerU2.5 performs better than the previous version on rotated tables.
- Table: MinerU2.5 performs better than the previous version on tables with merged cells.
- Formula: MinerU2.5 performs better than the previous version on formulas containing Chinese.
- Formula: MinerU2.5 performs better than the previous version on multiline and complex formulas.
A.2.3 Layout&OCR
The comparisons highlight finer layout detection and fewer detection omissions on watermark pages for MinerU2.5 relative to its previous version.
- Layout&OCR: MinerU2.5 achieves finer layout detection than its previous version.
- Layout&OCR: MinerU2.5 achieves fewer detection omissions than its previous version on watermark pages.
A.3.1 Table
The section presents comparisons across challenging document examples and describes the two-stage inference outputs for layout, OCR, formulas, and tables. Layout analysis uses resized images, while recognition retains native resolution subject to an image-token limit.
- Table: MinerU2.5 is compared with other systems on full-page, content-dense, colored, tightly arranged, irregularly merged, and frameless tables.
- Table: Additional comparisons cover nested conditional expressions, complex and nested matrices, academic literature, financial reports, and textbooks with watermarks.
- Two-stage inference: The layout detection output includes each element’s relative coordinates, category, and rotation direction in sequence.
- Two-stage inference: The layout-analysis input is resized to 1036 × 1036, while content-recognition inputs retain native resolution with image tokens limited to 4 to 2048.
- Two-stage inference: Formula recognition extracts formulas into LaTeX, and table recognition outputs structured OTSL representations.