PaperlayerView exact revision on arXiv

Source-linked AI summary

MinerU2.5-Pro: Pushing the Limits of Data-Centric Document Parsing at Scale

Bin Wang, Tianyao He, Linke Ouyang, Fan Wu, Zhiyuan Zhao, Tao Chu, Yuan Qu, Zhenjiang Jin, Weijun Zeng, Ziyang Miao, Bangrui Xu, Junbo Niu, Mengzhang Cai, Jiantao Qiu, Qintong Zhang, Dongsheng Ma, Yuefeng Sun, Hejun Dong, Wenzheng Zhang, Jutao Xiao, Jiayong Shi, Pengyu Liao, Xiaomeng Zhao, Huaping Zhong, Liqun Wei, Jing Yu, Jie Yang, Wei Li, Shasha Wang, Qianqian Wu, Xuanhe Zhou, Weijia Li, Zhenxiang Li, Zhongying Tu, Jiang Wu, Lijun Wu, Chao Xu, Kai Chen, Wentao Zhang, Yu Qiao, Bowen Zhou, Dahua Lin, Conghui He

arXiv:2604.04771v2cs.CVcs.CL

TL;DR

Document parsing research has emphasized architecture and inference efficiency, although diverse models share failures on the same hard samples, indicating a training-data bottleneck. MinerU2.5-Pro keeps MinerU2.5’s 1.2B-parameter architecture fixed and improves data engineering, training strategy, and evaluation. It reaches 95.69 on OmniDocBench v1.6, 2.71 points above the same-architecture baseline, while surpassing existing methods.

  • Problem

    Diverse state-of-the-art parsers share failures on the same hard samples, while training data has insufficient coverage and unreliable annotations.

  • Method

    MinerU2.5-Pro uses a coverage-, informativeness-, and accuracy-oriented Data Engine, progressive training, and OmniDocBench v1.6 while keeping the 1.2B-parameter architecture unchanged.

  • Results

    95.69 on OmniDocBench v1.6 improves the same-architecture baseline of 92.98 by 2.71 points and surpasses all existing methods.

  • Takeaways & Limitations

    At the current stage of architectural maturity, co-optimizing training-data coverage, informativeness, and annotation accuracy is presented as a primary lever for progress.

  • Takeaways & Limitations

    Element matching remains limited because equivalent formats and structurally different but valid representations can be difficult to distinguish fairly.

Abstract

from arXiv · show

Current document parsing methods advance primarily through model architecture innovation, while systematic engineering of training data remains underexplored. Yet state-of-the-art models spanning diverse architectures and parameter scales exhibit highly consistent failure patterns on the same set of hard samples, suggesting that the performance bottleneck stems from shared deficiencies in training data rather than from architectural differences. Building on this finding, we present MinerU2.5-Pro, which advances the state of the art purely through data engineering and training strategy design while retaining the 1.2B-parameter architecture of MinerU2.5 unchanged. At its core is a Data Engine co-designed around coverage, informativeness, and annotation accuracy: Diversity-and-Difficulty-Aware Sampling expands training data from under 10M to 65.5M samples while mitigating distribution shift; Cross-Model Consistency Verification leverages output consensus among heterogeneous models to assess sample difficulty and generate reliable annotations; the Judge-and-Refine pipeline improves annotation quality for hard samples through render-then-verify iterative correction. A three-stage progressive training strategy--large-scale pre-training, hard sample fine-tuning, and GRPO alignment--sequentially exploits these data at different quality tiers. On the evaluation front, we rectify element-matching biases in OmniDocBench v1.5 and introduce a Hard subset, establishing the more discriminative OmniDocBench v1.6 protocol. Without any architectural modification, MinerU2.5-Pro achieves 95.69 on OmniDocBench v1.6, improving over the same-architecture baseline by 2.71 points and surpassing all existing methods, including those based on models with over 200x more parameters.

1 Introduction

State-of-the-art document parsers share failures on the same hard samples, indicating that shared training-data deficiencies—not architecture—are the main bottleneck. MinerU2.5-Pro addresses this through data engineering, improved evaluation, and progressive training while retaining the 1.2B-parameter architecture.

  • Motivation: Models spanning architectures and parameter scales exhibit highly similar failures on the same hard samples, pointing to shared training-data deficiencies.The analysis identifies insufficient coverage and unreliable annotations as central data problems.
  • Motivation: Under 10M training pages leave long-tail scenarios underrepresented, while hard-sample annotations remain especially error-prone.Examples include complex nested tables and dense formula layouts.
  • Results: 95.69 on OmniDocBench v1.6 improves the fixed-architecture baseline of 92.98 by 2.71 points and surpasses existing methods.OmniDocBench v1.6 corrects element-matching biases and adds a Hard subset within a Base/Hard/Full framework.
  • Approach: 65.5M pages are produced through a Data Engine co-designed for coverage, informativeness, and annotation accuracy.Its components are DDAS, CMCV, and a Judge-and-Refine annotation pipeline.
  • Approach: A three-stage strategy uses large-scale pre-training, hard-sample fine-tuning, and GRPO format alignment across data-quality tiers.The strategy is designed to exploit the progressively refined data produced by the Data Engine.

2 Related Work

Document parsing has evolved through pipeline, end-to-end VLM, and decoupled VLM paradigms, with architecture and inference efficiency dominating methodological development. This work positions systematic data engineering and evaluation-protocol design as complementary research dimensions.

  • Existing paradigms: Pipeline methods separate layout, text, table, and formula subtasks but suffer from error propagation and inter-module information loss.Their modular structure permits independent component optimization.
  • Existing paradigms: End-to-end VLM methods map document images directly to structured output, avoiding cascading pipeline errors.Representative systems include Nougat, GOT-OCR 2.0, Ocean-OCR, olmOCR, and dots.ocr.
  • Existing paradigms: Decoupled VLM methods separate layout analysis from content recognition, combining pipeline controllability with VLM semantic modeling.MinerU2.5 uses this paradigm in a single 1.2B-parameter model with native-resolution support.
  • Data-centric gap: Across these paradigms, architecture design and inference efficiency have received more attention than systematic training-data engineering.The paper treats coverage, informativeness, and annotation accuracy as a unified data-construction problem.
  • Data-centric gap: CMCV extends ensemble-style difficulty assessment by using multi-model disagreement and coupling it to downstream annotation in a closed loop.This approach adapts active-learning and query-by-committee principles to document parsing.
  • Evaluation: OmniDocBench v1.6 addresses overlooked evaluation-protocol bias by correcting element matching with Multi-Granularity Adaptive Matching.Matching choices affect scores because systems differ in granularity, segmentation, and output conventions.

3 Data Engine

MinerU2.5-Pro’s Data Engine jointly optimizes data coverage, difficulty informativeness, and annotation accuracy through coordinated sampling, cross-model verification, and hard-sample refinement. It expands and stratifies training data while assigning different annotation and training treatments to Easy, Medium, and Hard samples.

  • 3.1 Diversity-and-Difficulty-Aware Sampling: DDAS jointly balances diversity and difficulty at page and element granularity, producing training data across layout, text, formula, and table subtasks.Page sampling uses clustering and CMCV-derived difficulty; element sampling independently applies both dimensions to text, formula, and table blocks.
  • 3.1 Diversity-and-Difficulty-Aware Sampling: DDAS couples distributional clustering with CMCV difficulty labels to increase training-signal density while controlling total data volume.Balanced sampling downsamples large clusters, upsamples small clusters, and upweights Medium and Hard samples.
  • 3.2 Cross-Model Consistency Verification: CMCV uses agreement among MinerU2.5, PaddleOCR-VL, and Qwen3-VL-30B to classify samples as Easy, Medium, or Hard.It applies task-specific consistency metrics including edit distance, TEDS, and CDM.
  • 3.2 Cross-Model Consistency Verification: Medium samples are prioritized during sampling because external-model agreement indicates they are learnable while exposing MinerU2.5’s capability gaps.The preferred difficulty mix varies by subtask: formula and table recognition are more sensitive to Hard samples, whereas text recognition benefits more from Medium samples.
  • 3.3 Annotation Pipeline for Hard Case: Hard samples require Judge-and-Refine correction followed by targeted expert annotation because model disagreement prevents reliable direct annotation.The pipeline addresses self-reflection’s tendency to accept erroneous outputs by using iterative correction before routing residual failures to experts.
  • Data Engine overview: 65.5M Easy and Medium samples receive CMCV-based automatic annotations, while 192K expert-annotated Hard samples support later training stages.The stratified dataset assigns large-scale automatic data to Stage 1 and high-quality Hard data to Stages 2 and 3.

4 Progressive Training Strategy

MinerU2.5-Pro retains MinerU2.5’s 1.2B-parameter architecture and progressively trains it on increasingly targeted data. The stages move from broad CMCV-annotated pre-training to expert-annotated Hard-sample fine-tuning and metric-aligned GRPO.

  • Architecture: The unchanged model retains MinerU2.5’s 1.2B-parameter decoupled coarse-to-fine architecture.It uses the NaViT-675M vision encoder and Qwen2-0.5B language model without structural modification.
  • Stage 1: Large-Scale Pre-Training: Stage 1 pre-trains on large-scale CMCV auto-annotated data to build comprehensive foundational capabilities.The training set contains approximately 65.5M samples across recognition, layout, formula, table, and image-analysis tasks.
  • Stage 2: Hard-Sample Fine-Tuning: Stage 2 fine-tunes on 192K expert-annotated Hard samples mixed with Stage 1 replay data to improve hard-scenario performance while preserving generalization.Hard-to-replay ratios differ by subtask, reflecting differences in hard-sample volume and baseline performance.
  • Stage 3: Reinforcement Learning with GRPO: Stage 3 uses GRPO to optimize task-level metrics that token-wise cross-entropy does not directly reflect.GRPO samples output groups, computes task-specific rewards, and updates the policy using within-group relative advantages without a separate reward model.
  • Stage 3: Reinforcement Learning with GRPO: GRPO rewards directly use edit distance, CDM, TEDS, and category IoU for text, formula, table, and layout tasks.This reward design matches the optimization signals to the evaluation metrics.

5 OmniDocBench v1.6

OmniDocBench v1.6 addresses v1.5’s granularity-matching bias and weak Hard-sample coverage. It combines Multi-Granularity Adaptive Matching with a Base/Hard/Full protocol for more discriminative evaluation.

  • 5.1 Motivation: Fixed-granularity one-to-one matching in v1.5 can sharply penalize semantically correct predictions whose segmentation differs from the ground truth.The problem affects multi-line formulas and dense text when predictions split or merge regions differently from annotations.
  • 5.1 Motivation: Hard samples are virtually absent from v1.5, so top models cluster on mostly low-to-medium difficulty documents.This limits the benchmark’s discriminative power.
  • 5.1 Motivation: v1.6 introduces Multi-Granularity Adaptive Matching and a dedicated Hard subset to address matching bias and coverage gaps.The protocol retains the original set while adding challenging samples.
  • 5.2 Multi-Granularity Adaptive Matching: MGAM keeps ground truth fixed and searches prediction-side segmentation candidates before selecting the globally best matching.Its three stages evaluate original granularity, split predictions, and enumerated intermediate partitions.
  • 5.2 Multi-Granularity Adaptive Matching: The final matching selects the candidate with the highest aggregate score, after which the task-specific metric is computed.Examples include CDM for formulas and edit distance for text.
  • 5.2 Multi-Granularity Adaptive Matching: With MGAM, evaluation becomes neutral to output granularity and format preferences, removing a systematic source of scoring variance.The same matching approach is also reused for dense text elements.
  • 5.3 Hard Subset: The Hard subset contains 296 pages covering complex nested tables, dense mathematical layouts, and unconventional structures.All Hard-subset samples are excluded from MinerU2.5-Pro training.
  • 5.3 Hard Subset: OmniDocBench v1.6 reports Base, Hard, and Full tiers containing 1,355, 296, and 1,651 pages, respectively.Base preserves historical comparability, Hard increases sensitivity where standard evaluations saturate, and Full combines both.

6 Experiments

MinerU2.5-Pro is evaluated under a unified OmniDocBench v1.6 protocol across end-to-end and element-specific parsing tasks, with ablations isolating progressive training contributions. It leads overall and shows especially strong performance on hard scenarios, tables, formulas, and reading order.

  • End-to-End Document Parsing: 95.69 on OmniDocBench v1.6 Full ranks MinerU2.5-Pro first, improving over the same-architecture MinerU2.5 baseline by 2.71 points.The comparison uses a unified evaluation environment and reports text, formula, table, and reading-order extraction.
  • End-to-End Document Parsing: 94.08 on the Hard subset exceeds GLM-OCR and PaddleOCR-VL-1.5, both at 92.01, by 2.07 points.The Base subset is near saturation, with the top three models within 0.5 points, whereas the Hard subset separates performance more clearly.
  • End-to-End Document Parsing: 97.29 CDM, 93.42 TEDS, 95.92 TEDS-S, and 0.120 reading-order score are the best reported sub-metric results for MinerU2.5-Pro.The evaluation averages text edit distance, table TEDS, and formula CDM, while also reporting additional sub-metrics.
  • Training Stage Ablation: +1.31 is the largest single-stage gain from large-scale supervised fine-tuning, while hard-sample fine-tuning adds +0.96 and GRPO adds +0.45.Hard-sample fine-tuning raises table TEDS from 90.37 to 92.87, while GRPO raises formula CDM from 96.48 to 97.29.
  • Element-Specific Parsing: 0.019 text edit distance on Full is a 30.5% reduction from MinerU2.5’s 0.028 baseline.Without category priors, end-to-end models such as DeepSeek-OCR 2 and FireRed-OCR show significant degradation.
  • Element-Specific Parsing: MinerU2.5-Pro leads on five formula-recognition dimensions and is second on the remaining four, while reaching 99.20 CDM on OmniDocBench Base.It trails on handwritten formulas against Qwen3.5-397B and on SCE against GLM-OCR.
  • Element-Specific Parsing: 91.10 Overall TEDS and 94.48 TEDS-S rank MinerU2.5-Pro first, with a 4.18-point Hard-subset advantage over MinerU2.5.The strongest table-recognition advantage appears on hard samples, while GLM-OCR is slightly better on OmniDocBench Base and CCOCR.

7 Conclusion

The paper concludes that systematic data engineering can improve document parsing while keeping the 1.2B-parameter architecture fixed, but identifies unresolved evaluation and structural-understanding limitations. It therefore presents data construction and evaluation tools as complementary to architectural innovation.

  • Conclusion: 95.69 versus 92.98 on OmniDocBench v1.6 improves the overall score solely through data engineering with the 1.2B-parameter architecture fixed.The paper reports that MinerU2.5-Pro surpasses all existing methods under this unchanged architecture.
  • Conclusion: The Data Engine expands training data from under 10M to 65.5M pages, while progressive training matches data quality tiers.The evaluation protocol also corrects matching biases and adds Base, Hard, and Full tiers.
  • Limitations and Future Directions: OmniDocBench v1.6 still cannot fully resolve format and structural ambiguity because equivalent content can have different valid representations.The paper identifies semantic-equivalence-aware evaluation as an open problem.
  • Limitations and Future Directions: Vertical domains such as finance, legal, and medical require domain-specific evaluation sets, while annotation precision becomes more challenging near human-level performance.These concerns define boundaries for the coverage and reliability of the current evaluation protocol.
  • Limitations and Future Directions: The work measures content accuracy, leaving hierarchical, cross-reference, and cross-page structural relationships as a next step toward structured semantic understanding.The paper distinguishes content extraction from deeper document-level semantic structure.

A Prompt Design and Task Examples

MinerU2.5-Pro uses a unified image-plus-task prompt interface across layout, text, formula, table, and image analysis, with task-specific structured outputs. These tasks support localization, recognition, serialization, and content extraction for diverse document elements.

  • A Prompt Design and Task Examples: The five task suffixes pair a single <image> token with plain-text instructions and require no few-shot examples or structured metadata.The interface covers Layout Detection, Text Recognition, Formula Recognition, Table Recognition, and Image Analysis.
  • A.1 Layout Detection: Layout Detection localizes page regions and assigns semantic categories, rotation flags, and normalized bounding boxes in reading order.Coordinates use a [0, 999] grid, and regions are emitted top-to-bottom and left-to-right for left-to-right scripts.
  • A.1 Layout Detection: Layout Detection outputs newline-delimited region descriptors containing coordinates, category labels, and rotation directions.Each descriptor uses box, reference, and rotation control tokens.
  • A.2 Text Recognition: Text Recognition transcribes original-resolution crops into raw plain text while preserving whitespace, punctuation, and inline symbols.The crops are produced by Stage 1 Layout Detection.
  • A.3 Formula Recognition: Formula Recognition converts cropped formula regions into compilable LATEX, supporting inline, display, equation-numbered, and multi-line formulas.Display formulas use delimiters, equation numbers use \tag{}, and multi-line outputs concatenate independently recognized lines in reading order.
  • A.4 Table Recognition: Table Recognition serializes cropped tables as OTSL token sequences containing cell content and row boundaries, then converts them to HTML.Cells may contain plain text, inline LATEX, or both; <fcel> marks cells and <nl> separates rows.
  • A.5 Image-Aware Parsing: Image Analysis first classifies image regions into fine-grained types, then extracts captions and embedded text, formulas, tables, or combinations.Its output contains primary and secondary classes, caption text, and structured content fields.
  • A.5 Image-Aware Parsing: Image Analysis represents each result with class, subclass, caption, and content fields delimited by special tokens.A standalone formula can be classified as a pure formula with empty caption and directly extracted content.

B Extended Parsing Capabilities

MinerU2.5-Pro extends parsing beyond recognition accuracy with image-aware extraction, paragraph and cross-page table merging, and preservation of images embedded in tables. These capabilities target complex multi-page documents and improve end-to-end completeness even though they are outside OmniDocBench’s single-page scoring scope.

  • Practical impact: These extensions target multi-page, richly illustrated, structurally complex documents and improve parsing completeness and usability without affecting OmniDocBench scores.OmniDocBench focuses on single-page content recognition.
  • Image-aware parsing: Image-aware parsing classifies image regions into subtypes and applies differentiated extraction strategies for charts, text images, and table-like images.The release has not yet applied Data Engine optimization to image-analysis data, leaving room for improvement.
  • Truncated paragraph merging: Truncated paragraph merging reconnects text fragments split by columns, intervening figures or tables, or unusually wide line spacing.The merge decision is integrated into Layout Detection and reassembles continuous paragraphs during rendering.
  • Cross-page table merging: Cross-page table merging uses rule-based candidate detection followed by semantic understanding at page boundaries to reconstruct complete tables.Compatible column counts and structural patterns identify candidate table pairs.
  • Cross-page table merging: Per-column merge decisions distinguish direct concatenation from semantic merging when different columns require different treatments.This handles tables where some columns continue across a page break while others remain separate rows.
  • In-table image detection: Embedded table images are detected inside table bounds, masked during Table Recognition, and restored as linked HTML image references.The three-step process preserves image-to-cell correspondence while preventing images from interfering with table structure recognition.
  • In-table image detection: The system preserves embedded images as HTML <img> elements with unique identifiers linking them to extracted image content blocks.Restoration occurs after the OTSL sequence is generated.

C.1 Detailed Results on Base Subset

On OmniDocBench v1.6 Base, leading models perform closely, while MinerU2.5-Pro ranks second overall and leads several component metrics.

  • Base subset: 96.12 places MinerU2.5-Pro second on the Base Overall score, only 0.07 points behind GLM-OCR.The top six Overall scores span 94.49–96.19.
  • Base subset: 0.033 gives MinerU2.5-Pro the leading Base Text Edit Distance result over GLM-OCR’s 0.039.Lower values indicate the reported leading result for this metric.
  • Base subset: 94.49 gives MinerU2.5-Pro the leading Base Table TEDS result over GLM-OCR’s 94.36.The comparison is reported among the detailed Base-subset results.
  • Base subset: 0.109 gives MinerU2.5-Pro the leading Base Reading Order result over GLM-OCR’s 0.122.The reported Base results show tightly clustered overall performance.

C.2 Detailed Results on Hard Subset

The OmniDocBench v1.6 Hard subset separates model capabilities more clearly than Base, with MinerU2.5-Pro achieving the highest score and strongest reported robustness.

  • Hard subset: 94.08 makes MinerU2.5-Pro the Hard-subset leader, 1.60 points ahead of PaddleOCR-VL’s 92.48.MinerU2.5-Pro also achieves the best Formula CDM, Table TEDS, and Reading Order scores.
  • Hard subset: 97.54, 89.91, and 0.170 are MinerU2.5-Pro’s best Hard-subset scores for Formula CDM, Table TEDS, and Reading Order, respectively.These are the reported best scores for the three component metrics.
  • Hard subset: 4.18 points separate GLM-OCR’s Base score of 96.19 from its Hard score of 92.01, while HunyuanOCR declines 9.76 points from 92.45 to 82.69.The Hard ranking differs markedly from the Base ranking.
  • Hard subset: 2.04 points is MinerU2.5-Pro’s decline from Base to Hard, the smallest reported drop and the strongest robustness among the compared models.Gemini 3 Pro/Flash also perform strongly on Hard at 91.99 and benefit from strong hard-formula scores.

D Qualitative Comparison with SOTA Methods

MinerU2.5-Pro shows stronger qualitative parsing than competing methods across complex tables, multi-line formulas, matrices, and image-aware chart parsing. Its comparisons emphasize structural recovery, row-by-row formula analysis, and extracting structured chart content.

  • Formulas and matrices: Row-by-row formula analysis improves multi-line formula recognition, while end-to-end approaches merge or misalign equation lines.
  • Formulas and matrices: MinerU2.5-Pro captures complex matrix structure and alignment, while competing models exhibit symbol errors or structural collapse.
  • Charts and diagrams: Image-aware parsing extracts structured content from diverse chart types, whereas other models ignore or misinterpret chart content.
  • Complex tables: MinerU2.5-Pro correctly recovers rotated-table structure, while competing models produce misaligned rows or missing cells.
  • Complex tables: MinerU2.5-Pro preserves long merged-cell spans, whereas other models incorrectly split or duplicate merged cells.
Loading 2604.04771v2…