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.
Figures & tables
Figure 1: Performance comparison on OmniDocBench v1.6, which comprises Base (standard samples), Hard (challenging samples), and Full (overall) splits. Built upon MinerU2.5 [ 25 ] with its 1.2B-parameter architecture entirely unchanged, MinerU2.5-Pro improves the overall score from 92.98 to 95.69 purely through data engineering and training strategy design, outperforming both specialized document parsing models ( e.g . GLM-OCR [ 11 ] , PaddleOCR-VL-1.5 [ 7 ] , Youtu-Parsing [ 45 ] ) and general-purpose VLMs ( e.g . Gemini 3 Pro, Qwen3-VL-235B [ 43 ] ). Detailed results are presented in Table 2 .
Figure 2: Overview of the Data Engine pipeline. The system co-optimizes three dimensions—Coverage, Informativeness, and Accuracy—through three synergistic stages: Diversity-and-Difficulty-Aware Sampling (DDAS), Cross-Model Consistency Verification (CMCV), and the Annotation Pipeline for Hard Case.
Figure 3: The DDAS pipeline operates at two granularity levels. Upper : Page-level sampling for layout detection data—pages from the PDF pool are embedded via ViT-base, clustered, and resampled by jointly weighting cluster diversity and CMCV-derived difficulty, yielding about 60M pages with balanced distribution and difficulty coverage. Lower : Element-level sampling—the selected pages are parsed by layout detection models into text, formula, and table blocks; each element type is independently clustered and assessed by CMCV, then sampled to balance both diversity and difficulty at the element granularity. The two levels are combined to produce the final training data for layout, text, formula, and table subtasks.
Category
Parameter
Stage 1
Stage 2
Stage 3
Vision
Max Resolution
2048 × 28 × 28
2048 × 28 × 28
2048 × 28 × 28
#Tokens per Image
64–2048
64–2048
64–2048
Data
Dataset Type
Layout & OCR † & Image Analysis
Layout & OCR † & Image Analysis
Layout & Text & Formula & Table
#Samples
65.5M
3.9M (192K human-labeled)
192K
Model
Trainable
All
All
All
Sequence Length
8192
8192
8192
Table 1: Training configurations for the three-stage progressive strategy. All three stages share the same model architecture and resolution settings; they differ in data source, data scale, and optimization hyperparameters, reflecting the progression from broad coverage (Stage 1) to targeted refinement (Stage 2) to metric-level alignment (Stage 3).
Figure 4: Examples of element-matching bias in OmniDocBench v1.5. Semantically correct predictions receive low scores due to granularity mismatch between predicted and ground-truth segmentation.
Model Type
Methods
Param
Overall ↑
Text Edit ↓
Formula CDM ↑
Table TEDS ↑
Table TEDS-S ↑
Read Order Edit ↓
Specialized VLMs
MinerU2.5-Pro
1.2B
95.69
0.036
97.29
93.42
95.92
0.120
GLM-OCR [ 11 ]
0.9B
95.15
0.044
96.99
92.83
95.39
0.133
PaddleOCR-VL-1.5 [ 7 ]
0.9B
94.87
0.038
96.69
91.67
94.37
0.130
PaddleOCR-VL [ 6 ]
0.9B
94.11
0.040
95.70
90.65
93.74
0.135
Youtu-Parsing [ 45 ]
2.5B
93.68
0.044
93.45
92.02
95.00
0.116
Logics-Parsing-v2 [ 4 ]
4B
93.27
0.041
95.47
88.42
91.98
0.137
Table 2: Performance comparison of document parsing methods on OmniDocBench v1.6 Full across text, formula, table, and reading order extraction tasks.
Stage
Base
Hard
Full
Δ Full
Text ↓
CDM ↑
TEDS ↑
MinerU2.5 (baseline)
93.23
91.65
92.98
—
0.045
95.59
87.88
Stage 1: Large-Scale SFT
94.54
93.10
94.29
+1.31
0.039
96.40
90.37
+ Stage 2: Hard-Sample SFT
95.60
93.84
95.25
+0.96
0.036
96.48
92.87
+ Stage 3: GRPO
96.12
94.08
95.69
+0.45
0.036
97.29
93.42
Table 3: Training stage ablation on OmniDocBench v1.6.
Model
Type
Base ↓
Hard ↓
Full ↓
MinerU2.5-Pro
Decoupled
0.015
0.048
0.019
Qwen3.5-397B [ 33 ]
General
0.016
0.052
0.020
GLM-OCR [ 11 ]
Decoupled
0.016
0.053
0.021
Qwen3-VL-235B [ 43 ]
General
0.017
0.049
0.021
PaddleOCR-VL-1.5 [ 7 ]
Decoupled
0.018
0.056
0.022
MinerU2.5 [ 25 ]
Decoupled
0.023
0.066
0.028
Table 4: Text recognition (Edit Distance ↓ ) on OmniDocBench v1.6.
Model
OmniDoc
Public
Inhouse
Base
Hard
CPE
HWE
SCE
SPE
LaTeX-80M
Chinese
Fuzzy
MinerU2.5-Pro
99.20
98.79
98.97
95.38
97.04
99.44
97.23
95.28
94.90
PaddleOCR-VL-1.5 [ 7 ]
98.76
97.22
98.84
92.27
94.95
99.27
92.77
94.06
89.73
GLM-OCR [ 11 ]
98.75
98.28
96.74
95.10
97.77
98.42
95.39
94.35
93.75
PaddleOCR-VL [ 6 ]
98.72
97.64
98.93
94.45
95.88
99.30
93.67
94.35
91.56
Qwen3.5-397B [ 33 ]
98.19
97.25
98.32
97.59
95.87
99.41
95.17
78.24
90.53
Table 5: Formula recognition (CDM ↑ ) across multiple benchmarks. CPE, HWE, SCE, and SPE are the Complex Printed, Handwritten, Screen-Captured, and Simple Printed Expression subsets from UniMERNet [ 34 ] , respectively.
Figure 5: Layout Detection examples. The model localizes content regions with bounding boxes, category labels, and rotation flags on diverse document pages.
Figure 6: Text Recognition examples across Chinese, English, and mixed-language text regions.
Figure 7: Formula Recognition examples including single-line display formulas and complex multi-line equations.
Figure 8: Table Recognition examples showing OTSL token output and the corresponding rendered HTML for tables with varying complexity.
Figure 9: Image-aware parsing examples. The model classifies image regions into fine-grained subtypes and extracts structured content accordingly.
Figure 10: Illustration of truncated paragraph merging. In multi-column and complex layouts, Layout Detection splits continuous paragraphs into separate bounding boxes. The merge label predicted by the model reassembles them into coherent paragraphs during Markdown rendering.
Figure 11: Cross-page table merging example. The model performs semantic understanding at the junction of two table fragments identified by rule-based heuristics, producing per-column merge decisions to reconstruct the complete table.
Model
Overall ↑
Text Edit ↓
Formula CDM ↑
Table TEDS ↑
TEDS-S ↑
Read Order ↓
GLM-OCR
96.19
0.039
98.10
94.36
96.59
0.122
MinerU2.5-Pro
96.12
0.033
97.16
94.49
96.63
0.109
PaddleOCR-VL-1.5
95.72
0.032
97.18
93.17
95.31
0.118
Youtu-Parsing
94.87
0.038
94.32
94.07
96.56
0.101
Ovis2.6-30B-A3B
94.56
0.031
95.11
91.64
94.33
0.125
PaddleOCR-VL
94.49
0.035
95.43
91.51
94.39
0.123
Table 7: Detailed results on OmniDocBench v1.6 Base subset.
Model
Overall ↑
Text Edit ↓
Formula CDM ↑
Table TEDS ↑
TEDS-S ↑
Read Order ↓
MinerU2.5-Pro
94.08
0.052
97.54
89.91
93.61
0.170
PaddleOCR-VL
92.48
0.066
96.24
87.84
91.60
0.189
GLM-OCR
92.01
0.066
94.81
87.81
91.44
0.186
PaddleOCR-VL-1.5
92.01
0.065
95.74
86.75
91.30
0.181
Gemini 3 Flash
91.99
0.085
96.68
87.83
92.51
0.214
Gemini 3 Pro
91.99
0.083
97.23
87.03
91.68
0.198
Table 8: Detailed results on OmniDocBench v1.6 Hard subset.
Figure 12: In-table image detection and recognition. Embedded images within table cells are masked with placeholders during Table Recognition and restored as <img> references in the final HTML output.
Figure 13: Qualitative comparison on rotated table recognition. MinerU2.5-Pro correctly recovers the rotated structure, while competing models produce misaligned rows or missing cells.
Figure 14: Qualitative comparison on tables with long merged cells. MinerU2.5-Pro preserves the span structure, whereas other models incorrectly split or duplicate merged cells.
Figure 15: Qualitative comparison on complex matrix recognition. MinerU2.5-Pro accurately captures the matrix structure and alignment, while other models exhibit symbol errors or structural collapse.
Figure 16: Qualitative comparison on multi-line formula recognition. The row-by-row analysis of MinerU2.5-Pro faithfully reproduces each line, whereas competing models merge or misalign lines.
Figure 17: Qualitative comparison on image-aware chart parsing (Part 1). MinerU2.5-Pro extracts structured content from diverse chart types, while other models either ignore or misinterpret chart content.
Figure 18: Qualitative comparison on image-aware chart parsing (Part 2). Additional chart types demonstrating the generalization of MinerU2.5-Pro’s image analysis pipeline.
Document parsing models are improving rapidly, but existing benchmarks are still insufficient for a comprehensive evaluation of their capabilities. We introduce PureDocBench, a benchmark with 10 domains, 66 subcategories, and 1,475 pages. Each page has matched Clean, Digital, and Real versions, covering clean renders, simulated degradation, and real capture or sharing conditions. The resulting 4,425 images have source-derived annotations for text, formulas, tables, and reading order, checked through automated validation and human review. Keeping page content and annotations fixed allows direct comparisons across image conditions. We evaluate 58 models, including pipeline-based parsers, end-to-end parsers, and general vision-language models. The best three-track average score is 79.69 out of 100, and the mean across models is 63.84. Scores drop by 3.34 points on Digital and 11.61 points on Real on average, relative to Clean. TeleOCR leads on Clean, GLM-5.3-Flash on Digital, and Gemini 3.6 Flash on Real. The benchmark dataset, source documents, generation tools, and evaluation code are publicly available at https://github.com/zhihengli-casia/PureDocBench.
Zhiheng Li, Zongyang Ma, Jiaxian Chen +12
CASIA · UCAS · Beijing Key Laboratory of Super Intelligent Security of Multi-Modal Information +6
We introduce PaddleOCR-VL-1.6, an upgraded compact document parsing model built upon PaddleOCR-VL-1.5. Although PaddleOCR-VL-1.5 establishes a strong 0.9B baseline, its remaining errors concentrate in under-optimized regions where model behavior is unstable, data coverage is sparse, or supervision is unreliable. Rather than expanding the training corpus indiscriminately, PaddleOCR-VL-1.6 introduces a region-aware data optimization framework that identifies weak regions from the previous model, applies targeted enhancement to these regions, and improves the reliability of supervision signals. It further adopts a progressive post-training recipe based on curated data selection and reinforcement learning, pushing model performance to a higher level through staged optimization. PaddleOCR-VL-1.6 achieves a new state-of-the-art score of 96.33% on OmniDocBench v1.6, demonstrates strong competitiveness against top-tier VLMs, and provides a practical post-training recipe for the PaddleOCR-VL series.
We introduce OvisOCR2, a 0.8B document parsing model. OvisOCR2 is designed as an end-to-end parser: given a document page image, it generates a Markdown representation in natural reading order, covering text, formulas, tables, and visual regions. We build a data engine that combines filtered real-document annotations with synthetic pages whose rendered images and Markdown targets are derived from the same HTML source. The training recipe includes supervised fine-tuning, reinforcement learning on a 4B branch with a multi-component reward design, on-policy distillation into the 0.8B model, and model fusion. On OmniDocBench v1.6, OvisOCR2 achieves a state-of-the-art overall score of 96.58, placing an end-to-end model at the top of this leaderboard previously dominated by pipeline methods and highlighting the potential of end-to-end document parsing. On PureDocBench, OvisOCR2 also achieves the highest Avg3 score of 75.06. Beyond these two public benchmarks, we evaluate OvisOCR2 on an in-house benchmark designed to cover a broader set of long-tail and challenging scenarios. OvisOCR2 obtains the best overall performance among the compared methods, providing further evidence of its generalization and robustness. OvisOCR2 is available at https://huggingface.co/ATH-MaaS/OvisOCR2.