Evolution and Challenges in Document Parsing Technology

Document parsing sits at the intersection of computer vision and natural language processing, facing two fundamental challenges: balancing model accuracy with computational efficiency, and maintaining robust recognition across complex layouts and diverse text formats. Traditional methods employing sequential “detection + recognition” pipelines suffer from error propagation and contextual information loss.

While emerging vision-language models (VLMs) enable joint optimization of layout analysis and content recognition through unified architectures, large-parameter models like Qwen3-VL (235B parameters) face significant deployment barriers due to computational costs and latency issues.

The MinerU2.5-Pro model developed by a research team achieves breakthrough performance by combining an innovative two-stage parsing strategy with meticulous data engineering. With only 1.2B parameters, it scores 95.69 on the OmniDocBench v1.6 benchmark - a 2.71-point improvement over its predecessor and surpassing models 200x larger in parameter count. This validates the technical feasibility of prioritizing data quality over model scale.

Two-Stage Parsing Architecture

Unified Vision-Language Model Design

The model employs a modified Transformer architecture from Qwen2VL, comprising:

• Visual Encoder: Hierarchical feature extraction through 4 downsampling stages, generating 1/32 resolution feature maps while preserving spatial relationships
• Language Encoder: Supports 4096-token context modeling with optimized attention efficiency for long documents
• Cross-Modal Attention: Dynamic routing mechanism that automatically adjusts visual-language feature fusion ratios based on document complexity (70% visual for simple tables, 60% language for complex academic papers)

Two-Stage Parsing Workflow

Stage 1: Global Layout Analysis
Prompt engineering guides the model to detect semantic regions with structured JSON output:

{
  "regions": [
    {
      "type": "title",
      "bbox": [0.1, 0.05, 0.9, 0.15],
      "confidence": 0.98
    },
    {
      "type": "table",
      "bbox": [0.2, 0.2, 0.8, 0.5],
      "confidence": 0.95
    }
  ]
}

Stage 2: Local Content Recognition
Region-specific prompts generate structured outputs (e.g., Markdown tables for tabular data):

| Product | Sales | Growth |
|---------|-------|--------|
| A       | 1200  | 15%    |
| B       | 980   | 8%     |

This decoupled approach reduces average computational complexity from O(n) to O(k) (k = detected regions), achieving 3.2x faster inference while maintaining >98% region detection recall.

Data Engineering Breakthroughs

Three-Dimensional Data Quality Framework

The team developed a systematic data construction methodology:

1. Coverage Optimization: 217 document types collected (contracts, invoices, academic papers) with ≥5,000 samples per category
2. Diversity Enhancement: 12 data augmentation techniques including font substitution (300+ families), layout perturbation (0.8-1.2x scaling), and noise injection
3. Accuracy Assurance: Four-round human validation with 99.2% final consistency

Dynamic Hard Sample Mining

The algorithm identifies error patterns in low-accuracy samples (<85%) and generates targeted augmentations:

• Small text: 5,000 samples at 6-8pt font sizes
• Complex tables: 3,000 variants with 1-3干扰 lines (translated as “distraction lines”)
  This approach improved Hard subset performance by 5.3 points, closing 78% of the accuracy gap with large models.

Performance Benchmarking

Comprehensive Evaluation Results

On OmniDocBench v1.6 end-to-end testing:
| Model | Parameters | Score | Speed (pages/sec) |
|——————————-|—————-|———-|—————————-|
| MinerU2.5-Pro | 1.2B | 95.69 | 12.5 |
| GLM-OCR | 13B | 95.15 | 8.2 |
| PaddleOCR-VL-1.5 | 7B | 94.87 | 9.1 |
| Gemini 3 Pro | 256B | 92.85 | 3.7 |
| Qwen3-VL-235B | 235B | 89.78 | 2.1 |

The model demonstrated particular strength in complex documents, leading the Hard subset by 2.07 points with a 94.08 score.

Technical Insights

Key findings include:

1. Model efficiency: 1.2B parameters can match/exceed 200B models through data optimization
2. Task decoupling: Two-stage design improves efficiency by 40% over single-stage models
3. Data impact: Systematic data engineering contributes 15-20% performance gains vs. architectural changes

Deployment Recommendations

Typical Use Cases

1. Financial sector: Automated extraction of key terms and values from contracts/reports
2. Healthcare: Structured database construction from medical records and test results
3. Legal domain: Evidence classification and relationship mapping in litigation materials
4. Education: Automated grading of open-ended exam questions

Edge Device Optimization

For resource-constrained environments:

1. Model quantization: INT8 conversion reduces model size by 75% and speeds up inference by 2.8x
2. Dynamic batching: Adjusts batch sizes based on document complexity
3. Region cropping: Processes large documents in segments to reduce visual tokens

On NVIDIA Jetson AGX Orin, the optimized model achieves 8.2 pages/second processing capability.

Future Development Directions

The research opens new avenues including:

1. Multilingual expansion: Coverage of 100+ languages for cross-border document processing
2. Real-time interaction: Incremental inference with user correction capabilities
3. 3D document processing: Content recognition in scanned 3D document models
4. Privacy preservation: Federated learning frameworks for secure model training

This work demonstrates the viability of combining algorithmic optimization with data engineering to enable lightweight models for complex cognitive tasks, providing a reference architecture for next-generation document parsing technologies.