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

[Dixon](https://medium.com/@huguosuo)


Follow

4 min read

·

Oct 5, 2025

1

How a two-stage, coarse-to-fine approach redefines document parsing for speed, accuracy, and scalability

**Summary**

MinerU2.5 introduces a breakthrough in high-resolution document understanding by decoupling global layout analysis from local content recognition, enabling both high accuracy and exceptional computational efficiency

2509.22186v2

. Built with only 1.2 billion parameters, the model achieves state-of-the-art results on the OmniDocBench benchmark — outperforming large general-purpose models such as GPT-4o, Gemini-2.5 Pro, and Qwen2.5-VL-72B across key tasks like text, table, and formula recognition.

Unlike monolithic vision-language models (VLMs) that directly process massive, high-resolution pages — incurring high memory and latency costs — MinerU2.5 uses a two-stage “coarse-to-fine” pipeline. The first stage analyzes a downsampled page to detect global layout and structure; the second stage zooms in on cropped, high-resolution regions to perform fine-grained parsing. This separation reduces computational load by over an order of magnitude while maintaining pixel-level precision in complex regions

2509.22186v2

The model’s success also relies on a comprehensive data engine, which systematically curates and refines massive document datasets through iterative mining and human-in-the-loop annotation, ensuring robust performance across diverse formats, languages, and layouts.

**💡 Intuition**

Most vision-language models struggle with documents because they treat every pixel as equally important. MinerU2.5 “looks smartly” — first getting the big picture at low resolution, then focusing attention only where detail matters. It’s like reading a page by first scanning its layout and then zooming in to understand formulas, tables, and text precisely. This strategy yields high efficiency without sacrificing fidelity, making it practical for real-world use.

**🎯 Problem**

Document parsing — especially for scientific papers, invoices, and complex PDFs — poses three core challenges:

1. High Resolution & Dense Layouts — Documents contain large, intricate regions with text, figures, and tables at fine granularity.

2. Inefficient Processing — End-to-end VLMs suffer from quadratic token complexity (O(N²)), leading to huge computational costs.

3. Hallucination & Inconsistency — Models often misread or “hallucinate” content when interpreting dense, multi-column text or math-heavy pages.

Existing OCR and layout systems either rely on modular pipelines that propagate errors across stages or on monolithic VLMs that are accurate but computationally prohibitive

2509.22186v2

**🛠️ Solution**

Press enter or click to view image in full size


MinerU2.5 addresses these limitations with a decoupled, two-stage vision-language framework and a closed-loop data engine that fuels continual improvement.

# **1. Two-Stage Parsing Strategy**

• **Stage I — Layout Analysis**:Performs rapid structural parsing on downsampled (1036×1036 px) images, identifying elements such as text blocks, figures, tables, and formulas while minimizing computation.
• **Stage II — Content Recognition**:Uses the detected layout to crop high-resolution patches from the original page, enabling fine-grained recognition (text OCR, formula-to-LaTeX conversion, table-to-HTML generation) at native resolution
• 2509.22186v2

# **2. Model Architecture**

• Vision Encoder: **NaViT-based native-resolution encoder** (675M parameters) with dynamic 2D positional embeddings.
• Language Model: **Qwen2-Instruct-0.5B**, integrated via a lightweight patch-merging adaptor.
• The result: an efficient 1.2B-parameter VLM that balances speed and performance better than any prior document parser.

# **3. Data Engine for Robustness**

MinerU2.5’s performance is reinforced by a multi-stage data pipeline that builds diverse, high-quality training data:

• **Stage 1:** Automatic large-scale pretraining on curated document datasets with refined annotations.
• **Stage 2:** **Iterative Mining via Inference Consistency (IMIC)** identifies “hard cases” (low-consistency predictions) for expert correction, focusing annotation efforts where the model struggles most.
• **Stage 3:** Final fine-tuning with these curated, challenging samples, improving robustness in edge cases (rotated tables, long formulas, mixed-language pages).

# **4. Task-Specific Innovations**

• **Layout Analysis:** Introduces *PageIoU*, a perceptually aligned metric that measures page-level spatial consistency better than standard IoU.
• **Formula Parsing:** Proposes the *Atomic Decomposition and Recombination (ADR)* framework — splitting long equations into atomic parts for LaTeX reconstruction.
• **Table Recognition:** Uses *Optimized Table Structure Language (OTSL)* to cut redundant tokens and halve sequence length while improving accuracy.

Together, these advances make MinerU2.5 a general-purpose OCR engine capable of outperforming specialized and general models alike across document types.

**🚧 Limitations and Future Opportunities**

While MinerU2.5 is significantly faster and more efficient than prior approaches, it still faces challenges:

• **Inference Cost:** The two-stage pipeline introduces additional latency compared to single-pass models in small-document cases.
• **Model Size vs. Accessibility:** Despite being lightweight for a VLM, the 1.2B parameter scale can be heavy for on-device or edge deployments.
• **Limited Multimodality:** Current design focuses on text-centric documents; extending to charts or multi-page reasoning remains future work.

The authors suggest potential future directions such as:

• Integrating **multimodal reasoning** (text + diagrams + handwriting).
• Developing **adaptive-resolution attention mechanisms** for dynamic token selection.
• Creating **open data benchmarks** that better capture real-world document diversity.

In short, MinerU2.5 redefines document parsing by thinking like a human reader — first scanning, then focusing. Its decoupled architecture, enriched data engine, and clever efficiency optimizations bring industrial-grade parsing to accessible scales, marking a new frontier for vision-language models in structured document understanding.