Why splitting parsing into two stages changes the game
The core idea behind decoupled vision-language models (VLMs) is to stop treating a whole page as one giant image. Instead, they first do a quick, low-resolution scan to find the layout—where the headings, tables, and formulas are—and then zoom into only those specific regions at full resolution for detailed reading. This avoids the quadratic explosion in computational cost that comes from processing every pixel at high resolution, which was a major bottleneck for older models.
Evidence from two independent systems shows this works. MinerU2.5, a 1.2-billion-parameter model, uses this coarse-to-fine strategy and achieves state-of-the-art accuracy on multiple benchmarks while keeping computational overhead low [2]. Similarly, PaddleOCR-VL introduces a 'Valid Region Focus Module' that identifies the important parts of a page and skips the background, allowing a compact 0.9-billion-parameter model to outperform larger, more expensive systems [1]. Both papers converge on the same conclusion: you don't need to process the whole image at full resolution to get accurate results—you just need to know where to look.
Speed and efficiency: the practical payoff
For real-world use, the biggest win is speed. By decoupling layout analysis from content recognition, models can avoid the sequential bottleneck of generating every token for a whole page one by one. A new approach called HPD-Parsing replaces full-page autoregressive generation with a hierarchical parallel decoding scheme, where a main branch handles layout and multiple concurrent branches read different blocks at the same time. This achieves 4,752 tokens per second—2.62 times the throughput of the fastest existing document parsing model and 3.06 times that of a vanilla autoregressive baseline—while maintaining competitive accuracy [6].
Other models also report major speedups. HunyuanOCR-1.5, a lightweight end-to-end OCR model, uses a decoding optimization called DFlash to achieve a 6.37x speedup in transformer inference and 2.14x under vLLM, making it the fastest among lightweight OCR VLMs [5]. These gains matter because they make high-resolution parsing feasible on standard hardware, enabling real-time or near-real-time document processing in applications like digitizing archives, automated form filling, and searchable PDFs.
The catch: real-world photos still trip up models
The biggest challenge isn't clean digital PDFs—it's photographs of documents taken with a phone. A new benchmark called DocPTBench, which includes over 1,300 high-resolution photographed documents, found that moving from digital-born to photographed documents causes a substantial performance drop: popular multimodal LLMs lose an average of 18% accuracy in end-to-end parsing and 12% in translation, while specialized document parsing models drop by an average of 25% [3]. This shows that geometric distortions, lighting changes, and shadows are still major hurdles.
The good news is that researchers are already tackling this. NaviDC-OCR introduces deformation-aware learning to incorporate geometric perception into VLMs, and it achieves state-of-the-art scores on Wild-OmniDocBench (88.53) and PureDocBench (78.41), outperforming other models on camera-captured documents [4]. This suggests that over the next two years, we'll see models that are not only faster and cheaper but also more robust to the messy conditions of real-world use—though the 18-25% gap indicates there's still work to do.
About These Sources
This answer is built on 6 studies (2 peer-reviewed, 4 preprints) — published from 2025 to 2026, 6 from 2024 or later — selected as the most relevant from 8 studies that passed quality screening, drawn from 43 papers retrieved from a database of over 500 million.
Sources used in this answer
Boosting Document Parsing Efficiency and Performance with Coarse-to-Fine Visual Processing
PaddleOCR-VL uses a coarse-to-fine architecture with a Valid Region Focus Module to skip redundant visual regions, achieving state-of-the-art parsing accuracy with a 0.9B model and fewer vision tokens.
MinerU2.5: A Decoupled Vision-Language Model for Efficient High-Resolution Document Parsing
MinerU2.5, a 1.2B-parameter model, decouples global layout analysis from local content recognition, achieving state-of-the-art accuracy on multiple benchmarks with lower computational overhead.
DocPTBench: Benchmarking End-to-End Photographed Document Parsing and Translation
DocPTBench, with over 1,300 photographed documents, shows that MLLMs drop 18% in parsing and 12% in translation accuracy, and specialized models drop 25%, when moving from digital to photographed documents.
NaviDC-OCR: Navigating Document Parsing Across Digital and Camera-Captured Documents
NaviDC-OCR incorporates deformation-aware learning and adaptive sampling, achieving state-of-the-art scores on Wild-OmniDocBench (88.53) and PureDocBench (78.41), and ranking first in the ICDAR 2026 Sci-ImageMiner Challenge.
HunyuanOCR-1.5: Making Lightweight OCR VLMs Faster and Better
HunyuanOCR-1.5 uses DFlash to achieve a 6.37x transformer inference speedup and 2.14x under vLLM, making it the fastest lightweight OCR VLM while improving long-tail capabilities.
HPD-Parsing: Hierarchical Parallel Document Parsing
HPD-Parsing uses hierarchical parallel decoding to achieve 4,752 tokens per second, 2.62x the throughput of the fastest existing model and 3.06x of a vanilla autoregressive baseline, with competitive accuracy.