Research map / Vision-language and multimodal / Multimodal LLMs

Boosting Document Parsing Efficiency and Performance with Coarse-to-Fine Visual Processing

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Interleaved image-text generation with MLLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes PaddleOCR-VL, a coarse-to-fine document parser where a lightweight layout/reading-order module (VRFM, RT-DETR plus pointer network) crops valid regions that are then recognized by a 0.9B VLM, reaching SOTA on OmniDocBench v1.5 with fewer vision tokens.

Problem

High-resolution document parsing with VLMs yields quadratically many vision tokens because large parts of document pages are redundant background, raising compute and latency, while uniform token compression hurts fine-grained layout accuracy.

Approach

Decouple layout analysis from recognition: a lightweight Valid Region Focus Module detects layout elements and predicts reading order, then only the cropped valid regions are fed to a compact native-resolution VLM (PaddleOCR-VL-0.9B), avoiding processing redundant background.

Key results

• PaddleOCR-VL-L reaches an overall score of 92.62 on OmniDocBench v1.5 with 2561 vision tokens, versus 90.67 for MinerU2.5 with 3256 tokens.
• PaddleOCR-VL-L achieves 1.62 pages/s and 2470.7 tokens/s on A100 with vLLM, about 53.1% higher page throughput than MinerU2.5.
• Formula-block CDM of 0.9404 versus 0.9187 for MinerU2.5; in-house chart RMS-F1 of 0.8440 versus 0.7300 for Qwen2.5-VL-72B.

Surprising findings

• DocKylin-style region-cropping token reduction on Qwen3-VL-8B lowered accuracy by about 1.85 points, suggesting naive token pruning hurts parsing.
• PaddleOCR-VL-0.9B with VRFM uses fewer tokens than a 72B-scale general VLM while outperforming it on OmniDocBench v1.5.

Limitations

• Training data is largely in-house or synthetic and not released, limiting reproducibility of the data pipeline.
• Chart recognition is evaluated only on an in-house benchmark, and the public test sets were judged to have poor annotation quality.
• Reported efficiency is from specific hardware and the vLLM backend; FastDeploy integration is still in progress.

Experimental setup

Datasets: OmniDocBench v1.5 CASIA-HWDB UniMER-1M MathWriting ChartQA PlotQA Unichart Chart2Text DVQA Beagle ChartINFO visText

Benchmarks: OmniDocBench v1.5 OmniDocBench-OCR-block OmniDocBench-Formula-block OmniDocBench-Table-block In-house chart benchmark Multilingual OCR benchmark

Metrics: Overall score Text Edit distance Formula CDM Table TEDS Table TEDS-S Reading Order Edit Edit distance RMS-F1 Pages/s Tokens/s

Baselines: MinerU2.5 dots.ocr MonkeyOCR-pro-3B Qwen2.5-VL-72B Gemini-2.5 Pro GPT-4o DeepSeek-OCR MinerU2-VLM Dolphin PP-StructureV3

Builds on: RT-DETR LLaVA PP-StructureV3 PP-DocLayout_Plus-L ERNIE-4.5-0.3B

Closest papers by meaning

• Towards Real-World Document Parsing via Realistic Scene Synthesis and Document-Aware Training (CVPR 2026 · similarity 0.87)
• Read, Parse, Describe: Unified Document Parsing with Visual Element Description Generation (NeurIPS 2026 · similarity 0.87)
• LensVLM: Selective Context Expansion for Compressed Visual Representation of Text (NeurIPS 2026 · similarity 0.87)
• DocFormerv2: Local Features for Document Understanding (AAAI 2024 · similarity 0.86)
• DocKylin: A Large Multimodal Model for Visual Document Understanding with Efficient Visual Slimming (AAAI 2025 · similarity 0.86)
• PP-OCRv5: A Specialized 5M-Parameter Model Rivaling Billion-Parameter Vision-Language Models on OCR Tasks (CVPR 2026 · similarity 0.86)