How PaddleOCR‑VL‑1.6’s 0.9B Model Achieved 96.33% SOTA on OmniDocBench v1.6

Document Parsing and Model Overview

Document parsing converts scanned pages into machine‑readable structures (text, tables, formulas, charts, seals, reading order, layout). A compact 0.9 B‑parameter model, PaddleOCR‑VL‑1.6, achieved a 96.33 % overall score on OmniDocBench v1.6, leading the leaderboard.

Diagnosing Weak Points

Analysis of the predecessor PaddleOCR‑VL‑1.5 (94.93 % on OmniDocBench v1.5) showed that remaining errors cluster in three regions:

Boundary‑Fragile Regions : Minor visual perturbations (pixel shift, JPEG compression, slight blur) cause large output changes, indicating unstable decision boundaries.

Coverage‑Sparse Regions : Samples that appear in the training set are still mispredicted because the surrounding data distribution is sparse, causing long‑tail patterns to be overwhelmed by dominant distributions.

Unreliable‑Supervision Regions : High‑confidence errors arise from incorrect labels. Three external expert models (Qianfan‑OCR, GLM‑OCR, MinerU2.5‑Pro) are used to cross‑validate and correct these labels.

Targeted high‑value annotations were added for each region.

Three‑Step Training Pipeline

The optimization framework combines a model‑driven data engine with a progressive post‑training strategy:

CPT (Continued Pre‑Training) : 16.8 M samples, including long‑tail documents (ancient books, rare characters, industrial tables) and corrected annotations, broaden the model’s coverage.

SFT (Supervised Fine‑Tuning) : 7.3 M hard samples are selected via Uncertainty‑Aware Cluster Sampling (UACS), expert‑disagreement cases, and corrected unreliable‑supervision samples to sharpen performance on fragile regions.

RL (Reinforcement Learning) : 49 K samples are processed with GRPO. A high‑potential sample mining strategy evaluates candidates on learning potential, uncertainty, and reward variance, ensuring only informative samples influence the 0.9 B model.

Reward functions map complex document‑parsing outputs to verifiable signals, incorporating legality, structural correction constraints, and real‑score dimensions.

Model Architecture

The system consists of two models: PP‑DocLayout V3 for layout analysis and PaddleOCR‑VL‑1.6‑0.9B for visual‑language understanding. The architecture mirrors PaddleOCR‑VL‑1.5, featuring a Native Resolution Visual Encoder, Adaptive MLP Connector, and ERNIE‑4.5‑0.3B language model. No architectural changes or parameter increases were made; gains stem solely from smarter data strategies and refined training.

Benchmark Results

OmniDocBench v1.6 adds Multi‑Granularity Adaptive Matching (MGAM) and a 296‑page Hard subset covering nested tables, dense formulas, and unconventional structures. PaddleOCR‑VL‑1.6 achieved:

Total score: 96.33 % (↑1.4 pts vs. 1.5)

Text edit distance: 0.033

Formula CDM: 97.49 %

Table TEDS: 94.76 %

Table‑structure TEDS: 97.11 %

Reading‑order score: 0.127

On Real5‑OmniDocBench (simulating real‑world captures with scanning, bending, phone photos, lighting changes, and tilt), PaddleOCR‑VL‑1.6 scored 93.19 % (↑1.14 pts), outranking 235 B Qwen3‑VL, 241 B InternVL3.5, 1 T KimiK2.5, and GPT‑5.2.

Task‑Specific Gains

Hard table recognition (1 258 samples, 20 table types): TEDS 91.71, structure TEDS 94.67 (≈+2 pts over MinerU2.5‑Pro).

Chart parsing (1 801 samples, 11 chart types): RMS‑F1 91.74, Chinese chart F1 93.37 (↑11 pts over previous generation).

Text localization (9 dimensions): total 87.47, with improvements on ancient scripts, Japanese, and handwritten Chinese.

Seal recognition: NED 0.119, far better than Qwen3‑VL‑235B’s 0.382.

The methodology—diagnosing weak zones, precisely augmenting data, and applying progressive CPT‑SFT‑RL training—demonstrates a viable path for continuously improving performance of compact models, enabling low‑cost deployment on edge devices for document digitization, invoice processing, and archive management.

References

https://huggingface.co/PaddlePaddle/PaddleOCR-VL-1.6

https://modelscope.cn/models/PaddlePaddle/PaddleOCR-VL-1.6

https://github.com/PaddlePaddle/PaddleOCR

https://arxiv.org/pdf/2606.03264