PaperlayerView exact revision on arXiv

Source-linked AI summary

PaddleOCR-VL-1.6: Expanding the Frontier of Document Parsing with Under-Optimized Region Refinement and Progressive Post-Training

Zelun Zhang, Hongen Liu, Suyin Liang, Yubo Zhang, Yiqing Xiang, Jiaxuan Liu, Ting Sun, Manhui Lin, Yue Zhang, Changda Zhou, Tingquan Gao, Cheng Cui, Yi Liu, Dianhai Yu, Yanjun Ma

arXiv:2606.03264v1cs.CV

TL;DR

Document parsing errors remain concentrated in difficult regions with sparse coverage or unreliable supervision. PaddleOCR-VL-1.6 targets these regions with optimized data and progressive post-training, achieving 96.33% on OmniDocBench v1.6.

  • Problem

    Document parsing errors remain concentrated in difficult regions with sparse coverage, unstable behavior, or unreliable supervision despite strong existing models.

  • Method

    PaddleOCR-VL-1.6 mines under-optimized regions, enhances their data and supervision, and applies progressive CPT, SFT, and RL post-training.

  • Results

    96.33% on OmniDocBench v1.6 establishes state-of-the-art overall performance for PaddleOCR-VL-1.6.

  • Takeaways & Limitations

    Targeted data optimization and staged post-training unlock further potential in compact document VLMs without indiscriminate model scaling.

Abstract

from arXiv · show

We introduce PaddleOCR-VL-1.6, an upgraded compact document parsing model built upon PaddleOCR-VL-1.5. Although PaddleOCR-VL-1.5 establishes a strong 0.9B baseline, its remaining errors concentrate in under-optimized regions where model behavior is unstable, data coverage is sparse, or supervision is unreliable. Rather than expanding the training corpus indiscriminately, PaddleOCR-VL-1.6 introduces a region-aware data optimization framework that identifies weak regions from the previous model, applies targeted enhancement to these regions, and improves the reliability of supervision signals. It further adopts a progressive post-training recipe based on curated data selection and reinforcement learning, pushing model performance to a higher level through staged optimization. PaddleOCR-VL-1.6 achieves a new state-of-the-art score of 96.33% on OmniDocBench v1.6, demonstrates strong competitiveness against top-tier VLMs, and provides a practical post-training recipe for the PaddleOCR-VL series.

1. Introduction

PaddleOCR-VL-1.6 targets residual document-parsing errors concentrated in unstable, sparsely covered, and unreliably supervised regions. It combines Under-Optimized Region Mining, high-precision annotation, and progressive post-training while preserving a compact 0.9B scale and achieving state-of-the-art performance on OmniDocBench v1.6.

  • Problem: Residual errors concentrate in long-tail layouts, rare scripts, dense formulas, complex tables, and noisy supervision rather than uniformly distributed noise.These difficult regions remain underrepresented or unreliable even when the overall training corpus is large.
  • Method: Under-Optimized Region Mining diagnoses boundary-fragile, coverage-sparse, and unreliable-supervision regions specific to PaddleOCR-VL-1.5.Boundary-fragile regions show unstable predictions, while coverage-sparse regions occupy low-density neighborhoods in feature semantic space.
  • Post-training: Progressive post-training uses curated data in successive Continued Pre-training and Supervised Fine-Tuning stages instead of one mixed training stage.Continued Pre-training broadens distributional coverage and corrects supervision; Supervised Fine-Tuning emphasizes high-difficulty, high-quality samples.
  • Contributions: 0.9B is the preserved compact model scale of PaddleOCR-VL-1.6, which achieves state-of-the-art performance on OmniDocBench v1.6.The upgrade is built upon improved data strategies and a refined post-training pipeline.

2. PaddleOCR-VL-1.6 Overview

PaddleOCR-VL-1.6 retains the compact PaddleOCR-VL system and lightweight 0.9B architecture while focusing improvements on targeted data optimization and progressive post-training. Its overview combines under-optimized-region diagnosis, expert-guided refinement, and a CPT-SFT-RL training pipeline.

  • System design: The system comprises unchanged PP-DocLayoutV3 for layout analysis and PaddleOCR-VL-1.6-0.9B for vision-language understanding.The upgrade focuses on improving PaddleOCR-VL-1.6-0.9B rather than modifying PP-DocLayoutV3.
  • System design: PaddleOCR-VL-1.6-0.9B preserves the lightweight architecture of PaddleOCR-VL-1.5-0.9B, using a Native Resolution Visual Encoder, Adaptive MLP Connector, and ERNIE-4.5-0.3B Language Model.Its main upgrade is a targeted data engine and refined post-training process, not model enlargement or architectural modification.
  • Practical tasks: The model supports document parsing and text spotting, with document parsing using PP-DocLayoutV3 for high-precision localization followed by PaddleOCR-VL-1.6-0.9B understanding.Text spotting directly uses PaddleOCR-VL-1.6-0.9B for end-to-end detection and recognition across documents, identification cards, manuscripts, posters, screenshots, signboards, and multilingual text images.
  • Improvement framework: The data engine diagnoses fragile predictions, sparse coverage, and unreliable labels to retrieve new unlabeled samples and identify annotations needing correction.It specifically targets boundary-fragile, coverage-sparse, and external-support-deficient regions.
  • Improvement framework: Retrieved samples receive expert-consensus labeling and iterative Judge-and-Refine processing, followed by progressive post-training through a CPT-SFT-RL pipeline and reusable high-value data selection.Hard cases with insufficient expert consensus are further refined before the progressive post-training stages.

3. Under-Optimized Region Driven Data Engine

PaddleOCR-VL-1.6 targets Under-Optimized Regions where PaddleOCR-VL-1.5 lacks reliable image-to-structured-output mappings, rather than indiscriminately expanding generic training data. Its data engine mines unstable, coverage-sparse, and unreliable-supervision cases for targeted refinement and reliable labeling.

  • Model-oriented data optimization: The data engine focuses on Under-Optimized Regions where the current model has not learned a reliable mapping from document images to structured outputs.These regions are motivated by unstable predictions, sparse distributional support, and unreliable supervision identified in PaddleOCR-VL-1.5.
  • Boundary-Fragile Regions: 128 predictions per sample produce 8128 pairwise discrepancy scores, from which the largest 128 distances are averaged into the Boundary-Fragility Score.The final selection takes the top 1% of ranked samples and adds samples exhibiting model degeneration in any prediction.
  • Boundary-Fragile Regions: Boundary-Fragile Region mining combines late-checkpoint prediction instability with sensitivity to semantic-invariant visual perturbations.The method retains eight checkpoints from the last 8% of training and evaluates perturbations including pixel shifts, JPEG compression, noise, blur, and non-uniform scaling.
  • Coverage-Sparse Regions: Coverage-Sparse Region mining uses document-specific feature representations and progressively tightened similarity thresholds to expose small, weakly connected outlier neighborhoods.This density-oriented strategy preserves rare document modes that fixed-K clustering can absorb into nearby dense groups.
  • Unreliable-Supervision Regions: Expert agreement yields high-confidence corrected data, while unresolved cases are treated as difficult examples or forwarded to manual annotation as pre-annotations.This process mines and improves Unreliable-Supervision Regions inherited from PaddleOCR-VL-1.5 while reducing human effort.

4. Progressive Post-Training Recipe

PaddleOCR-VL-1.6 progressively post-trains from the PaddleOCR-VL-1.5 checkpoint through continued pre-training, selective supervised fine-tuning, and reinforcement learning. The recipe expands coverage, concentrates learning on difficult reliable samples, and selects high-potential RL data with task-balanced, verifiable rewards.

  • Progressive recipe: PaddleOCR-VL-1.6 starts from the PaddleOCR-VL-1.5 checkpoint and allocates Under-Optimized Region Driven Data Engine data across three progressive post-training stages.The recipe prioritizes efficiently absorbing high-value data instead of restarting large-scale pre-training.
  • Continued Pre-Training: 16.8M training samples form the continued-pretraining corpus by combining PaddleOCR-VL-1.5 data with all newly retrieved samples and latest annotations.The corpus broadens coverage while improving supervision quality across previously under-optimized regions and long-tail document scenarios.
  • Supervised Fine-Tuning: SFT concentrates supervised learning on difficult samples mined by UACS, expert-disagreement cases, and other cases requiring reliable labels.This stage refines task behavior rather than reusing all curated data uniformly.
  • Reinforcement Learning: RL regularizes output styles, improves generalization, and suppresses out-of-distribution degeneration, but compact-model sensitivity makes careful data selection essential.The 0.3B language-model component makes PaddleOCR-VL-1.6-0.9B especially vulnerable to casual RL sampling and poor sample efficiency.
  • Reinforcement Learning: The GRPO-oriented mining strategy generates 16 rollouts per candidate, filters overly difficult and overly easy samples, and scores the remainder using reward headroom, uncertainty, and reward diversity.The learning-potential gap is defined as rmax(x) − rmean(x), while task-balanced top-ranked samples are selected across OCR, chart parsing, table recognition, formula recognition, seal recognition, and text spotting.
  • Reinforcement Learning: A representation-aware verifiable reward maps outputs and references into task-specific canonical representations to provide task-aligned scalar feedback while enforcing strict validity constraints.The design addresses the limited learning signal from overly sparse binary rewards in the compact model.

5. Evaluation

PaddleOCR-VL-1.6 achieves state-of-the-art document parsing results on OmniDocBench v1.6 and Real5-OmniDocBench, while also leading evaluations of hard tables, charts, text spotting, and seals. Progressive post-training improves OmniDocBench v1.6 from 94.93% to 96.33%, with CPT and SFT providing the largest gains and RL adding a final improvement.

  • OmniDocBench v1.6: 96.33% overall on OmniDocBench v1.6 establishes PaddleOCR-VL-1.6 as state of the art, improving over PaddleOCR-VL-1.5’s 94.93%.It also improves Text-Edit distance, CDM Score, Table-TEDS, and Table-TEDS-Struct by 0.5%, 0.6%, 3.09%, and 2.74%, respectively.
  • Real5-OmniDocBench: 93.19% overall on Real5-OmniDocBench is a new state-of-the-art result, improving 1.14 points over PaddleOCR-VL-1.5’s 92.05%.Despite its compact 0.9B parameter scale, the model outperforms substantially larger general-purpose VLMs, including Qwen3-VL-235B and Gemini-3 Pro.
  • Hard table recognition: 91.71 Overall TEDS and 94.67 Structural TEDS are the highest scores on the In-house-Table benchmark.The benchmark contains 1,258 challenging table samples spanning 20 categories and varied formats and conditions.
  • Chart parsing: 91.74 overall RMS-F1, including 90.11 on English charts and 93.37 on Chinese charts, gives PaddleOCR-VL-1.6 the strongest In-house-Chart performance.The evaluation covers 1,801 manually reviewed samples across 11 chart categories.
  • Text spotting: PaddleOCR-VL-1.6 achieves the highest spotting accuracy across all 9 evaluated text-spotting dimensions, demonstrating generalization across diverse visual conditions and layouts.The benchmark covers common scenes, Japanese text, degraded images, handwriting, tables, ancient documents, and Traditional Chinese.
  • Seal recognition: 0.119 NED on seal recognition substantially outperforms Qwen3-VL’s 0.382 despite PaddleOCR-VL-1.6’s compact 0.9B scale versus 235B parameters.The benchmark contains 300 images featuring diverse seal shapes and challenging conditions such as overlapping text and low-contrast impressions.
  • Progressive post-training: 0.69 points from CPT and SFT provide the largest progressive post-training gains, while RL raises OmniDocBench v1.6 from 96.25% to 96.33%.RL also increases Formula-CDM from 97.37% to 97.49% after CPT and SFT leave less optimization headroom.

6. Conclusion

PaddleOCR-VL-1.6 preserves the efficient 0.9B architecture of PaddleOCR-VL-1.5 while improving performance through under-optimized-region-driven data optimization and progressive post-training. It achieves state-of-the-art OmniDocBench v1.6 performance, robustness on Real5-OmniDocBench, and gains across key document-understanding capabilities.

  • Model and training strategy: PaddleOCR-VL-1.6 builds on PaddleOCR-VL-1.5 while preserving its efficient 0.9B architecture.The model avoids indiscriminate scaling and instead targets remaining weaknesses.
  • Model and training strategy: Its data engine targets under-optimized regions, and its progressive post-training pipeline covers CPT, SFT, and RL.The approach combines targeted data optimization with staged post-training.
  • Results: The resulting model achieves state-of-the-art performance on OmniDocBench v1.6 and strong robustness on Real5-OmniDocBench.These benchmarks summarize the model’s overall performance and robustness.
  • Results: It delivers consistent gains in hard table recognition, chart parsing, text spotting, and seal recognition.The improvements span several key document-parsing sub-capabilities.
Loading 2606.03264v1…