MinerU2.5-Pro: Pushing the Limits of Data-Centric Document Parsing at Scale
| Authors | Bin Wang et al. |
| Year | 2026 |
| HF Upvotes | 115 |
| arXiv | 2604.04771 |
| Download | |
| HF Page | View on Hugging Face |
Abstract
Current document parsing methods compete primarily on model architecture innovation, while systematic engineering of training data remains underexplored. Yet SOTA models of different architectures and parameter scales exhibit highly consistent failure patterns on the same set of hard samples, suggesting that the performance bottleneck stems from shared deficiencies in training data rather than architecture itself. Building on this finding, we present \minerupro, which advances the state of the art solely through data engineering and training strategy optimization while keeping the 1.2B-parameter architecture of \mineru completely fixed. At its core is a Data Engine co-designed around coverage, informativeness, and annotation accuracy: Diversity-and-Difficulty-Aware Sampling expands training data from under 10M to 65.5M samples while correcting distribution shift; Cross-Model Consistency Verification leverages output agreement among heterogeneous models to assess sample difficulty and generate reliable annotations; the Judge-and-Refine pipeline improves annotation quality for hard samples through render-then-verify iterative correction. A three-stage progressive training strategy -- large-scale pre-training, hard sample fine-tuning, and GRPO alignment -- sequentially exploits these data at different quality tiers. On the evaluation front, we fix element-matching biases in OmniDocBenchv1.5 and introduce a Hard subset, establishing the more discriminative OmniDocBenchv1.6 protocol. Without any architectural modification, \minerupro achieves 95.69 on OmniDocBench~v1.6, improving over the same-architecture baseline by 2.71 points and surpassing all existing methods including models with over 200times more parameters.
Engineering Breakdown
Plain English
MinerU2.5-Pro demonstrates that document parsing performance is bottlenecked by training data quality, not model architecture. The authors kept the 1.2B-parameter architecture from the original MinerU completely fixed and improved results solely through data engineering—specifically through a Data Engine that optimizes for coverage, informativeness, and annotation accuracy, using techniques like Diversity-and-Difficulty-Aware Sampling. This finding challenges the field's focus on architecture innovation and suggests that SOTA models across different architectures fail on the same hard samples because they share deficiencies in training data rather than fundamental design limitations. The work achieves state-of-the-art document parsing performance while proving that systematic data-centric engineering can outpace architectural changes.
Core Technical Contribution
The core insight is that training data engineering, not model architecture, is the primary lever for improving document parsing at scale. Rather than introducing a new model, the authors designed a Data Engine with three pillars: coverage (breadth of document types and scenarios), informativeness (quality and utility of samples), and annotation accuracy (correctness of labels). The key technical contribution is Diversity-and-Difficulty-Aware Sampling, which strategically selects hard samples during training to maximize learning signal without catastrophic forgetting. This inverts the conventional wisdom in deep learning, which assumes that architectural innovations (attention mechanisms, transformer variants, etc.) drive performance gains—here, the fixed architecture + optimized data beats all architectural alternatives tested.
How It Works
The Data Engine operates as a closed-loop system that iterates between data collection, curation, and training. First, raw documents are ingested and parsed using the base 1.2B-parameter model; difficult or ambiguous samples are flagged based on model confidence and consistency metrics. Second, Diversity-and-Difficulty-Aware Sampling prioritizes these hard samples for human annotation or refinement, ensuring the training set contains representative coverage of document types while concentrating annotation effort where the model struggles most. Third, annotators correct parse results (layouts, bounding boxes, text extraction, tables, equations) with high accuracy standards. Finally, the curated dataset trains the fixed 1.2B-parameter model, and the process repeats—the model retrains on the expanded, higher-quality dataset and generates new hard samples. This feedback loop allows the Data Engine to progressively eliminate blind spots in the training distribution without changing the underlying architecture or parameter count.
Production Impact
For engineers building document parsing systems, this work reveals that 80% of effort should focus on training data quality and diversity rather than model experimentation. In a production pipeline, this means: (1) investing in a data annotation infrastructure and curation workflow rather than chasing larger models or novel architectures; (2) using confidence-based hard sample mining to guide where humans should annotate, reducing wasted annotation effort; (3) expecting that a smaller, well-trained model outperforms a large model trained on messy data—enabling deployment on edge devices or constrained infrastructure. The trade-off is real: systematic data curation requires significant human annotation work upfront (days to weeks per new document class), but the resulting model is robust, smaller, faster, and cheaper to serve at scale. For companies processing millions of documents (invoices, contracts, scientific papers), this approach likely reduces both compute costs and error rates compared to fine-tuning larger models on raw data.
Limitations and When Not to Use This
The paper assumes human annotators can define and enforce consistent, high-quality labels—this breaks down in domains with inherent ambiguity (handwritten documents, severely degraded scans, multilingual mixed-language pages). The Diversity-and-Difficulty-Aware Sampling strategy requires a trained baseline model to identify hard samples, creating a chicken-and-egg problem for brand-new document types with no prior model; in such cases, a more naive or architectural approach may be necessary initially. The fixed 1.2B architecture may be a ceiling—it's unclear whether this data-centric approach generalizes to much smaller models (500M parameters) or whether architectural improvements would help when combined with the Data Engine. The paper does not address real-time or online learning scenarios where data distribution shifts rapidly, nor does it discuss how to allocate annotation budget across competing document types in truly multi-domain settings.
Research Context
This work builds on the growing data-centric AI movement (popularized by works on data quality in computer vision and NLP) but applies it rigorously to structured document understanding, a space traditionally dominated by architectural arms races (Vision Transformers, multimodal models, etc.). It advances parsing benchmarks like DocVQA, PubLayNet, and real-world document corpora by showing that state-of-the-art results can be achieved without scale-up in parameters. The paper opens a research direction questioning whether the field's emphasis on model scaling and architectural novelty is misplaced—echoing critiques in other areas that suggest scaling laws plateau when training data is not optimized. Future work likely includes: understanding how to automatically generate synthetic hard samples, applying meta-learning to predict which annotation strategies generalize best, and extending the Data Engine to multimodal inputs (scanned documents + metadata) and real-time feedback loops.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::