Mdpbench Eval

mdpbench-eval

MDPBench: A Benchmark for Multilingual Document Parsing in Real-World Scenarios — Zhang Li et al. (2026) (arXiv:2603.28130, 2026)

What this evaluates

MDPBench evaluates the capability of document parsing models to accurately extract text, formulas, tables, and layout structures from multilingual document images under real-world conditions. It specifically probes robustness to photographic degradation, non-Latin scripts, right-to-left reading orders, and cross-lingual generalization without prior language or image-type knowledge.

Datasets

• MDPBench — total 3400; splits: test (3400); repo https://github.com/Yuliang-Liu/MultimodalOCR

Metrics

• accuracy (primary) — range: percent
  • Percentage of correctly parsed documents out of the total evaluated. Higher is better.
• Edit — range: other
  • Edit distance between predicted and ground-truth text strings. Lower is better.
• CDM — range: [0, 1]
  • Formula recognition metric (Character-level Document Match or similar). Higher is better.
• TEDS — range: [0, 1]
  • Tree Edit Distance based Similarity for table structure recognition. Higher is better.
• PageIoU — range: [0, 1]
  • Page Intersection over Union for layout detection. Higher is better.

Input / output format

Input: Document images (digital-born or photographed) without metadata indicating language or capture type.

Output: Parsed document representation including text content, layout bounding boxes, and recognized formulas/tables.

Scoring recipe

def compute_accuracy(predictions, gold):
    correct = sum(1 for p, g in zip(predictions, gold) if p == g)
    return (correct / len(gold)) * 100

# Component metrics use standard implementations:
# Edit: levenshtein_distance(pred_text, gold_text)
# CDM: character-level match score for formulas
# TEDS: tree_edit_distance_similarity(pred_table, gold_table)
# PageIoU: intersection_over_union(pred_layout, gold_layout)

Common pitfalls

• Models are evaluated without prior knowledge of the input language or whether the image is photographed or digital; providing this metadata artificially inflates performance.
• Component-level evaluations (e.g., cropped text blocks) favor models trained on block-level data, while end-to-end full-page evaluations favor models trained on full-document images, leading to inconsistent rankings across settings.
• Right-to-left scripts (e.g., Arabic) are often misparsed due to left-to-right processing assumptions in standard OCR pipelines.

Evidence (verbatim from paper)

As shown in Tab.[2], the top-performing proprietary model, Gemini-3-Pro*[gemini3pro]*, achieves an overall accuracy of 86.4%, reaching state-of-the-art (SOTA) results in 14 of 17 languages. In contrast, the best open-source model, dots.mocr, attains 80.5% overall accuracy, revealing a clear gap between proprietary and open-source approaches.

Citation

@misc{zhang2026mdpbench,
  title={MDPBench: A Benchmark for Multilingual Document Parsing in Real-World Scenarios},
  author={Zhang Li et al. (2026)},
  year={2026},
  note={arXiv:2603.28130}
}

• arXiv: 2603.28130