LongDocBench - Page-Level Parsing: leaderboard
LongDocBench (Ant Group and Zhejiang University, 2026) evaluates document-level structure recovery in 85 long financial reports, textbooks and academic papers. Page-level document parsing under the OmniDocBench v1.6 protocol on all 2,582 pages; overall score in percent, higher is better.
Metric: OmniDocBench v1.6 overall score (%; mean of 100 x (1 - text edit distance), formula CDM and table TEDS) of end-to-end document parsing on all 2,582 LongDocBench pages. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | MinerU2.5-Pro | 95.96 | |
| 2 | LingDT-VL-OCR-4B | 95.07 | |
| 3 | GLM-OCR | 93.86 | |
| 4 | PaddleOCR-VL-1.6 | 93.67 | |
| 5 | PaddleOCR-VL-1.5 | 93.54 |
Interactive version: theaggregate.ai/benchmark?slug=longdocbench-page-level-parsing · How It Works · Data refreshed daily, snapshot 2026-11-12.