Dr. DocBench - Text Edit Distance: leaderboard
Dr. DocBench (2026) parses 4,514 difficult pages sampled by parser failure from long books in 52 subject domains, with expert annotations of text, formulas, tables and reading order. This board is the normalized edit distance of the extracted page text against the expert transcription, times 100; lower is better.
Metric: Normalized Levenshtein edit distance of the extracted page text against the expert transcription, times 100 (0-100), 4,514 expert-annotated difficult pages sampled by parser failure from long books in 52 subject domains; each page parsed with the model's own document-to-markdown prompt (specialized parsers through their official APIs); lower is better. Source: arxiv.org. Saturation forecast: Around February 2027. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 19 | #49 |
| 2 | Kimi K2.5 | 19 | #256 |
| 3 | Gemini 3.1 Pro (Preview) | 22 | #94 |
| 4 | Claude Opus 4.6 | 22 | #102 |
| 5 | Qwen 3.5 122B A10B | 23 | #331 |
| 6 | Qwen 3.5 Plus | 25 | #232 |
| 7 | Qwen 3.5 Flash | 26 | #435 |
| 8 | GPT-4o | 37 | #741 |
| 9 | Nemotron Nano 12B v2 VL | 62 | #2231 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=dr-docbench-text-edit-distance · How It Works · Data refreshed daily, snapshot 2026-11-13.