PureDocBench - Real Degraded - Reading Order Edit Distance: leaderboard
Metric: Reading-order normalized edit distance (%) between the predicted and reference block sequences, times 100, on the real degraded track (the same pages through four physical acquisition chains: phone photos of prints and of photocopies, screen photographs, and screenshots with social-media compression) of 1,475 source pages in 10 domains and 66 subcategories, rendered from LLM-written HTML/CSS with source-linked annotations for text, formulas, tables and reading order; lower is better. Source: arxiv.org. Saturation forecast: Not forecast. 58 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Qwen 3.8 27B | 36.4 |
| 2 | Gemini 3.1 Pro (Preview) | 38.6 |
| 3 | Qwen 3.5 27B | 39 |
| 4 | Qwen 3.5 397B A17B | 39.9 |
| 5 | Claude Opus 5 | 40 |
| 6 | Qwen 3.5 122B A10B | 40.1 |
| 7 | Gemini 3.6 Flash | 40.6 |
| 8 | Qwen 3.5 35B A3B | 41.9 |
| 9 | GLM-5.3 Flash | 42.3 |
| 10 | Qwen 3.5 9B | 43.7 |
| 11 | Qwen 3.6 35B A3B | 46.9 |
| 12 | Qwen3.8-Flash-Next | 47.2 |
| 13 | Qwen 3.5 4B | 47.7 |
| 14 | Kimi K2.6 (Non-reasoning) | 48.1 |
| 15 | GPT-5.6 Sol | 50.2 |
Interactive version: theaggregate.ai/benchmark?slug=puredocbench-real-degraded-reading-order-edit-distance · How It Works · Data refreshed daily, snapshot 2026-10-07.