PureDocBench - Real Degraded - Reading Order Edit Distance: leaderboard

Reading-order normalized edit distance on the real degraded track of PureDocBench (2026), a document-parsing benchmark of 1,475 source-traceable pages (pages captured by phone from prints, photocopies and screens, or as compressed screenshots). Percent; lower is better.

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: Around 2030. 58 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1TeleOCR35.2
2Qwen 3.8 27B36.4#221
3Gemini 3.1 Pro (Preview)38.6#99
4Qwen 3.5 27B39#375
5FD-RL39.1
6Dots.mocr39.3
7Qwen 3.5 397B A17B39.9#264
8Claude Opus 540#26
9Qwen 3.5 122B A10B40.1#346
10Gemini 3.6 Flash40.6#81
11OvisOCR240.7
12WeVisDoc-4B41.1
13Logics-Parsing-v241.6
14Qwen 3.5 35B A3B41.9#492
15GLM-5.3 Flash42.3#133

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=puredocbench-real-degraded-reading-order-edit-distance · How It Works · Data refreshed daily, snapshot 2026-11-21.