Scores are one VLM judge’s preferences, not measured transcription accuracy. The pages are a diagnostic sample, not a random sample of all textbooks. Confidence intervals overlap for the top two models; use page-level inspection and independent human transcripts before choosing a winner.
01 / OVERVIEW
994 judged · 14 automatic tiesLeaderboard
| Rank | Model | ELO (95% interval) | Wins | Losses | Ties |
|---|
48 pages × 21 model pairs = 1,008 comparisons. The English-only rubric used Qwen3.5-35B-A3B (Novita). Images shown below are resized copies for viewing; the judge used images capped at 1,024 pixels on the long side and OCR text capped at 2,500 characters per output.
02 / PAGE BY PAGE
Inspect a comparison
JUDGE VERDICT