XL-DocBench (OCR): leaderboard
1,519 questions over professional documents in six domains, from regulations and annual reports to clinical guidelines, with contexts running to 2,303 pages. Every answer was re-answered and page-cited by human experts, 194 in all, and 72.6 percent need evidence from more than one page. This board feeds the model text extracted from the pages, which beats rendered page images for every model evaluated both ways. Microsoft and Wuhan University, 2026.
Metric: Accuracy (%). Source: officeintelligence.github.io. Saturation forecast: Around April 2028. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 | 39.8 | #63 |
| 2 | GPT-5.4 | 37.3 | #79 |
| 3 | Kimi K2.5 | 35.8 | #144 |
| 4 | Qwen 3.5 35B A3B | 32 | #256 |
| 5 | GPT-5.2 | 30.9 | #110 |
| 6 | DeepSeek V3.2 | 29.6 | #205 |
| 7 | Qwen 3.5 9B | 22.3 | #379 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=xl-docbench-ocr · How It Works · Data refreshed daily, snapshot 2026-10-20.