DocAttriBench - VISA: leaderboard
Grounded answer generation on the Paper-VISA and Wiki-VISA pages test split of DocAttriBench (De Grandis et al., 2026), which pairs each question with automatically attributed and filtered evidence regions such as text blocks, tables and figures. An answer counts only when its text matches the reference and its predicted evidence box overlaps an attributed layout element with IoU of at least 0.5; zero-shot, overall answer correctness in percent, higher is better.
Metric: Overall answer correctness (%; grounded answer generation on Paper-VISA and Wiki-VISA (2,779 test questions): an answer counts only when its text matches the reference and its evidence box is correct; zero-shot, with model-specific prompts and box parsers; element-level evidence boxes from the MAPPET attribution pipeline over Docling layout regions, a box counting when IoU >= 0.5). Source: arxiv.org. Saturation forecast: Around 2030. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2.5 VL 32B Instruct | 0.6 | #1055 |
| 2 | InternVL3-38B | 0.4 | #905 |
| 3 | InternVL3-8B | 0 | #1695 |
| 4 | Qwen 2.5 VL 7B Instruct | 0 | #1842 |
| 5 | InternVL2.5-2B | 0 | #6868 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=docattribench-visa · How It Works · Data refreshed daily, snapshot 2026-11-12.