DocAttriBench - Answer Locating: leaderboard
Answer locating in DocAttriBench (De Grandis et al., 2026): given only a question about a document image, predict the layout elements that contain the evidence for its answer. Box F1 in percent at IoU 0.5, averaged over the seven document VQA test sets; higher is better.
Metric: Box F1 (%; answer locating: given only the question, predict the evidence boxes, mean over the seven test sets DocVQA, VisualMRC, VISA, SlideVQA, VisualWebBench, LongDocURL and MMLongBench-Doc; 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: Estimated already saturated. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2.5 VL 32B Instruct | 19.3 | #1055 |
| 2 | InternVL3-38B | 18.6 | #905 |
| 3 | Qwen 2.5 VL 7B Instruct | 17.9 | #1842 |
| 4 | InternVL3-8B | 11.1 | #1695 |
| 5 | InternVL2.5-2B | 0.4 | #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-answer-locating · How It Works · Data refreshed daily, snapshot 2026-11-12.