PaveVQA - Localization Token-F1: leaderboard
PaveVQA is the vision-language question answering part of PaveBench (2026), built on real highway pavement inspection images with single-turn, multi-turn and expert-corrected questions about distress recognition, localization, quantification and maintenance. This board is token-level F1 in percent of short spatial descriptions on localization questions, answered zero-shot; higher is better.
Metric: Token-level F1 (%) of short spatial descriptions on localization questions, of the zero-shot model on PaveVQA, the vision-language question answering part of PaveBench (real highway pavement images, single-turn, multi-turn and expert-corrected questions); higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 3 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek-VL2-Small | 40.36 | #2464 |
| 2 | LLaVA-OneVision-7B | 27.71 | #2981 |
| 3 | Qwen 2.5 VL 3B | 16.39 | #2639 |
Interactive version: theaggregate.ai/benchmark?slug=pavevqa-localization-token-f1 · How It Works · Data refreshed daily, snapshot 2026-11-21.