GAVEL - Grounding Accuracy (IoU 0.5): leaderboard
Localizing the visual evidence for a caption error: given an image and a caption sentence containing a hallucination, the model draws a box around the region that shows the inconsistency; GAVEL (Gao, Hashimoto and Saito, 2026). Accuracy in percent, the share of boxes with an IoU of at least 0.5 against the human-annotated box; higher is better.
Metric: Accuracy (%; share of predicted evidence boxes with IoU of at least 0.5 against the annotated box). Source: arxiv.org. Saturation forecast: Around 2029. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 30B A3B Instruct | 22.8 | #870 |
| 2 | GPT-5 | 17.4 | #163 |
| 3 | GPT-5 Mini | 14.7 | #349 |
| 4 | Qwen 3 VL 8B Instruct | 12.1 | #982 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=gavel-grounding-accuracy-iou-0-5 · How It Works · Data refreshed daily, snapshot 2026-11-12.