GAVEL - Explanation Score: leaderboard
Explaining a caption error: given an image and a caption sentence containing a hallucination, the model states what is wrong with it; GAVEL (Gao, Hashimoto and Saito, 2026). GPT-5 rates each explanation against the human annotation from 0 to 5, rescaled to 0-100 and averaged over the test sentences; higher is better.
Metric: LLM score (0-100; GPT-5 rates each explanation against the annotated one on a 0-5 scale, x20, averaged over all test sentences). Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 70 | #163 |
| 2 | GPT-5 Mini | 70 | #349 |
| 3 | Qwen 3 VL 30B A3B Instruct | 62 | #870 |
| 4 | Qwen 3 VL 8B Instruct | 54 | #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-explanation-score · How It Works · Data refreshed daily, snapshot 2026-11-12.