RefereeBench - Foul and Penalty Reasoning: leaderboard
RefereeBench (2026) tests multimodal models as sports referees on 925 officiating videos across 11 sports with 6,475 human-annotated multiple-choice questions. This board is accuracy in percent on questions about why a foul or penalty applies, averaged across sports; higher is better.
Metric: Accuracy (%) on foul and penalty reasoning questions (mean of Q3 and Q5), averaged across 11 sports; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 11 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 53.1 | #175 |
| 2 | Gemini 3 Pro | 53 | #145 |
| 3 | Seed 1.8 | 50.7 | #252 |
| 4 | GPT-5 | 47.8 | #178 |
| 5 | Claude Sonnet 4.5 | 42.5 | #261 |
| 6 | GPT-4o | 39.1 | #751 |
| 7 | Claude Haiku 4.5 | 38.7 | #583 |
| 8 | InternVL3.5-8B | 37.1 | #1604 |
| 9 | Qwen 3 VL 8B | 37.1 | #917 |
| 10 | VideoLLaMA3-7B | 32.3 | #2050 |
| 11 | LLaVA-Video-7B | 24.9 | #2276 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=refereebench-foul-and-penalty-reasoning · How It Works · Data refreshed daily, snapshot 2026-11-21.