LongVQUBench - Open-Ended Relevance: leaderboard
Open-ended video quality questions of LongVQUBench (2026): a GPT-5 judge rates how directly each answer addresses the question from 0 to 1, averaged over the three levels. Reported in percent, higher is better.
Metric: Relevance score (%; how directly the answer addresses the question; GPT-5 judge score of each free-form answer against the expert reference answer, rated 0 to 1 and averaged over the local, cross-event and global levels, reported in percent; open-ended questions of the 60% held-out test split, same frame budgets as the multiple-choice runs). Source: arxiv.org. Saturation forecast: Estimated already saturated. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 88.9 | #167 |
| 2 | Qwen-VL-Max | 83.7 | #1222 |
| 3 | Qwen 3 VL 8B Instruct | 79.3 | #981 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=longvqubench-open-ended-relevance · How It Works · Data refreshed daily, snapshot 2026-11-13.