LongVQUBench: leaderboard
Long-term video quality understanding benchmark (2026): multiple-choice accuracy on the held-out test split, averaged over twelve question dimensions spanning local distortion events, cross-event reasoning and global quality judgments of long videos. Higher is better, in percent.
Metric: Accuracy (%; overall mean of the 12 question dimensions; multiple-choice accuracy on the 60% held-out test split of LongVQUBench (1,200 long videos with synthetic or authentic spatial and temporal distortions); uniformly sampled frames at the frame budget each model selected on the validation split, one forward pass, one answer option). Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 69.5 | #167 |
| 2 | Qwen-VL-Max | 64 | #1222 |
| 3 | Qwen 3 VL 8B Instruct | 60.3 | #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 · How It Works · Data refreshed daily, snapshot 2026-11-13.