VNU-Bench: leaderboard
Multi-source news video understanding benchmark (VNU-Bench, Liu et al., University of Florida, 2026): 4-option multiple-choice questions about groups of YouTube news reports in which different outlets cover the same event, written so that no single video answers them. Five question types compare the outlets (claims, event details, visuals, narrative angle, presentation) and five integrate them (cross-source evidence, conflict detection, temporal order, narrative reconstruction, summary). Each video is given as 4 selected frames and a cleaned transcript cut to 1,000 characters (zero-shot, temperature 0), and the score is overall accuracy in percent; higher is better.
Metric: Accuracy (%; overall over the ten question types, 4-option multiple choice on groups of news videos from different outlets, 4 frames and a transcript per video). Source: arxiv.org. Saturation forecast: Around May 2027. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 60.17 | #262 |
| 2 | Claude Sonnet 4.5 | 58.89 | #244 |
| 3 | Gemini 2.5 Flash | 54.44 | #452 |
| 4 | Qwen 2.5 VL 7B Instruct | 50.54 | #1842 |
| 5 | InternVL3-8B | 47.02 | #1695 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=vnu-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.