WorldSense
Multimodal · 2025-02-06
WorldSense evaluates omni-modal video understanding with 3,172 four-option multiple-choice questions over 1,662 audio-visual synchronized videos spanning 8 domains and 26 tasks, all manually annotated. Models receive video frames with the audio track, and scores are multiple-choice accuracy.
Top models (higher is better)
| Model | Score |
|---|---|
| LFM3-Omni-2B | 47.1 |
| Phi-4-multimodal-instruct | 42.4 |