Video Preference Alignment on GenAI-Bench
64.26Alignment Accuracy (w/ties)GT-SVJ
Evaluation results
| Method | Links | ||
|---|---|---|---|
| GT-SVJHuman Dataset Size=30.4K, Backbone Size=2B2026.02 | 64.26 | 75.15 | |
| UnifiedReward2026.01 | 54.8 | 70.9 | |
| VisionRewardHuman Dataset Size=2000K, Backbone Size=19B2026.02 | 51.56 | 72.41 | |
| VideoRewardHuman Dataset Size=182K, Backbone Size=2B2026.02 | 49.41 | 72.89 | |
| GT-SVJ (p = 0)Human Dataset Size=30.4K, Backbone Size=2B2026.02 | 49.36 | 62.41 | |
| VideoReward2026.01 | 49.1 | 72.8 | |
| VideoScoreHuman Dataset Size=37.6K, Backbone Size=8B2026.02 | 49.03 | 71.69 | |
| GT-SVJ (No DM)Human Dataset Size=30.4K, Backbone Size=2B2026.02 | 47.62 | 59.12 | |
| VideoScore22026.01 | 39.1 | 61.6 | |
| Qinsighttype=image-based evaluator2026.01 | 37.6 | 57.1 | |
| REACT2026.01 | 37.6 | 58.1 |