Video Preference Alignment on VideoReward-Bench
61.26Alignment Accuracy (w/ties)VideoReward
Evaluation results
| Method | Links | ||
|---|---|---|---|
| VideoRewardHuman Dataset Size=182K, Backbone Size=2B2026.02 | 61.26 | 73.59 | |
| GT-SVJHuman Dataset Size=30.4K, Backbone Size=2B2026.02 | 57.01 | 68.57 | |
| VisionRewardHuman Dataset Size=2000K, Backbone Size=19B2026.02 | 56.77 | 67.59 | |
| GT-SVJ (p = 0)Human Dataset Size=30.4K, Backbone Size=2B2026.02 | 55.42 | 65.37 | |
| GT-SVJ (No DM)Human Dataset Size=30.4K, Backbone Size=2B2026.02 | 53.14 | 63.21 | |
| VideoScoreHuman Dataset Size=37.6K, Backbone Size=8B2026.02 | 41.8 | 50.22 |