Preference Alone Falls Short: Arena's Research Reveals Limits of Human Reward Models for Image Generation

Arena, the AI evaluation platform, has published research showing that standard human preference reward models are insufficient for training frontier text-to-image systems—revealing that models optimized to please users often miss fine details, add unrequested content, or fail on compositional accuracy.