preference optimization

Rubric-Based DPO for Visual Preference Tuning

rDPO uses instance-specific rubrics to make visual preference optimization more fine-grained, improving filtering and benchmark results.