#Reinforcement Learning from Human Feedback

Rubric-Driven DPO for Visual Tasks

A new rDPO framework uses instance-specific rubrics to create high-quality preference data, dramatically improving multimodal AI evaluation and performance.