Tag
This paper introduces DeScore, a video reward model that decouples reasoning and scoring processes to improve training efficiency and generalization. It addresses the limitations of existing discriminative and generative reward models by using a 'think-then-score' paradigm with multimodal large language models.