Multi Dimensional Preference Score

Multi-dimensional preference scores aim to capture the nuanced and multifaceted nature of human preferences, moving beyond single, overall scores that inadequately represent the complexity of judgment. Current research focuses on developing models, often leveraging large language models or adapting existing architectures like CLIP, to learn these preferences across multiple dimensions (e.g., aesthetics, semantic alignment, quality) from large datasets of human evaluations. This work is significant for improving the evaluation and development of systems like text-to-image generators and recommendation systems, as well as for advancing our understanding of how humans make complex decisions involving multiple criteria.