Human Preference Modeling

The field develops ways to learn what people value from pairwise comparisons, ratings, choices, response times, and written explanations, then uses those preferences to evaluate or guide AI systems. Common approaches include Bradley–Terry models, learned reward models, and preference optimization methods such as RLHF and DPO, applied to language, image and video generation, recommendations, and robot behavior. A central challenge is that preferences vary across people and contexts, while aggregated labels and automated judges can miss disagreement, uncertainty, or tacit criteria. Current work explores personalization and pluralistic models, but must balance capturing diverse judgments with collecting reliable feedback at scale and preserving safety constraints.