Semiparametric Preference Optimization: Your Language Model is Secretly a Single-Index Model