Training a model to match a teacher model's probability distributions over possible next items, capturing nuanced user preferences.