Multi-Preference / Plackett-Luce DPO (mDPO / K-Pair DPO)

Listwise preference objectives use more than one chosen–rejected pair at a time. Plackett–Luce models assign probabilities to rankings; K-order Ranking Preference Optimization extends this approach to top-K rankings, where reliable feedback may exist only for the leading candidates.

Keep the feedback structure explicit: a full ordering, a top-K ordering and a set of preferred candidates are different observations. The KPO paper studies ranking tasks and adaptive K; it does not establish that every multi-response DPO method uses the same objective.