Abstract:Reward modeling of human preferences is one of the cornerstones of building usable generative large language models (LLMs). While traditional RLHF-based alignment methods explicitly maximize the expected rewards from a separate reward model, more recent supervised alignment methods like Direct Preference Optimization (DPO) circumvent this phase to avoid problems including model drift and reward overfitting. Although popular due to its simplicity, DPO and similar direct alignment methods can still lead to degenerate policies, and rely heavily on the Bradley-Terry-based preference formulation to model reward differences between pairs of candidate outputs. This formulation is challenged by non-deterministic or noisy preference labels, for example human scoring of two candidate outputs is of low confidence. In this paper, we introduce DRDO (Direct Reward Distillation and policy-Optimization), a supervised knowledge distillation-based preference alignment method that simultaneously models rewards and preferences to avoid such degeneracy. DRDO directly mimics rewards assigned by an oracle while learning human preferences from a novel preference likelihood formulation. Our experimental results on the Ultrafeedback and TL;DR datasets demonstrate that policies trained using DRDO surpass previous methods such as DPO and e-DPO in terms of expected rewards and are more robust, on average, to noisy preference signals as well as out-of-distribution (OOD) settings.