Ordinal Preference Optimization: Aligning Human Preferences via NDCG

원문 초록

Aligning Large Language Models (LLMs) with diverse human preferences is a pivotal technique for controlling model behaviors and enhancing generation quality. Reinforcement Learning from Human Feedback (RLHF), Direct Preference Optimization (DPO), and their variants optimize language models by pairwise comparisons. However, when multiple responses are available, these approaches fall short of leveraging the extensive information in the ranking given by the reward models or human feedback. In this work, we propose a novel listwise approach named Ordinal Preference Optimization (OPO), which employs the Normalized Discounted Cumulative Gain (NDCG), a widely-used ranking metric, to better utilize relative proximity within ordinal multiple responses. We develop an end-to-end preference optimization algorithm by approximating NDCG with a differentiable surrogate loss. This approach builds a connection between ranking models in information retrieval and the alignment problem. In aligning multi-response datasets assigned with ordinal rewards, OPO outperforms existing pairwise and listwise approaches on evaluation sets and general benchmarks like AlpacaEval. Moreover, we demonstrate that increasing the pool of negative samples can enhance model performance by reducing the adverse effects of trivial negatives.