Generalized Preference Optimization: A Unified Approach to Offline Alignment
This work provides a unified theoretical and algorithmic framework for offline alignment in AI, offering new variants and insights for practitioners, though it is incremental as it builds upon and generalizes existing methods.
The authors tackled the problem of offline preference optimization for aligning large models by proposing Generalized Preference Optimization (GPO), a unified framework that encompasses existing methods like DPO, IPO, and SLiC, and reveals connections between offline regularization and KL divergence regularization, with experiments showing similar trade-offs between regularization and performance across variants.
Offline preference optimization allows fine-tuning large models directly from offline data, and has proved effective in recent alignment practices. We propose generalized preference optimization (GPO), a family of offline losses parameterized by a general class of convex functions. GPO enables a unified view over preference optimization, encompassing existing algorithms such as DPO, IPO and SLiC as special cases, while naturally introducing new variants. The GPO framework also sheds light on how offline algorithms enforce regularization, through the design of the convex function that defines the loss. Our analysis and experiments reveal the connections and subtle differences between the offline regularization and the KL divergence regularization intended by the canonical RLHF formulation. In a controlled setting akin to Gao et al 2023, we also show that different GPO variants achieve similar trade-offs between regularization and performance, though the optimal values of hyper-parameter might differ as predicted by theory. In all, our results present new algorithmic toolkits and empirical insights to alignment practitioners.