Robust LLM Alignment via Distributionally Robust Direct Preference Optimization

TL;DR

This paper introduces two distributionally robust algorithms, WDPO and KLDPO, to improve large language model alignment with human preferences under distribution shift, demonstrating superior performance on benchmark datasets.

Contribution

The paper develops novel distributionally robust preference optimization algorithms for LLM alignment, addressing distribution shift issues with scalable learning methods.

Findings

WDPO and KLDPO outperform existing methods under preference distribution shifts

The algorithms are scalable and suitable for large models

Empirical results show significant alignment improvements

Abstract

A major challenge in aligning large language models (LLMs) with human preferences is the issue of distribution shift. LLM alignment algorithms rely on static preference datasets, assuming that they accurately represent real-world user preferences. However, user preferences vary significantly across geographical regions, demographics, linguistic patterns, and evolving cultural trends. This preference distribution shift leads to catastrophic alignment failures in many real-world applications. We address this problem using the principled framework of distributionally robust optimization, and develop two novel distributionally robust direct preference optimization (DPO) algorithms, namely, Wasserstein DPO (WDPO) and Kullback-Leibler DPO (KLDPO). We characterize the sample complexity of learning the optimal policy parameters for WDPO and KLDPO. Moreover, we propose scalable gradient descent-style learning algorithms by developing suitable approximations for the challenging minimax loss functions of WDPO and KLDPO. Our empirical experiments using benchmark data sets and LLMs demonstrate the superior performance of WDPO and KLDPO in substantially improving the alignment when there is a preference distribution shift.