Towards Robust Alignment of Language Models: Distributionally Robustifying Direct Preference Optimization

TL;DR

This paper introduces Dr. DPO, a robust training method for LLMs that mitigates noise in preference data by leveraging distributionally robust optimization principles, leading to improved text quality and response accuracy.

Contribution

It presents a novel framework, Dr. DPO, that extends DPO with distributionally robust optimization to handle both pointwise and pairwise noise in preference datasets.

Findings

Dr. DPO significantly improves text quality and response accuracy.

Theoretical analysis shows DPO's inherent robustness to pointwise noise.

Empirical results demonstrate enhanced performance in noisy environments.

Abstract

This study addresses the challenge of noise in training datasets for Direct Preference Optimization (DPO), a method for aligning Large Language Models (LLMs) with human preferences. We categorize noise into pointwise noise, which includes low-quality data points, and pairwise noise, which encompasses erroneous data pair associations that affect preference rankings. Utilizing Distributionally Robust Optimization (DRO), we enhance DPO's resilience to these types of noise. Our theoretical insights reveal that DPO inherently embeds DRO principles, conferring robustness to pointwise noise, with the regularization coefficient $\beta$ playing a critical role in its noise resistance. Extending this framework, we introduce Distributionally Robustifying DPO (Dr. DPO), which integrates pairwise robustness by optimizing against worst-case pairwise scenarios. The novel hyperparameter $\beta'$ in Dr. DPO allows for fine-tuned control over data pair reliability, providing a strategic balance between exploration and exploitation in noisy training environments. Empirical evaluations demonstrate that Dr. DPO substantially improves the quality of generated text and response accuracy in preference datasets, showcasing enhanced performance in both noisy and noise-free settings. The code is available at https://github.com/junkangwu/Dr_DPO.