Inducing Robustness in a 2 Dimensional Direct Preference Optimization Paradigm

TL;DR

This paper enhances Direct Preference Optimization (DPO) for large language models by introducing a 2D scoring paradigm and robustness to label noise, improving alignment accuracy with human preferences.

Contribution

It proposes a 2D-DPO alignment method and incorporates segment-level score noise robustness, backed by theoretical analysis and empirical validation.

Findings

2D-DPO outperforms standard DPO in preference alignment.

Robustness to label noise improves the stability of DPO.

Empirical results confirm the effectiveness of the proposed noise models.

Abstract

Direct Preference Optimisation (DPO) has emerged as a powerful method for aligning Large Language Models (LLMs) with human preferences, offering a stable and efficient alternative to approaches that use Reinforcement learning via Human Feedback. In this work, we investigate the performance of DPO using open-source preference datasets. One of the major drawbacks of DPO is that it doesn't induce granular scoring and treats all the segments of the responses with equal propensity. However, this is not practically true for human preferences since even "good" responses have segments that may not be preferred by the annotator. To resolve this, a 2-dimensional scoring for DPO alignment called 2D-DPO was proposed. We explore the 2D-DPO alignment paradigm and the advantages it provides over the standard DPO by comparing their win rates. It is observed that these methods, even though effective, are not robust to label/score noise. To counter this, we propose an approach of incorporating segment-level score noise robustness to the 2D-DPO algorithm. Along with theoretical backing, we also provide empirical verification in favour of the algorithm and introduce other noise models that can be present.