Inducing Robustness in a 2 Dimensional Direct Preference Optimization Paradigm
This work addresses a practical issue in aligning LLMs for applications like chatbots, but it is incremental as it builds on existing 2D-DPO methods.
The paper tackles the lack of robustness to label noise in 2D-DPO, a method for aligning LLMs with human preferences using granular scoring, and proposes a segment-level noise robustness approach that improves win rates by 5-10% over standard DPO in noisy conditions.
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.