Reward-Augmented Data Enhances Direct Preference Alignment of LLMs

TL;DR

This work identifies critical limitations of direct preference alignment, notably ignoring reward magnitudes and gaps that lead to unlearning of high-quality rejected responses and poor generalization to rare high-reward outputs. It introduces reward-conditioned LLM policies via a simple reward-augmented data relabeling scheme, which constructs a twofold, goal-conditioned preference dataset that can be optimized with existing direct alignment methods like DPO. Theoretical guarantees accompany the approach, and extensive experiments on UltraFeedback across multiple model families show consistent improvements on instruction-following and academic benchmarks, along with ablations that demonstrate the method's data-efficiency and robustness to reward scaling. The proposed framework enables learning from the full spectrum of response quality, including multi-attribute rewards, and offers practical, plug-and-play enhancement to current preference-based alignment pipelines with tangible real-world impact on controllable, high-quality LLM behavior.

Abstract

Preference alignment in Large Language Models (LLMs) has significantly improved their ability to adhere to human instructions and intentions. However, existing direct alignment algorithms primarily focus on relative preferences and often overlook the qualitative aspects of responses, despite having access to preference data that includes reward scores from judge models during AI feedback. Striving to maximize the implicit reward gap between the chosen and the slightly inferior rejected responses can cause overfitting and unnecessary unlearning of the high-quality rejected responses. The unawareness of the reward scores also drives the LLM to indiscriminately favor the low-quality chosen responses and fail to generalize to optimal responses that are sparse in data. To overcome these shortcomings, our study introduces reward-conditioned LLM policies that discern and learn from the entire spectrum of response quality within the dataset, helping extrapolate to more optimal regions. We propose an effective yet simple data relabeling method that conditions the preference pairs on quality scores to construct a reward-augmented dataset. The experiments across various benchmarks and diverse models demonstrate that our approach consistently boosts DPO by a considerable margin. Through comprehensive ablation studies, we demonstrate that our method not only maximizes the utility of preference data but also mitigates the issue of unlearning, demonstrating its broad effectiveness beyond mere data expansion. Our code is available at https://github.com/shenao-zhang/reward-augmented-preference.

Figures (4)

• Figure 1: Illustration of our method: construction of reward-augmented preference datasets followed by direct alignment algorithms.
• Figure 2: Instruction-following performance of the base models, the models trained with DPO on UltraFeedback, and the models trained with DPO on reward-augmented UltraFeedback on AlpacaEval 2.0, MT-Bench, and Arena-Hard-Auto benchmarks. Our method demonstrates considerable improvements consistently across all benchmarks. The complete table is deferred to Appendix \ref{['sec_full']}.
• Figure 3: Our method helps mitigate the unlearning issue of DPO.
• Figure 4: Comparisons with DPA, SteerLM, and (Info)NCA.

Theorems & Definitions (6)

• Theorem 4.1: Informal version
• Theorem 1.4: Suboptimality of Algorithm \ref{['alg:theory']}
• Lemma 1.5: Oracle optimal KL-regularized policy
• Theorem 1.6
• Lemma 1.7: Uniform concentration
• Lemma 1.8: Difference of Sigmoid functions