Reward-Augmented Data Enhances Direct Preference Alignment of LLMs

Abstract: Preference alignment in 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.

• The paper introduces a novel reward-augmented data method that enhances LLM direct preference alignment by retaining high-quality responses.
• It employs reward-conditioned data relabeling to train models on both chosen and rejected responses, reducing overfitting and boosting generalization.
• Empirical results across benchmarks like AlpacaEval 2.0 and GSM8K confirm significant performance gains over standard DPO techniques.

Overview of "Reward-Augmented Data Enhances Direct Preference Alignment of LLMs"

The paper "Reward-Augmented Data Enhances Direct Preference Alignment of LLMs" addresses the challenges and limitations inherent in current preference alignment methodologies for LLMs. Specifically, it introduces a novel approach that leverages reward-augmented data to improve direct preference alignment, thereby enhancing the ability of LLMs to not only follow human instructions but also generalize to responses that achieve higher rewards.

Motivation and Problem Statement

The study begins by examining the prevalent issues with direct alignment algorithms in LLMs, which primarily focus on relative preferences and often neglect qualitative differences between responses. This oversight can lead to overfitting and the unintentional unlearning of high-quality but rejected responses. Furthermore, the models may fail to generalize effectively to optimal responses due to the sparsity of high-reward data points. The study aims to resolve these challenges by introducing reward-conditioned policies.

Proposed Methodology

The researchers propose a data relabeling methodology that constructs a reward-augmented dataset. This strategy involves conditioning preference pairs on the quality scores assigned by judge models, which evaluate AI feedback. The approach is designed to work seamlessly with existing direct alignment algorithms, and it is applicable to any preference dataset.

The core idea is to optimize the LLM to discern and learn from the entire spectrum of response quality. By conditioning responses on reward scores, the model becomes more adept at recognizing patterns across varying qualities. This enables it to retain high-value information from both chosen and rejected responses, facilitating better generalization to sparse, high-quality responses.

Empirical Evaluation

The paper details comprehensive experiments across several benchmarks, such as AlpacaEval 2.0, MT-Bench, and Arena-Hard-Auto, involving multiple LLMs including Zephyr, Mistral, Qwen2, Llama3.1, Gemma2, and SPPO. The experimental results consistently show that the proposed method substantially improves the performance of Direct Preference Optimization (DPO), extending far beyond mere dataset expansion. On average, the method also enhances accuracy across academic benchmarks like GSM8K, GPQA, MUSR, TruthfulQA, BBH, and ARC.

Implications and Speculations

The implications of this research are twofold: practical and theoretical. Practically, it offers a scalable approach to improving LLM alignment with minimal computational overhead, which is critical given the increasing deployment of LLMs in real-world applications. Theoretically, it emphasizes the significance of incorporating qualitative data considerations into model training, which could reshape future alignment strategies.

Looking forward, this study sets a precedent for further exploration into reward-conditioned learning mechanisms. Future developments could explore more complex reward formulations and their impact on other aspects of LLM behavior, potentially improving robustness, safety, and fairness in AI systems.

Conclusion

In conclusion, this paper makes a significant contribution by addressing a critical gap in the direct alignment of LLMs using a reward-augmented data approach. The enhancements in model performance across diverse benchmarks underscore the viability and effectiveness of this strategy, paving the way for more refined alignment techniques in the domain of large-scale LLMs. This work represents a valuable step forward in aligning AI outputs more closely with intended human preferences.