Alignment-Aware Decoding

Alignment of large language models remains a central challenge in natural language processing. Preference optimization has emerged as a popular and effective method for improving alignment, typically through training-time or prompt-based interventions. In this paper, we introduce alignment-aware decoding (AAD), a method to enhance model alignment directly at inference. Theoretically, AAD can be interpreted as implicit reward optimization, yet it requires no specialized training beyond the standard DPO setup. Empirically, AAD consistently outperforms strong baselines across diverse alignment benchmarks and model scales. Moreover, in data-constrained settings, AAD can produce high-quality synthetic data to improve alignment under standard decoding, providing a practical solution when labeled data is limited. Large language models (LLMs) are the backbone of modern natural language processing, powering applications ranging from open-ended dialogue to complex reasoning tasks. Despite their impressive capabilities, aligning these models with human preferences remains a central challenge. Misaligned models can produce harmful, biased, or simply unhelpful outputs, motivating a growing body of work on alignment, i.e., the process of training models to better reflect human values and preferences (Ziegler et al., 2019; Ouyang et al., 2022; Amodei et al., 2016).