Adaptive Contrastive Decoding for Instruction-Following in Large Language Models

Large language models (LLMs) fine-tuned with instructions demonstrate promising capabilities in adhering to constraints embedded within prompts. However, their performance remains limited when handling more complex constraints. Moreover, existing studies suggest that imposing additional constraints can adversely impact the model’s reasoning ability and understanding of domain-specific knowledge. To address these challenges, we propose a novel approach called Adaptive Contrastive Decoding (ACD). This method decouples instructions from constraints, allowing the model to generate initial predictions based solely on unconstrained instructions. Subsequently, constrained instructions are used to verify the generated tokens in parallel, and constraint-sensitive tokens are refined using contrastive decoding. Experimental results on three standard benchmarks for instruction-following tasks demonstrate that ACD significantly outperforms existing methods (e.g. DOLA, SCD) in constraint adherence. Additionally, ACD achieves superior inference speed compared to traditional contrastive decoding methods, effectively balancing constraint adherence and inference efficiency.