APCD: Adaptive Path-Contrastive Decoding for Reliable Large Language Model Generation
A research paper introduced Adaptive Path-Contrastive Decoding (APCD) to enhance the reliability of large language model outputs by addressing error accumulation during generation. APCD employs an entropy-driven approach to decide when to explore alternative token paths and uses divergence-aware contrast to manage interactions between these paths. The method aims to improve factual accuracy and generation efficiency, as demonstrated on eight benchmarks. AI
IMPACT Introduces a novel decoding strategy that could lead to more reliable and factually accurate LLM outputs.