APCD: Adaptive Path-Contrastive Decoding for Reliable Large Language Model Generation
The APCD framework improves LLM generation reliability by using two mechanisms: Entropy-Driven Path Expansion and Divergence-Aware Path Contrast. It prevents premature branching by waiting for high uncertainty before exploring multiple paths, and it manages multi-path exploration by dynamically weakening interactions between paths as they diverge. This results in more accurate outputs while maintaining efficiency.