3/ Contrastive Decoding Improves Reasoning in Large Language Models – shows that contrastive decoding leads Llama-65B to outperform Llama 2 and other models on commonsense reasoning and reasoning benchmarks.

→ View original post on X — @dair_ai