Dual Traits in Probabilistic Reasoning of Large Language Models