A Unified Representation Underlying the Judgment of Large Language Models