控制令牌

Disentangling Memory and Reasoning Ability in Large Language Models