Thinking Without Output: Toward Modal Cognition in Language Models — LessWrong