Mechanistic Decoding of Cognitive Constructs in Large Language Models