Can Large Language Models Understand Intermediate Representations?