Enhancing Automated Interpretability with Output-Centric Feature Descriptions