A novel method that discovers interpretable yet discriminative sets of attributes for visual recognition by integrating large language models and evolutionary search.