Automated methods for identifying subsets of a model's components (neurons, attention heads) that are responsible for specific behaviors.