Identifiable Steering via Sparse Autoencoding of Multi-Concept Shifts