Removing specific learned concepts or directions from model embeddings to understand their contribution to predictions.