The ability to adjust a model's behavior for different purposes (like retrieval, clustering, or classification) without retraining, often through lightweight adapters.