A model's ability to decide when to rely on external context and when to ignore it based on whether the context is helpful or misleading.