A model's ability to recall factual knowledge even when the exact wording or phrasing differs from training data.