The degree to which abstract concepts can be extracted or recovered from a model's internal representations.