Lack of abstraction formation — Models memorize instead of compressing concepts. Early signal: Sample efficiency remains poor.