LLMs can't perform "genuine logical reasoning," Apple researchers suggest