Apple questions capabilities of AI reasoning models in new research paper