Apple researchers found that today’s most advanced AI reasoning models, though better than standard LLMs on moderately complex tasks, ultimately fail at higher complexities. That exposes fundamental limits in their ability to generalize reasoning.