The latest large reasoning models (LRMs) experience “complete accuracy collapse” when faced with highly complex tasks, according to a new paper