Reasoning is not model improvement