How can one know that's not coming from the pre-trained data. The paper is trying to evaluate whether the LLM has general problem solving ability.