Can Large Language Models generalize analogy solving like people can?