Understanding Large Language Models' Ability on Interdisciplinary Research