Vision language models are unable to handle most “real-world chemistry and materials science tasks,” according to a new study.