There's still a lot of research to do for improving compositional reasoning in large (and small) language models.