What is the predictive log-likehood

Thanks for the excellent work, I am very interested in it. Since the paper said because there is no predictive log-likehood, so it is difficult to estimate the generalization ability. How to understand this statement? Any advice including explanation and paper reference would be appreciated. Thank you!