What they are doing is just conditioning on the topic, author, institution and media attention to predict replicability. But they predict averages, can't make reliable predictions about specific papers. And the study is focusing only on Psychology, even though they trained their embeddings on 2m paper abstracts.