Or it's just some weird property of the specific training and inference setting the author was working with. The epistemic claims the author is making aren't s

Or it's just some weird property of the specific training and inference setting the author was working with. The epistemic claims the author is making aren't supported by the work. The work is interesting for other reasons of course.