In the absence of a very strong justification, my assumption for any random technique in an AI paper is that they tried a bunch of different things and whatever gave the highest evals made it into the paper (even though that performance is likely random not a genuine improvement)