> Which itself is a bit concerning, as that implies that modest gains in accuracy require exponentially more compute time

This is more of what I was getting at. I agree they should label the axis regardless, but I think the scaling relationship is interesting (or rather, concerning) on its own.