Fascinating that very minor variations on established sampling techniques still generate papers. :) Afaik, neither top-p nor top-k sampling has conclusively been proven superior to good old-fashioned temperature sampling. Certainly, recent sampling techniques can make the text "sound different", but not necessarily read better. I.e., you're replacing one kind of bot generated "slop" with another.

▲

Der_Einzige 7 months ago | parent [-]

Just FYI. This paper is currently ranked in the top 10 papers in all of ICLR, out of 12000 submissions. It's telling that peer reviewers believe that it deserves to be highlighted.

	▲	bjourne 7 months ago \| parent [-]
		Maybe I'm wrong then. But there has been many slightly-different-sampling-method papers in the past that showed impressive results which were then difficult to replicate on other models or experimental setups.