https://hackerpaper.kaipfister.de/p/2606.12790/Collected HN comments and submissions linking this paper.enMon, 02 Nov 2026 11:16:22 GMThttps://news.ycombinator.com/item?id=50110856urn:hackerpapers:mention:2606.12790:50110856Thu, 15 Oct 2026 21:49:44 GMT<p>That second &quot;Mind the Gap&quot; paper seems to be doing something very much like that, though their &quot;conformative decoding&quot; (CD) seems very simple: they make both the RLHFed and the base model produce a probability distribution for a given context, then apply top-p truncation to the RLHFed probabilities as one normally does, and then they look up what logprobs the base model assigned to those same tokens, and sample according to the linear combination of the two models&#x27; logprobs. This seems to work a moderate amount for them. Looking at the cited-by I found a paper that evaluated a bunch of such…</p><p><a href="https://hackerpaper.kaipfister.de/p/2606.12790/">Paper and discussion history</a>. A mention is not an endorsement.</p>cs.CL