RL debate give coarse-grained signal: we saw uplift in our last post, but protocol, RL algorithm & seed noise are hard to disentangle. Luckily best-of-N at 10 samples already yields 20–40pp accuracy uplift and judge-hacking. https://x.com/jacob_pfau/status/2073063892817731592 https://www.lesswrong.com/posts/hb8pv3zyAHGJpwz9F/debate-with-self-play-best-of-n-optimization [2/n] https://x.com/jacob_pfau/status/2075552387209503000/photo/1

@idavidrein https://x.com/jacob_pfau/status/2075552384932077616?s=20