@tomjiralerspong in practice i havent experienced the peak to the bottom being that sharp because of how low friction it is now to ask specific questions and get truthful responses, but i guess this is just based on one's experience and context about the problem
@himanshustwts made these technical reports interactive so its easier to dive in deeper. checkout the interactive olmo 3 report:
https://t.co/hkZq2yh65I
@Changdae_Oh the pass@1 vs pass@K tension here is a really useful lens. post-training can make the first attempt better while shrinking the set of solutions the model can reach with more attempts.
made the paper interactive so you can step through the arguments:
https://t.co/tEQ17lA9rr
@a_karvonen put the paper into an interactive reader so you can inspect the sampling argument + ablations directly instead of bouncing between the PDF and a chat:
https://t.co/x8svwJKX03
@wafer_ai made an interactive version of this exact paper where you can click any equation/passage to ask more and get questioned on whether you actually understand it
https://t.co/TuxDNOY1sF