META, Stanford, Harvard, NYU: NEW RL & SFT Training Algo

Your AI Base Model is a Genius. Your Post-Training is Destroying it. How an Old Statistical Trick (MCMC) is Fixing Modern AI. We finally understand what RL actually does to LLMs (And it's terrifying). Why the Future of AI isn't Better Models, it's Better Sampling. All rights w/ authors: Sharpening Tax in Post-Training Changdae Oh1,2,∗, Qi Zeng1, Qi Qi1, Andrey Zhmoginov, Deren Lei1, Yun He1, Hoang Phan1,3,∗, Hangoo Kang4, Azalia Mirhoseini4, Sharon Li2 from 1 Meta Superintelligence Labs, 2 Un