AI RESEARCH

(How) Learning Rates Regulate Catastrophic Overtraining

arXiv CS.LG

ArXi:2604.13627v1 Announce Type: new Supervised fine-tuning (SFT) is a common first stage of LLM post-