Stability as a Liability:Systematic Breakdown of Linguistic Structure in LLMs

Xianzhe Meng, Qiangsheng Zeng, Ling Luo +7

Training stability is typically regarded as a prerequisite for reliable optimization in large language models. In this work, we analyze how stabilizing training dynamics affects th…