作者位次 #5/5ACL 2026Poster· Natural Language Processing
A Layer-wise Analysis of Supervised Fine-Tuning
Qinghua Zhao · Xueling Gong · Xinyu Chen · Zhongfeng Kang · Xinlu Li
摘要
While critical for alignment, Supervised Fine-Tuning (SFT) incurs the risk of catastrophic forgetting, yet the layer-wise emergence of instruction-following capabilities remains elusive. We investigate this mechanism via a comprehensive analysis utilizing information-theoretic, geometric, and opt...