Tara Esmaeilbeig

This paper analyzes PEFT of LLMs via linearization, showing how an explicit inductive bias connects fine-tuning dynamics to NTKs and offering theoretical and empirical insights into when linearization accurately predicts adaptation performance.