Tag
This paper presents Iso-LoRA, an optimizer that enhances LoRA by promoting even energy distribution across singular directions through spectral descent on tangent-space perturbations, improving effective rank and downstream performance in language model adaptation.