Abstract

Systematically optimizing adapter placement and rank in LoRA enhances code and math reasoning under tight computational budgets.