Cut Fine-Tuning Cost: Adapt LMs to Multi-Modal Tasks with <1% New Params

GLUE and V&L results show near–full-tune accuracy, strong few-shot transfer, and far lower per-task storage than current adapter/prefix methods.