A parameter-efficient fine-tuning method that prepends learnable tokens to input sequences to adapt model behavior.