A parameter-efficient fine-tuning technique that adapts a pre-trained model by learning orthonormal transformations, preserving the model's original knowledge while specializing it for new tasks.