A technique applied during model inference (not training) to steer or correct the model's behavior, such as prompting or decoding modifications.