Providing preference weights or trade-off parameters as input to a model to control its behavior at inference time.