This paper advocates the use of virtual references for engine control loops instead of the commonly used physical variables. Virtual references are derived from measurements by the analysis of the sensitivity of the single target quantities with respect to the available process inputs. As these virtual references correspond locally to the target variables, decreasing them reduces the target variables as well, until the limits of the corresponding Pareto set are reached. The virtual variables can be seen as local mappings of the target variables onto an input space, and provide a precious insight on the contribution of the different input quantities. The performance of the method is shown experimentally on a 4 cylinder production engine at the well known example of the nitric oxides/particulate matter trade-off optimization in Diesel engines. An analysis during a torque transient is then used to confirm the validity of the choice of the variables.