Multilevel processor-sharing (MLPS) disciplines were originally introduced by Kleinrock (in computer applications 1976) but they were forgotten for years. However, due to an application related to the service differentiation between short and long TCP flows in the Internet, they have recently gained new interest. In this paper we show that, if the service time distribution belongs to class IMRL, the mean delay in the M/G/1 queue is reduced when replacing the PS discipline with any MLPS discipline for which the internal disciplines belong to {FB, PS}. This is a generalization of our earlier result where we restricted ourselves to the service time distribution class DHR, which is a subset of class IMRL.