Dicrete latent variable models (DLVM) are a unified framework for generative clustering models including mixture models and block models. We propose a new algorithm to greedily maximise the integrated likelihood criterion with respect to a partition in any DLVM where this quantity is analytic.