Heuristics are widely applied to modularity maximization models for the identification of communities in complex networks. We present an approach to be applied as a post-processing to heuristic methods in order to improve their performances. Starting from a given partition, we test with an exact algorithm for bipartitioning if it is worthwhile to split some communities or to merge two of them. A combination of merge and split actions is also performed. Computational experiments show that the proposed approach is effective in improving heuristic results.