A rollback refinement process for software testing using software models in a model-based testing environment. The method includes: identifying a model having a set of model elements assigned relevance scores based on correspondence between each model element in the set of model elements and a pre-defined suspect element, the model producing an expected behavior when running a test case; generating a reduced model on condition that the reduced model include only model elements assigned a threshold reduction score corresponding to the relevance scores; determining that the reduced model does not reproduce the expected behavior when running the test case; rolling back a model element into the reduced model to produce a revised model; and responsive to the revised model not reproducing the expected behavior, performing a rollback refinement of the revised model based on a number of model elements rolled back into the reduced model.