Abstract: Methods, systems and computer program products for evaluating performance of generative models are disclosed. One method includes providing a base model and a candidate model having observed variables and first and second conceptually related variables related to the observed variables, respectively, receiving observations assigned to a subset of the observed variables, and for each observation, evaluating the observation by the base model to produce a base assessment of the observation, evaluating the observation by the candidate model to produce a second assessment of the observation, determining a similarity measure of the assessment of the observation based on the base and second assessments, and selecting a subset of observations having low similarity measures for use in evaluating performance of the candidate model.