Relevance Vector Machine (RVM) is a machine learning technique relying on Bayesian inference that can be used to solve tomography image reconstruction problems under a probabilistic framework. By highlighting discrepancies between entropy estimates and inaccuracies of the posterior distribution covariance matrix estimates, we demonstrate how the adaptive RVM framework does not offer reliable and consistent operation for certain soft-field tomography problems, herein exemplified by electrical capacitance volume tomography. This has important consequences on the practical applicability of RVM for image reconstruction problems involving such sensor modalities.