Visual abstract reasoning problems are a difficult challenge for neural networks to tackle, due to the involvement of different levels of knowledge abstraction to be learnt. Visual properties must be correctly extracted and linked to high-level concepts, on top of which further elaboration is required to solve the problems. This thesis uses Variational Auto-Encoders, and explores their different variants to obtain meaningful and disentangled latent representations to address these problems. Experimental results on a public dataset show that this approach can adapt to data distribution shifts over time by consolidating the previously learnt knowledge, showing improvements in terms of generalization on Out-of-Distribution data.