Neurosymbolic Computation aims to unify the two main branches in AI, namely, neural networks and logic. Answer Set Program- ming (ASP) is a good candidate for the logic part as it offers a declara- tive and expressive language. We consider a Visual Question Answering (VQA) problem, where we want to answer a question using visual input, for which neural and neurosymbolic approaches achieved good results. Our interest is here with elaboration of the questions, when new pred- icates, of increasing sophistication, become available. To this end, we experimentally compare a neural-based approach against a neurosym- bolic one over the CLEVR dataset. The results show that the latter approach is more robust to the new questions, achieving high accuracy, and also provides the benefit of producing explainable answers. On the downside, the neurosymbolic approach requires that the semantics of the questions respective predicates have to be manually coded. Preliminary work is being done to relieve this by using Inductive Logic Programming to learn the semantics of the predicates. Introduction. Visual Question Answering (VQA) is the field of problems wh