Extension of the method to problems with higher (or lower) number of integer variables

[rezabayani]

@gasse Apologies if this sounds like a basic question: Does the method discussed in the paper "Exact Combinatorial Optimization ..." extend to problems with a different number of "integer variables" than seen during the training?
I have read the paper a few times and nowhere it is specifically mentioned whether "larger" and "harder" instances denote more integer variables. It is mentioned that the policy is learned in the GCNN structure and according to figure 2, the proposed structure gets rid of the constraints; but, it is still dependent on the size of integer variables. Does this mean "harder instances" implies the problem has only more constraints? If not, how does the GCNN assign probabilities to unseen variables as it is trained only for a fixed number of variables?