NN should in principle be able to use the leveling effect in evaluation. But the problem is that they will never do anything they are not trained to do. If their training set of positions does not contain sufficiently many N Queens vs M Knights with N = 2-3 and M = 3-7 they will essentially just produce random values as evaluation when they are presented such positions. And of course training them on positions from actual Chess games, whether from human game databases or computer games, will not even expose them to a single such position.
NN should in principle be able to use the leveling effect in evaluation. But the problem is that they will never do anything they are not trained to do. If their training set of positions does not contain sufficiently many N Queens vs M Knights with N = 2-3 and M = 3-7 they will essentially just produce random values as evaluation when they are presented such positions. And of course training them on positions from actual Chess games, whether from human game databases or computer games, will not even expose them to a single such position.