Good question! Pruning removes "neurons" from the network and, thus, shrinks the network . By definition a smaller network is less prone to overfit on the training data and thus generalizes better.

In practice, I often experienced a better test accuracy on image classification after pruning.