The paper in this post doesn't use neural networks at all. It also includes the time to train the model in the sorting time, which the paper you linked doesn't.

The paper in this post doesn't use neural networks at all. It also includes the time to train the model in the sorting time, which the paper you linked doesn't. NN-sort trains on historic data to sort future data, which requires the distribution to be roughly the same, whearas this algorithm learns online.