I tend to not believe unbelievable results in machine learning. It's too easy to unintenionally cause some kind of information leakage. I haven't read the paper in detail though, so their experimentation setup could be foolproof, this is not a critique of this paper specifically.