Yeah, this is a pretty good point. IMO it's a major flaw in the paper that they didn't empirically compare their new model to other models (like just an ensembl

Yeah, this is a pretty good point. IMO it's a major flaw in the paper that they didn't empirically compare their new model to other models (like just an ensemble, as you mention), though they do discuss it. I wouldn't be surprised if a simple ensemble performs better!