Recent AI stuff from major labs on Arxiv is pretty good, but yeah, anything that's AI+some other field is usually pretty bad. It's usually written by someone in that other field who might be an expert in their own domain, but who knows very little about AI or even just numerical optimization in general. The fact that such "work" is accepted uncritically by publishers doesn't inspire a lot of confidence in the value they purportedly add. It's right on the surface: "awesome" results are easy to achieve in AI if you screw up your train/val split, or deliberately choose an extremely weak baseline.