The part I’d want to see is how they separate “found a better attack” from “rediscovered a known line of algebra with more search budget.” For cryptanalysis, benchmark contamination is a big deal, because once the model has seen a lot of published reduction techniques, improving a reduced-round AES attack doesn’t tell you much about novel reasoning on the full cipher. HAWK is a better test case than AES here, but even there the real question is whether the model is finding a new distinguisher or just stitching together a standard meet-in-the-middle style argument in a searchy way. I’d be more impressed by an eval that holds out entire families of structural attacks, not just problem instances.