Those benchmarks don't tell much, they only check if a problem was solved, not how. Also there's surely a lot of benchmaxxing going on in the model training.