In my view as well it's not really cheating, it's just over fitting.

If a model doesn't do good in the benchmarks it will either be retrained until it does or you won't hear about it.