Impressive work! Their approach makes a ton of sense:
"by not directly optimizing for the benchmark, we show that it becomes much more representative"
and is evident in the generalization and robustness: ~74% higher on the adversarial benchmark and similar/superior results on the standard Imagenet