Not yet. For (b), we gave some advantage to Eve by (1) running two steps of training Eve for every step of training A&B; and (2) running multiple independent

Not yet. For (b), we gave some advantage to Eve by (1) running two steps of training Eve for every step of training A&B; and (2) running multiple independent retrains of Eve from scratch after freezing the Alice & Bob networks. Not quite the same as increasing the capacity of the network, but similar. As you noted - we mostly stuck to the regime in which a solution could be found in theory (or trivially by a human), to explore whether or not the adversarial formulation of the NNs could get anywhere near it. Many next steps indeed.