This seems like a really neat trick, I wonder how far this can be taken on larger models with multi-variate rewards?