I also think that a none linear reward curve would benefit Steem.

Maybe a sigmoid function as suggested by @twinner might be interesting as it would make self-upvotes less attractive but at the same time would prevent the incredible high rewards for some trending posts which resulted from the n^2 curve.