Their proposed solution, which is offered without extensive analysis, is to relay all blocks, even late ones, and then choose the preferred block in ties at random. Ignoring collateral vulnerabilities which a hasty implementations of forward-all might create, I believe this proposal has a problem in that it creates a selfishness advantage even without any sybil attack at all, so long as the selfish miner has enough hashrate.
But even without their fix, and assuming gamma=0 (i.e. the selfish pool never wins a block propagation race) then if I'm reading it right the authors' result purports to show that a pool with enough hash rate (more than a third of the total hash power) still has a selfishess advantage, which gets quite substantial as the size of the pool grows larger than that.
(Not that I'm particularly supporting their change - just questioning the implication that could be drawn from your post that with the protocol as it stands today it's not possible to gain a selfishness advantage without mounting a sybil attack.)
roy