Now with enough tries it will happen that you find two blocks in a row - again happens all the time with pools. But what you do is that you wait until the other pools find theirs. Once they find the block (you already found it remember?), or they've wasted many cycles finding it, you publish yours that you've found already.
By doing this you kind of selectively publish the blocks you find, wasting other pools cycles and sort of giving you a heads start for the next block search. It's all actually about building a pool that selectively publishes blocks so that it can get you a heads start for the next mining cycle. Miners will likely join this selfish pool because they'll have a heads start and not feel as though their hashing power is going to waste since it's constantly being invalidated by the selfish pool (by publishing blocks at specific times).
The paper assumes there's only one selfish pool. I presume that if a selfish pool arises, there will also be a variety of other selfish pools arising at the same time. If all pools are acting this way, then there's really no incentive for miners to go to any specific pool (avoiding centralization). So it will just be like regular mining again.