My guess is your point is "is this effectively better" than beam search, I wondered the same thing. The paper doesn't mention it, which at least on the surface seems strange.
Beam search is like a breadth-limited breath-first search. This is more akin to a depth-first search where you give the model the ability to backtrack.