In the context of this paper, I think "generalizable reasoning" means that finding a method to solve the puzzle and thus being able to execute the method on puzzle instances of arbitrary complexity.