This seems relatively clearly false in the case of competition programming problems. Concretely, o3 with 50 submissions beats o1 with 10k submissions. (And o1 is presumably much better than the underlying instruct model.)
I'd guess this paper doesn't have the actual optimal methods.