They cite https://arxiv.org/abs/2503.23829 which is interesting though. If you have lots of tokens to burn, just try the task lots of times. Can find better solutions than reasoning would on its first try. Only tested on small models though.