Small Language Models Improve Giants by Rewriting Their Outputs