I'd argue that it's the same thing. Asking a language model to “reason step by step” only marginally improved the results before we steered the thinking traces with RLVR.