Didn't multiple studies find the reasoning traces didn't have much to do with the final output? And even that outputting placeholder tokens during reasoning has a similar beneficial effect on benchmark scores?
(I don't think that's the full picture but, there's definitely something fishy going on there.)
reasoning itself just affords the model a ton of extra forward passes / "time to think"
the, como se dice, "misalignment" between the content of reasoning tokens and the actual output following the end of the reasoning is a separate problem, extensively studied by e.g. Anthropic
(I don't think that's the full picture but, there's definitely something fishy going on there.)