I hadn't seen SwiReasoning (
https://swireasoning.github.io, paper and code), it looks like that works at generation time without any requirements on the model. It increases token-efficiency and accuracy, but at first skim it seems like this would be incompatible with multi-token prediction. For large reductions in token budget it could be worth it.