I suspect this has to do with newer training procedures for reasoning models, where they are injecting stuff like "wait a minute" to force the model to reason m

I suspect this has to do with newer training procedures for reasoning models, where they are injecting stuff like "wait a minute" to force the model to reason more, as described in the Deepseek R1 training docs