I thought they pretty clearly explained where the additional performance came from. If there is only 1 valid schema confirming option you can skip the LLM entir

I thought they pretty clearly explained where the additional performance came from. If there is only 1 valid schema confirming option you can skip the LLM entirely. If there are only a limited number of possible tokens (eg. only } or ,) then you run it on smaller subset of the model. Between these two you capture a large amount of the actual token count of most json.