Just curious if structured outputs/constrained generation improve model accuracy, e.g., for information extraction problems. Does anyone have experience with this and why?