It does talk about openai explicitly instruction tuning the llm to try to constrain the output and the limitations of such approaches.