What if we prompt the LLM to generate a response with citations, and then we have program which looks up the citations in a citation database to validate their correctness? Responses with hallucinated citations are thrown away, and the LLM is asked to try again. Then, we could retrieve the text of the cited article, and get another LLM to opine on whether the article text supports the point it is being cited in favour of. I think a lot of these problems with LLMs could be worked around with a few more moving parts.