Verification and validation of LLM output in this context would mean doing all the same research, training etc done today for human staff and then comparing the results line by line. It would actually take more time. How do you know if the LLM failed to apply one of hundreds of rules from a procedure unless you have a human trained on it who has also examined every relevant document and artifact from the process?
> one clear driver is continually increasing rules, regulations, and compliance, along with fears of audits and lawsuits
As mentioned by the GP posts the main problem is the increasing rules, regulations and compliance need to be processed the admin staff not the research contributions itself (these invention and innovation parts are performed by the graduate students and professors who are getting cuts by the limited budget).
This AI based system will include (not limited to) LLM with RAG (with relevants documents) that can perform the work of the tens if not hundreds jobs of the admin staff. The agent AI can also include rule based expert system for assessment of the procedures. It will be much faster than human can ever be with the on-demand AWS scale scaling (pardon the pun).
Ultimately it will need only a few expert admin staff for the compliance validation and compliance instead tens of hundreds as typical now in research organizations. The AI based system will even get better over time due to this RLHF and expert human-in-the-loop arrangement.
