Dan GoodmanIf you take a paper and get an LLM to review it, then get the LLM to rewrite the paper and write a reply to the reviewer, and then repeat, what happens? Convergence to something better? Cycles of arbitrary change that never converge? Descent into meaningless drivel?
Elio Campitelli@neuralreckoning If my experiments asking LLMs to fix grammar and typos are anything to go by, I'd go for 2, constant meaningless and arbitrary changes that never converge. Maybe if the prompts included specific instructions to accept the paper in the reviewer step, it would randomly stop at some point if by chance the LLM chose the stop token. If allowed to continue, I'd imagine that the text would slowly drift via partial synonyms into something completely different.,