[akomtu]

LLMs can't evaluate their own output. LLMs suggest possibilities, but can't evaluate them. Imagine an insane man who is rumbling something smart, but doesn't self-reflect. The evaluation is done against some framework of values that are considered true: the rules of a board game, the language syntax or something else. LLMs also can't fabricate evaluation because the latter is a rather rigid and precise model, a unlike natural language. Otherwise you could set up two LLMs questioning each other.

[candiddevmike]

Isn't this kind of the hope/dream of multi-agent systems where one LLM "coordinates" among others or checks the responses? In my experience it works about as well as you're describing.

[izabera]

oh boy do i have the paper for you https://proceedings.neurips.cc/paper_files/paper/2014/file/f...

[mdp2021]

Sorry, what do GANs have to do with this? It is not the same kind of "evaluation".

And anyway, there is no need to have two networks to iteratively refine output: one suffices (like we naturally are meant to do).