So the idea is that LLM1 looks at the output of LLM0 and builds a new set of constraints, and then LLM0 has to try again, rinse and repeat? (LLM0 could be the s

So the idea is that LLM1 looks at the output of LLM0 and builds a new set of constraints, and then LLM0 has to try again, rinse and repeat? (LLM0 could be the same as LLM1, and I think it is in the article?)