I'm not sure if that will hold up. The LLM is not going to do anything random and that is actually a powerful component that makes original output possible.
I wonder if a combination would be useful. Use an actual GA to do the mutation, and then let an LLM "fix" each mutated child.
I wonder if a combination would be useful. Use an actual GA to do the mutation, and then let an LLM "fix" each mutated child.