This is complimentary to LLMs. For example, LLMs often produce code that doesn't quite compile. What if with some tweaks we could make it run and produce some output? Then, we could assess if the output is right.