I don’t think LLMs can easily find errors in their output.
There was a recent meme about asking LLMs to draw a wineglass full to the brim with wine.
Most really struggle with that instruction. No matter how much you ask them to correct themselves they can’t.
I’m sure they’ll get better with more input but what it reveals is that right now they definitely do not understand their own output.
I’ve seen no evidence that they are better with code than they are with images.
For instance, if the time to complete only scales with length of the token and not the complexity of its contents then it probably safe to assume it’s not being comprehended.
In my experience, if you confuse an LLM by deviating from the the "expected", then all the shims of logic seem to disappear, and it goes into hallucination mode.
Tbf that was exactly my point. An adult might use 'inference' and 'reasoning' to ask clarification, or go with an internal logic of their choosing.
ChatGPT here went with a lexigraphical order in Python for some reason, and then proceeded to make false statements from false observations, while also defying its own internal logic.
"six" > "ten" is true because "six" comes after "ten" alphabetically.
No.
"ten" > "seven" is false because "ten" comes before "seven" alphabetically.
No.
From what I understand of LLMs (which - I admit - is not very much), logical reasoning isn't a property of LLMs, unlike information retrieval. I'm sure this problem can be solved at some point, but a good solution would need development of many more kinds of inference and logic engines than there are today.
Do you believe that the LLM understands what it is saying and is applying the logic that you interprets from its response, or do you think its simply repeating similar patterns of words its seen associated with the question you presented it?
If you take the time to build an (S?)LM yourself, you'll realize it's neither of these. "Understands" is an ill-defined term, as is "applying logic".
But a LLM is not "simply" doing anything. It's extremely complex and sophisticated. Once you go from tokens into high-dimensional embeddings... it seems these models (with enough training) figure out how all the concepts go together. I'd suggest reading the word2vec paper first, then think about how attention works. You'll come to the conclusion these things are likely to be able to beat humans at almost everything.
Start adding different prompts, different models and you get all kinds of ways to catch errors. Just like humans.