zlacker

>>koie+(OP)
No, they can't "correct reasoning errors", and that's a clickbait title.

>>einpok+xj
If you look at the paper, they only claim LLM can correct the errors if the mistake location is given. And the mistake finding part is not yet solved.

>>ming03+0A
They don't correct errors even then. They just generate something which sounds like what one might say in a conversation when constrained not to express the error. If there's essentially just one option for that, its the correct one - but then it's like telling someone that the answer to a yes/no question is not the one they generated. Not much "error correction" to do then.