Re:LLMs can't self-correct in reasoning tasks, DeepMind stud...
by jncfhnb on Tuesday October 10, 2023 @08:58PM
(#37839736)
No! This is a bad argument.
Your one off example of an LLM not being able to handle a request that is not solveable is not relevant to claim of whether LLMs can fundamentally do this.
Curren state LLMs do current have the ability to point out that some requests are impossible.
They’re not great at it. But they can do it. What you’re hitting is a current state quality problem.
There’s no particular reason to think that it’s unachievable with the same approach.
Your one off example of an LLM not being able to handle a request that is not solveable is not relevant to claim of whether LLMs can fundamentally do this.
Curren state LLMs do current have the ability to point out that some requests are impossible.
They’re not great at it. But they can do it. What you’re hitting is a current state quality problem.
There’s no particular reason to think that it’s unachievable with the same approach.