News for hackers. Stuff that matters.
Re:LLMs cannot find reasoning errors, but can correct them by mkl on Tuesday November 21, 2023 @02:11AM (#38360394)
Getting attached to a perspective despite evidence to the contrary would require perspective and distinguishing fact from fiction, but just copying humans protesting that they're right (regardless of context) seems plausible, as there's a lot of that to learn from.
Re:LLMs cannot find reasoning errors, but can correct them by kromem on Tuesday November 21, 2023 @11:39PM (#38374785)
> distinguishing fact from fiction

Actually this part does seem in recent research to be encoded in LLMs at an abstract level in a linear representation...

https://arxiv.org/abs/2310.06824