On: Large Language Models Are Neurosymbolic Reasoners

[CuriouslyC]

Chain of thought is basically reasoning as humans do it, the only difference is that unlike humans the model can't see that its output is wrong, abandon a line of reasoning and re-prompt itself (yet).

[naasking → CuriouslyC]

Various attempts at feeding their output back in to check itself have shown marked improvements in accuracy.

[skenderbeu → CuriouslyC]

Multi agent LLMs talking to each other can already do this. It's just not cost feasible yet because it can lead to infinite loops and no solutions