LLMs can't self-correct in reasoning tasks, DeepMind study finds

Re:LLMs can't self-correct in reasoning tasks, DeepMind stud... by mike00632 on Monday October 09, 2023 @03:32PM (#37824378)

I just asked Chat GPT 3.5 the following: "Please give an exponential equation that has integer solutions which is difficult to solve.", and the response is: "

Certainly, here's an exponential equation that can be challenging to solve for integer solutions:

2^x+3^y=7^z

This equation involves three variables, x, y, and z, and requires finding integer values for these variables that satisfy the equation. This type of equation is known as a Diophantine equation, and solving it can be quite challenging, especially for larger values of x, y, and z.

[ Parent | on HN ]

Re:LLMs can't self-correct in reasoning tasks, DeepMind stud... by shinycode on Monday October 09, 2023 @03:47PM (#37824600)

Can it reason about a math problem that no human ever thought about and solve it ?

[ Parent | on HN ]

Re:LLMs can't self-correct in reasoning tasks, DeepMind stud... by og_kalu on Monday October 09, 2023 @03:58PM (#37824734)

Reasoning about a math problem that no human has ever thought about and solving it is the bar for understanding now ? Can you do that ?

[ Parent | on HN ]

Re:LLMs can't self-correct in reasoning tasks, DeepMind stud... by crdrost on Monday October 09, 2023 @04:13PM (#37824906)

That's quite the nerd snipe.

I mean, it's definitely not a Diophantine equation and solving it is definitely not challenging -- (2, 1, 1) happens to be an easy solution -- but I want to say it probably doesn't have any other solutions but I don't see a great way to prove it...

[ Parent | on HN ]