eval approach: GPT4 provides alternative responses with reasoning and then has to pick the best one

[qrdlgit]

Describe the feature or improvement you're requesting

I think it would be interesting to see how GPT4 performs on more ideal scenarios, where COT is allowed extensively.

In particular, GPT4 is allowed to think step by step along several different approaches to solve a problem and provide its reasoning. Some prompting should be done to ensure that it tries to look at the problem from different angles / viewpoints / perspectives.

Then, in a second prompt, it's asked to pick from the approach it thinks is most likely to be correct.

If there has been a paper or effort done on evaluating GPT4 for something like this, I'd greatly appreciate a link. :)

Additional context

No response

[qrdlgit]

https://arxiv.org/pdf/2305.10601.pdf