It sounds like they're giving it some temperature to allow answer variation, then having it score its own output and taking the best answers. This is akin to c

It sounds like they're giving it some temperature to allow answer variation, then having it score its own output and taking the best answers. This is akin to common bootstrap methods in statistics that resample the data to estimate a distribution over outcomes based on sample variance. The key in my opinion is the self-scoring procedure, how that is implemented would really control how well this process works. Since it works pretty well here - a 70b non-moe model in striking distance of GPT4 is huge - I'm inclined to say they've happened upon a self-scoring procedure that's good at reinforcing desired traits in the model.