I wonder what would happen if this was applied to alpha code. Instead of generating 1 million codes for each problem during inference, do this kind of iteration

I wonder what would happen if this was applied to alpha code. Instead of generating 1 million codes for each problem during inference, do this kind of iteration during training and then generate less codes for each problem (or generate 1 million but hopefully better ones?).