Because that would be twice as computationally intensive. "Reasoning" models integrate some of that natively. In a way, they're trained to double check themsel

Because that would be twice as computationally intensive. "Reasoning" models integrate some of that natively. In a way, they're trained to double check themselves - which does improve accuracy at the cost of compute.