“On the GSM dataset, "checkability training" with small models successfully transferred to human evaluators, increasing their accuracy and decreasing the time they took to judge solutions.”