A fast AI model's average live-game score reportedly rose from 13.51 to 24.19 after offline coaching

A post describing a Peking University and Samsung paper says RealtimeBench has 180 tasks across six games that keep running while models think. GPT-6.1-sol beat the faster Jev in five of six paused games, but scored 1.51 to Jev's 13.51 out of 100 in live play. The post says Jev's average score rose to 24.19 after two rounds of offline coaching in which a slower model reviewed failures and wrote instructions.