北大与 Samsung 论文：慢模型离线指导快模型，实时游戏得分从 13.51 升至 24.19
北大和 Samsung 做了个 RealtimeBench，测试里 GPT-6.1-sol 思考太久反而输给快模型 Jev，再用 TORSI 让慢模型离线写指令带快模型，得分翻到 24.19。
北京大学与 Samsung 的论文指出，游戏评测常在模型思考时暂停世界，慢模型的延迟成本被忽略。他们构建了 RealtimeBench，包含 Mario、Snake、Tetris 等 6 款游戏共 180 个任务，推理期间游戏持续运行。暂停时 GPT-6.1-sol 在 6 款游戏中赢 5 款，实时场景下只拿到 1.51 分，低于快模型 Jev 的 13.51 分。他们提出 TORSI 方法，让慢模型在局间分析快模型的失败并撰写指令，2 轮后 Jev 平均分从 13.51 提高到 24.19，6 类游戏全部提升。
New Peking University and Samsung paper shows that a slow reasoning model can coach a fast model offline, raising its live game score from 13.51 to 24.19.
shows that the stronger reasoning model in paused games can fall far behind a faster model once the game keeps moving.
Many game evaluations pause the world while the model thinks, so slow answers cost nothing. In real settings, a correct action can arrive too late.
They built RealtimeBench, 180 tasks across 6 games like Mario, Snake, and Tetris that keep running during inference. GPT-6.1-sol beat the faster Jev in 5 of 6 games when paused, but scored 1.51 versus 13.51 out of 100 live.
Their fix, TORSI, has the slow model study the fast model's failures between games and write instructions the fast model follows during play. After 2 rounds, Jev's average score rose from 13.51 to 24.19, improving all 6 game types.
For real-time agents, keep a fast model in the loop and use the slow model offline to write and test its instructions.
