GeneralPoints test favors RL over SFT under rule and suit changes
A report on the GeneralPoints experiment describes a game similar to “24 points,” using four cards with addition, subtraction, multiplication and division. In a rule-change test, J, Q and K counted as 10 during training but as 11, 12 and 13 during testing; RL accuracy rose 3.5%, while SFT accuracy fell 8.1%. In a separate suit-change test, training used black suits and testing used red suits; RL rose 17.6% and SFT fell 9.9%. The report presents these results as an example of how RL and SFT can behave differently when inputs change, while cautioning that they do not show RL always generalizes better or SFT only memorizes.
The reported results illustrate how outcome-based RL and supervised fine-tuning can differ when rules or card suits change, while the two experiments do not support universal conclusions about either method.
原文:mp.weixin.qq.com/s/Mc-6RlqGn…(2026年10月11日)
做 AI 应用的人常有这种直觉:监督微调(SFT)出来的模型,遇到见过的题型特别稳,一遇到新花样就开始“死记硬背”地套旧答案。
这不是错觉。有一组像“教学演示”一样的实验,把这种差别清清楚楚地量了出来。
一个“24 点”游戏,照出了“背”和“懂”
这个游戏叫 GeneralPoints,规则和大家玩过的“24 点”很像:给你四张牌,用加减乘除凑出 24,每张牌用一次。