AI 导读
我们使用拒绝采样微调来模仿有用的步骤，并用提示引导的自蒸馏来纠正错误。 GLM 5.2 用相同的权重对同一段记录回合评分两次：一次带纠正提示，一次不带。训练让无提示的下一 token 预测与带提示时的预测对齐。
正文
We use rejection sampling fine-tuning to imitate useful steps and hint-guided self-distillation to correct errors.
GLM 5.2 scores the same recorded turn twice, using the same weights: once with a corrective hint and once without it. Training aligns the hint-free next-token predictions with those made with the hint.
来源：@perplexity_ai · x.com