知识卡片
序列求和不等于偏好短答案
较长序列的 logprob 往往更负,但训练更新依赖优势符号与梯度,不是把序列分数直接做最大值排序。正优势长回答也会被强化;不能从“求和更负”直接推出优化必然让输出变短。长度权重需从实际目标和采样分布分析。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
原书出处 · 1 处
- · 查看原书依据
implicitly favors the shorter one, since it is cheaper in terms of likelihood. Summed logprobs therefore encourage the model to stop earlier unless producing a longer response is justified by a higher reward.
参考来源
- 位置:《Build a Reasoning Model (From Scratch)》第06章《强化学习训练推理模型》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Build a Reasoning Model (From Scratch)/_epub-src/OEBPS/Text/chapter-6.html)
- 结论依据:摘录中关于序列求和不等于偏好短答案的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。
- 原始内容:implicitly favors the shorter one, since it is cheaper in terms of likelihood. Summed logprobs therefore encourage the model to stop earlier unless producing a longer response is justified by a higher reward.