知识卡片

普通读书笔记卡 · 6499.g · 生成模型:GPT-6

样本数增加不保证单调收益

理解等级:中级

更多候选提供更多观察,也消耗更多计算;在特定模型和解码分布下,有限评估中的投票准确率未必随样本数单调增加。应比较收益区间、成本和共同错误,不能把候选数当成自动提高正确性的旋钮。

这张卡片看不懂?

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

      Rows 5-7 show the self-consistency scaling results. We can see that increasing the number of samples from 3 to 10 further improves accuracy to 31.6%, but it also significantly increases runtime. In this case, there is almost no accuracy advantage to using 10 instead of 3 samples.

    参考来源

    - 位置:《Build a Reasoning Model (From Scratch)》第04章《推理时计算扩展》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Build a Reasoning Model (From Scratch)/_epub-src/OEBPS/Text/chapter-4.html) - 结论依据:摘录中关于样本数增加不保证单调收益的定义、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:Rows 5-7 show the self-consistency scaling results. We can see that increasing the number of samples from 3 to 10 further improves accuracy to 31.6%, but it also significantly increases runtime. In this case, there is almost no accuracy advantage to using 10 instead of 3 samples.