← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 6510.g · 生成模型:GPT-6

多轮修订要检验退化

理解等级:中级

继续批评和修订会消耗更多资源,也可能把正确稿改错。本章结果中两轮未总胜过一轮,因此应固定轮数或可验证停止条件,并比较初稿、各轮及最终接受稿质量;不能把迭代次数当作必然改善。

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

      In figure 5.21, the revised answer receives a higher logprob score (-0.377) than the initial answer (-1.258), which makes it reasonable to accept the revision. In practice, however, self-refinement can also produce worse answers, especially when run for multiple iterations. The scorer therefore provides a way to decide whether a revised answer represents an improvement.

    参考来源

    - 位置:《Build a Reasoning Model (From Scratch)》第05章《自我修正与推理时扩展》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Build a Reasoning Model (From Scratch)/_epub-src/OEBPS/Text/chapter-5.html) - 结论依据:摘录中关于多轮修订要检验退化的定义、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:In figure 5.21, the revised answer receives a higher logprob score (-0.377) than the initial answer (-1.258), which makes it reasonable to accept the revision. In practice, however, self-refinement can also produce worse answers, especially when run for multiple iterations. The scorer therefore provides a way to decide whether a revised answer represents an improvement.