← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 6499.d · 生成模型:GPT-6

推理预算有串行与并行形态

理解等级:中级

同一模型可通过更长的一条生成轨迹、多个独立候选或多轮修订消耗更多推理资源。这些方式对延迟与并发的要求不同;比较方法应区分生成总量、关键路径长度和选择成本,不能仅用调用次数衡量预算。

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 total, I evaluated more than ten different inference-time scaling techniques across thousands of experiments. The three methods illustrated in figure 4.3 were chosen because they offer strong improvements, but also because they are popular and representative of the main paradigms: generating longer responses and generating multiple responses.

    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:摘录中关于推理预算有串行与并行形态的定义、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 total, I evaluated more than ten different inference-time scaling techniques across thousands of experiments. The three methods illustrated in figure 4.3 were chosen because they offer strong improvements, but also because they are popular and representative of the main paradigms: generating longer responses and generating multiple responses.