← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8682.d · 生成模型:GPT-6

概率空间搜索需要质量信号

理解等级:高级

深度优先、宽度优先或MCTS可探索不同生成路径,但模型高概率不等于答案正确。搜索还需定义状态、扩展、评价与停止规则;没有可靠评价时,算力可能只让错误更充分展开。

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

      compute is to search through the generation probability distribution. The idea is that for every token being generated, there is a probability assigned to every token in the vocabulary, so if we search properly, we might end up traversing the probability space and choosing the right path to generate the correct answer. Search-based methods include depth first, breadth first, and Monte Carlo Tree Search (MCTS).

    参考来源

    - 位置:《Designing Large Language Model Applications》第08章《对齐训练与推理》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch08.html) - 结论依据:摘录中关于概率空间搜索需要质量信号的定义、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:compute is to search through the generation probability distribution. The idea is that for every token being generated, there is a probability assigned to every token in the vocabulary, so if we search properly, we might end up traversing the probability space and choosing the right path to generate the correct answer. Search-based methods include depth first, breadth first, and Monte Carlo Tree Search (MCTS).