知识卡片
生成候选与验证正确分工
当检验答案比从零求解容易时,可由生成器提出候选,再交给模型、工具或符号验证器核查。关键是验证器有可靠判据;若只是另一次同源猜测,不能把双模型一致当可靠证明。
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
- · 对齐训练与推理
原书出处 · 1 处
- · 查看原书依据
Verifiers for Improving Reasoning So, LLMs may not be all that great at producing the right answer to a question that requires multistep reasoning. But all hope is not lost. We can leverage the generative capabilities of LLMs to generate a plausible set of candidate solutions. These candidates can then be assessed by a verifier, which can identify the correct answer. This is possible in
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:摘录中关于生成候选与验证正确分工的定义、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:Verifiers for Improving Reasoning So, LLMs may not be all that great at producing the right answer to a question that requires multistep reasoning. But all hope is not lost. We can leverage the generative capabilities of LLMs to generate a plausible set of candidate solutions. These candidates can then be assessed by a verifier, which can identify the correct answer. This is possible in