知识卡片
反馈循环必须有终止预算
先生成、验证、回传错误再修订,直到通过或达到轮次上限。反馈需指出可操作问题,并保留失败轨迹;不设终止预算可能反复修改而无真实进展,通过条件也不能只依模型自评。
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
Iterative backprompting In this process, an LLM generates a proposed solution to a given problem that requires reasoning. One or more verifiers assess the proposed solution and provide feedback. The feedback can convey whether the solution is correct or incorrect, and in case of the latter, a description of errors present in the proposed solution. The LLM takes the feedback as input and generates the
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:摘录中关于反馈循环必须有终止预算的定义、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:Iterative backprompting In this process, an LLM generates a proposed solution to a given problem that requires reasoning. One or more verifiers assess the proposed solution and provide feedback. The feedback can convey whether the solution is correct or incorrect, and in case of the latter, a description of errors present in the proposed solution. The LLM takes the feedback as input and generates the