知识卡片
并行多候选受覆盖与验证限制
生成k个多样候选,再由验证器选择,成功依赖正确答案是否出现在候选集中及是否被识别。更高温度可能增加差异也增加无效答案,不能只增加k而忽略验证器误判及总成本。
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
Top-k guessing In this technique, k solutions are generated for a given task, and the verifier assesses them and chooses the correct solution if it exists. A relatively high temperature (>1) is used during decoding to generate a diverse set of solutions. Kambhampati et al. show that top-k guessing exhibits similar performance levels as iterative backprompting. Inference-Time Computation This might well be the most significant
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:摘录中关于并行多候选受覆盖与验证限制的定义、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:Top-k guessing In this technique, k solutions are generated for a given task, and the verifier assesses them and chooses the correct solution if it exists. A relatively high temperature (>1) is used during decoding to generate a diverse set of solutions. Kambhampati et al. show that top-k guessing exhibits similar performance levels as iterative backprompting. Inference-Time Computation This might well be the most significant