知识卡片
推理时算力分配要按任务衡量
可把预算分配给多次采样、顺序修订、较长中间过程或树搜索。收益取决于题目难度、模型与验证信号,不是增加token总能改善;应在固定成本或延迟约束下比较成功率。
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
Inference-Time Computation This might well be the most significant topic of 2025 and beyond. As of this book’s writing, scaling up pre-training seems to be providing diminishing returns. Therefore, there is a hunt for new scaling dimensions. The most promising among them is scaling up inference-time compute. The premise is simple. For a given query, instead of generating the final answer right away, what if
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:摘录中关于推理时算力分配要按任务衡量的定义、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:Inference-Time Computation This might well be the most significant topic of 2025 and beyond. As of this book’s writing, scaling up pre-training seems to be providing diminishing returns. Therefore, there is a hunt for new scaling dimensions. The most promising among them is scaling up inference-time compute. The premise is simple. For a given query, instead of generating the final answer right away, what if