知识卡片
训练投入与推理投入可以叠加
强化学习更新参数,推理时扩展则在固定参数上投入更多生成或选择计算。训练后的模型仍可使用采样、投票或修订,二者并不互斥。比较方案时要分开记录一次性训练成本和每次请求成本。
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
原书出处 · 1 处
- · 查看原书依据
they can be combined. For example, after improving a model's reasoning capabilities through reinforcement learning (RL), which is the core focus of this chapter, inference-time scaling techniques such as those introduced in chapters 4 and 5 can be applied to further boost performance.
参考来源
- 位置:《Build a Reasoning Model (From Scratch)》第06章《强化学习训练推理模型》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Build a Reasoning Model (From Scratch)/_epub-src/OEBPS/Text/chapter-6.html)
- 结论依据:摘录中关于训练投入与推理投入可以叠加的定义、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:they can be combined. For example, after improving a model's reasoning capabilities through reinforcement learning (RL), which is the core focus of this chapter, inference-time scaling techniques such as those introduced in chapters 4 and 5 can be applied to further boost performance.