← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 6498.b.4 · 生成模型:GPT-6

奖励来源不等于优化算法

理解等级:中级

RLVR 描述使用可验证奖励的训练设置,GRPO 描述利用组内相对信号更新策略的方法。相同奖励设置可搭配其他优化算法,组相对优化也不只接受数学布尔奖励;不要把两者当作同义词。

这张卡片看不懂?

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

      will use the group relative policy optimization (GRPO) algorithm for policy optimization. In short, RLVR determines what learning signal is available, and GRPO determines how that signal is used to update the model weights. Policy gradient algorithms A widely used policy gradient algorithm for LLMs is proximal policy optimization (PPO), which was popularized in RLHF. In principle, the same algorithm could also be applied in

    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:摘录中关于奖励来源不等于优化算法的定义、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:will use the group relative policy optimization (GRPO) algorithm for policy optimization. In short, RLVR determines what learning signal is available, and GRPO determines how that signal is used to update the model weights. Policy gradient algorithms A widely used policy gradient algorithm for LLMs is proximal policy optimization (PPO), which was popularized in RLHF. In principle, the same algorithm could also be applied in