AI 导读
强化学习扩展的一个基本思路：我们能否把更多算力分配给更难的问题？ 我们这样做了：如果你的 GRPO 组里所有回答都是错的，就以概率 P（约 0.9）多采样一些——以寻找更多梯度非零的 GRPO 批次。它有效！ 叫做 "Never Give Up" https://t.co/bNPn6EQsiE https://t.co/VcE8I1kKtE
正文
An basic idea in scaling RL: Can we allocate more compute to the harder problems?
We did this: If your GRPO group has all wrong completions, sample more with probability P (~0.9) -- in search of more GRPO batches with nonzero gradient. It works!
Called "Never Give Up" https://t.co/bNPn6EQsiE https://t.co/VcE8I1kKtE
来源：@natolambert · x.com