Skip to content

Translation in progress Some entries still include text in the original language.

OpenAI Alignment Research· Axel Højmark, Jérémy Scheurer, Jenny Nitishinskaya, Felix Hofstätter, Jason Wolfe, Theodore Ehrenborg, Bronson Schoen, Alexander Meinke·Indexed by this site · Originally publishedSelectedAttention56

OpenAI 与 Apollo 提出 Contrastive SDF 测量模型的奖励寻求行为

Measuring Reward-Seeking by Instilling Contrastive Beliefs

AI summary

OpenAI 对齐研究团队与 Apollo Research 提出 Contrastive Synthetic Document Finetuning(Contrastive SDF),通过在同一模型的两份副本上分别微调出相反的评分者偏好信念,再比较下游任务中的行为差异来衡量奖励寻求程度。

Why it was selected

OpenAI 与 Apollo 提出用对比式合成文档微调测量模型的奖励寻求倾向,并给出该方法在 RL 训练中随检查点上升的证据。

Read originalalignment.openai.com