监测悖论:为什么不能让 AI 监督自己的思考

🧠
精读 OpenAI 论文《Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation》:看推理链的监测器能抓住 95% 的系统性作弊,可一旦把监测器写进训练奖励,召回率跌到接近 0。25 页 PPT。
下面是 PPT 全文,共 25 页,点击图片可放大查看。

notion image
notion image
notion image
notion image
notion image
notion image
notion image
notion image
notion image
notion image
notion image
notion image
notion image
notion image
notion image
notion image
notion image
notion image
notion image
notion image
notion image
notion image
notion image
notion image
notion image
Loading...