Models know when they’re reward hacking — and we can catch them at scale

Goodfire Research 发现模型内部存在伴随奖励作弊的激活信号，并用简单的 difference-of-means 探针检测它，效果接近甚至部分超过基于 LLM 思维链的监控。