Tag: reward hacking

A new theoretical review argues that reward hacking, benchmark failure, and distribution shift share a single structural root: observation regimes ...

Researchers have developed a reinforcement learning method that trains multimodal AI models to detect hateful and propagandistic memes in English ...