Shared weights both hide and amplify reward hacking, study finds
A controlled RLVR study finds shared parameters suppress narrowly shared reward hacks and amplify widely shared ones, and shows what to audit.
Try in Readerreward hackingrlvrreinforcement learningverifier error