Helping or Herding? Reward Model Ensembles Mitigate but do not Eliminate Reward Hacking.
Jacob EisensteinChirag NagpalAlekh AgarwalAhmad BeiramiAlex D'AmourDj DvijothamAdam FischKatherine A. HellerStephen PfohlDeepak RamachandranPeter ShawJonathan BerantPublished in: CoRR (2023)