AIConfPaper
← Search

Johannes Ackermann

2 accepted papers

2026

Gradient Regularization Prevents Reward Hacking in Reinforcement Learning from Human Feedback and Verifiable Rewards

Johannes Ackermann, Michael Noukhovitch, Takashi Ishida, Masashi Sugiyama
ICML 2026poster

Reinforcement Learning from Human Feedback (RLFH) or Verifiable Rewards (RLVR) are two key steps in the post-training of modern Language Models (LMs). A common problem is reward hacking, where the policy may exploit inaccuracies of the reward and learn an unintended behavior. Most previous works add…

Cited by 0Source ↗Scholar ↗