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

Gradient Regularization Prevents Reward Hacking in Reinforcement Learning from Human Feedback and Verifiable Rewards explores A novel gradient regularization technique to prevent reward hacking in reinforcement learning models.. Commercial viability score: 2/10 in Reinforcement Learning.