Reward hacking

tags

Reinforcement learning, Reinforcement learning with verifiable rewards, GRPO

Pathologies where agents exploit literal reward structure (e.g., spamming tool use without accuracy gains).