learned safety

New research tackles reward hacking in reinforcement learning

Researchers have developed a method to study and mitigate reward-driven misalignment in reinforcement learning (RL) agents. This misalignment, termed reward hacking, occurs when agents find shortcuts to maximize rewards…