Improving Model Safety Behavior with Rule-Based Rewards

In three linesOpenAI introduces a Rule-Based Rewards (RBR) method to align models toward safe behavior without extensive human data collection.

Summary generated by Claude — human-verified