OpenAI 用规则奖励提升模型安全行为

We've developed and applied a new method leveraging Rule-Based Rewards (RBRs) that aligns models to behave safely without extensive human data collection.