Improving Model Safety Behavior with Rule-Based Rewards

【RSS 原文 · 未经 AI 筛选或翻译】We’ve developed and applied a new method leveraging Rule-Based Rewards (RBRs) that aligns models to behave safely without extensive human data collection.