Reinforcement learning towards broadly and persistently beneficial models

OpenAI Alignment

OpenAI reports that RL training on realistic scenarios targeting beneficial traits improves alignment across dozens of benchmarks. Gains generalize beyond training domains and persist under adversarial pressure. The dataset spans domains like health, science, education, and coding.

OpenAIReinforcement learning

OpenAI Alignment·Akshay V. JagadeeshRahul AroraMikhail TrofimovKhaled SaabAli MalikFoivos TsimpourlasKaran Singhal

Jun 18, 6:00 PM