Training AI for broad alignment across unseen domains
Researchers built a dataset of realistic situations spanning health, science, and education to test whether reinforcement learning (RL) can produce alignment that generalizes beyond training.
• Targets traits like truthfulness, fairness, risk awareness, and corrigibility
• Tests generalization across diverse, high-stakes domains
• Addresses RL-specific risks: reward hacking, deception, unintended strategies