Breaking safety paradox with feasible dual policy iteration

Published in International Conference on Learning Representations (**ICLR**), 2026