Ensuring the safety of language models (LMs) while maintaining their usefulness remains a critical challenge in AI alignment. Current approaches rely on sequential adversarial training: generating adversarial prompts and fine-tuning LMs to defend against them. We introduce a different paradigm: framing safety alignment as a non-zero-sum game between an Attacker LM and a Defender LM trained jointly via online reinforcement learning. Each LM continuously adapts to the other's evolving strategies, driving iterative improvement. Our method uses a preference-based reward signal derived from pairwise comparisons instead of point-wise scores, providing more robust supervision and potentially reducing reward hacking. Our RL recipe, AdvGame, shifts the Pareto frontier of safety and utility, yielding a Defender LM that is simultaneously more helpful and more resilient to adversarial attacks. In addition, the resulting Attacker LM converges into a strong, general-purpose red-teaming agent that can be directly deployed to probe arbitrary target models.
@article{arxiv.2512.20806,
title = {Safety Alignment of LMs via Non-cooperative Games},
author = {Anselm Paulus and Ilia Kulikov and Brandon Amos and Rémi Munos and Ivan Evtimov and Kamalika Chaudhuri and Arman Zharmagambetov},
journal= {arXiv preprint arXiv:2512.20806},
year = {2026}
}