Safety Alignment of LMs via Non-cooperative Games
Fuente:
arXiv
Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866911737512460288 |
|---|---|
| author | Paulus, Anselm Kulikov, Ilia Amos, Brandon Munos, Rémi Evtimov, Ivan Chaudhuri, Kamalika Zharmagambetov, Arman |
| author_facet | Paulus, Anselm Kulikov, Ilia Amos, Brandon Munos, Rémi Evtimov, Ivan Chaudhuri, Kamalika Zharmagambetov, Arman |
| contents | 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. Code at github.com/facebookresearch/advgame. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2512_20806 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Safety Alignment of LMs via Non-cooperative Games Paulus, Anselm Kulikov, Ilia Amos, Brandon Munos, Rémi Evtimov, Ivan Chaudhuri, Kamalika Zharmagambetov, Arman Artificial Intelligence 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. Code at github.com/facebookresearch/advgame. |
| title | Safety Alignment of LMs via Non-cooperative Games |
| topic | Artificial Intelligence |
| url | https://arxiv.org/abs/2512.20806 |