arXiv · 2026 · Preprint

Safety Alignment as Continual Learning: Mitigating the Alignment Tax via Orthogonal Gradient Projection

Guang-Long Sun, Siyuan Zhang, Li-Yuan Wang, Jun Zhu, Hang Su, Yi Zhong

Date
2026-02-08
Citations
4
arXiv
2602.07892
Cite
@misc{sun2026safety,
  title = {Safety Alignment as Continual Learning: Mitigating the Alignment Tax via Orthogonal Gradient Projection},
  author = {Guang-Long Sun and Siyuan Zhang and Li-Yuan Wang and Jun Zhu and Hang Su and Yi Zhong},
  year = {2026},
  eprint = {2602.07892},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2602.07892},
}