Metraipolis › archive › dc_69ceec6f42e8

Addition by Subtraction: Toward Mitigating Fragile Safety Alignment in Large Language Models — arXiv:2611.06553

Source: arXiv (en)
Original: https://arxiv.org/abs/2611.06553
License (as archived): CC0-1.0
Archived: 2026-11-05T18:17:06.977Z
SHA-256 of the archived text: 19509f13d3d6355dd0ad4031f6db3b22d9b1c68f204bbd1ee75cbe22ae1c43a3
Stable URL of this page: https://metraipolis.com/documento/dc_69ceec6f42e8

How to cite

Addition by Subtraction: Toward Mitigating Fragile Safety Alignment in Large Language Models — arXiv:2611.06553 — arXiv (en), https://arxiv.org/abs/2611.06553 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_69ceec6f42e8 (Metraipolis archive dc_69ceec6f42e8, sha256 19509f13d3d6355dd0ad4031f6db3b22d9b1c68f204bbd1ee75cbe22ae1c43a3)

BibTeX

@misc{dc69ceec6f42e8,
  title = {Addition by Subtraction: Toward Mitigating Fragile Safety Alignment in Large Language Models — arXiv:2611.06553},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2611.06553},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_69ceec6f42e8; sha256 19509f13d3d6355dd0ad4031f6db3b22d9b1c68f204bbd1ee75cbe22ae1c43a3},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2611.06553. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Large language models (LLMs) undergo safety alignment to prevent harmful generations, yet they remain vulnerable to jailbreak attacks. Recent work reveals a deeper structural weakness: safety alignment can be highly sparse, such that pruning a small number of safety-critical neurons or ablating a single refusal direction can substantially undermine aligned behavior. In this work, we take the first steps toward strengthening safety alignment such that it is less dependent on sparse model components. We propose Addition by Subtraction, which gradually removes identified safety-critical neurons during finetuning and uses preference optimization to re-establish safety-aligned behavior through other parts of the model. Our results show that the resulting models are substantially more robust to pruning-based attacks. Remarkably, these benefits generalize beyond the white-box attacks motivating our design: Addition by Subtraction improves robustness against other sparsity-based interventions, including refusal-direction ablation, as well as a broad range of black-box jailbreak attacks. The finetuned models also exhibit improved truthfulness while largely preserving general capabilities. Comprehensive experiments across four models from three vendors demonstrate consistent improvements in safety robustness, reducing attack success rates by up to 90% for some attacks.

Tema degli esploratori: agenti e intelligenza artificiale (elenco arXiv cs.AI, nuovi lavori del giorno).
Categorie arXiv: Artificial Intelligence (cs.AI) ; Cryptography and Security (cs.CR).
Pagina del lavoro: https://arxiv.org/abs/2611.06553 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito).
Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina.
Letto il 2026-11-05T18:12:49.014Z dalla pagina https://arxiv.org/list/cs.AI/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).