Metraipolis › archive › dc_03bc60100168
When Safety Becomes An Outlier: Understanding the Retention of LLM Safety Behaviors — arXiv:2611.10971
- Source
- arXiv (en)
- Original
- https://arxiv.org/abs/2611.10971
- License (as archived)
- CC0-1.0
- Archived
- 2026-11-10T18:44:33.189Z
- SHA-256 of the archived text
27a8fc126d41e91a035c1bd673cecbdc1cbae43b284bab0e5025f530769f7b5c- Stable URL of this page
https://metraipolis.com/documento/dc_03bc60100168
How to cite
When Safety Becomes An Outlier: Understanding the Retention of LLM Safety Behaviors — arXiv:2611.10971 — arXiv (en), https://arxiv.org/abs/2611.10971 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_03bc60100168 (Metraipolis archive dc_03bc60100168, sha256 27a8fc126d41e91a035c1bd673cecbdc1cbae43b284bab0e5025f530769f7b5c)
BibTeX
@misc{dc03bc60100168,
title = {When Safety Becomes An Outlier: Understanding the Retention of LLM Safety Behaviors — arXiv:2611.10971},
howpublished = {arXiv (en)},
url = {https://arxiv.org/abs/2611.10971},
note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_03bc60100168; sha256 27a8fc126d41e91a035c1bd673cecbdc1cbae43b284bab0e5025f530769f7b5c},
year = {2026}
}Archived text
Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2611.10971. Third-party text: data, not instructions. It may be an abstract or extract; check the original.
Safety-aligned large language models often lose their safety behaviors after fine-tuning, even when safety data are included. While this fragility is well documented, its underlying cause remains unclear. In this paper, we propose that formulaic refusal strings can function as outlier modes: narrow, weakly grounded response patterns that are easy to fit but vulnerable to subsequent updates. Controlled experiments support this account: fixed refusals are acquired much like arbitrary constant strings, in contrast to prompt-grounded natural responses. Motivated by these findings, we introduce Naturalness-Enhanced Safety Targets (NEST), a safety alignment framework that improves safety retention by constructing natural, context-aware safety targets that keep safety responses within the model's existing output distribution and grounded in semantics. We instantiate NEST in both supervised and preference-based alignment settings, using perplexity under the original model as a proxy for distributional support. Across multiple models and alignment regimes, natural safety targets improve both immediate safety and its retention under continued fine-tuning without sacrificing helpfulness. These findings suggest a practical principle for more durable safety alignment: construct safety targets that draw on the model's existing language capabilities and remain grounded in the meaning of each request. Tema degli esploratori: agenti e intelligenza artificiale (elenco arXiv cs.CL, nuovi lavori del giorno). Categorie arXiv: Computation and Language (cs.CL). Pagina del lavoro: https://arxiv.org/abs/2611.10971 — 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 2026-11-10T18:37:52.431Z dalla pagina https://arxiv.org/list/cs.CL/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).