AsFT: Anchoring Safety During LLM Fine-Tuning Within Narrow Safety Basin
Proceedings of the AAAI Conference on Artificial Intelligence · 14.03.2026 · 40 (40) · 34322-34330
DOI: 10.1609/aaai.v40i40.40729 ↗Kurz gefasst
Fachartikel in der Zeitschrift Proceedings of the AAAI Conference on Artificial Intelligence, online erschienen am 14.03.2026. Verfasst von Shuo Yang und 10 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref).
Abstract (Originalsprache)
Fine-tuning large language models (LLMs) improves performance but introduces critical safety vulnerabilities: even minimal harmful data can severely compromise safety measures. We observe that perturbations orthogonal to the alignment direction—defined by weight differences between aligned (safe) and unaligned models—rapidly compromise model safety. In contrast, updates along the alignment direction largely preserve it, revealing the parameter space as a "narrow safety basin". To address this, we propose AsFT (Anchoring Safety in Fine-Tuning) to maintain safety by explicitly constraining update directions during fine-tuning. By penalizing updates orthogonal to the alignment direction, AsFT effectively constrains the model within the "narrow safety basin," thus preserving its inherent safety. Extensive experiments on multiple datasets and models show that AsFT reduces harmful behaviors by up to 7.60%, improves task performance by 3.44%, and consistently outperforms existing methods across multiple tasks.
Zitieren
APA
Yang, S., Zhang, Q., Liu, Y., Huang, Y., Jia, X., Ning, K., Yao, J., Wang, J., Hailiang, D., Song, Y., & Yuan, L. (2026). AsFT: Anchoring Safety During LLM Fine-Tuning Within Narrow Safety Basin. Proceedings of the AAAI Conference on Artificial Intelligence, 40(40), 34322-34330. https://doi.org/10.1609/aaai.v40i40.40729
Vancouver
Yang S, Zhang Q, Liu Y, Huang Y, Jia X, Ning K, et al. AsFT: Anchoring Safety During LLM Fine-Tuning Within Narrow Safety Basin. Proceedings of the AAAI Conference on Artificial Intelligence. 2026;40(40):34322-34330. doi:10.1609/aaai.v40i40.40729
BibTeX
@article{Yang2026,
author = {Yang, Shuo and Zhang, Qihui and Liu, Yuyang and Huang, Yue and Jia, Xiaojun and Ning, Kun-Peng and Yao, Jia-Yu and Wang, Jigang and Hailiang, Dai and Song, Yibing and Yuan, Li},
title = {AsFT: Anchoring Safety During LLM Fine-Tuning Within Narrow Safety Basin},
journal = {Proceedings of the AAAI Conference on Artificial Intelligence},
year = {2026},
volume = {40},
number = {40},
pages = {34322--34330},
doi = {10.1609/aaai.v40i40.40729}
}
EUVIMED