Safety at One Shot: Patching Fine-Tuned LLMs with A Single Instance
Fuente:
arXiv
Salvato in:
| Autori principali: | , , , , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2026
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866912805086560256 |
|---|---|
| author | Zhang, Jiawen He, Lipeng Chen, Kejia Lou, Jian Liu, Jian Yang, Xiaohu Jia, Ruoxi |
| author_facet | Zhang, Jiawen He, Lipeng Chen, Kejia Lou, Jian Liu, Jian Yang, Xiaohu Jia, Ruoxi |
| contents | Fine-tuning safety-aligned large language models (LLMs) can substantially compromise their safety. Previous approaches require many safety samples or calibration sets, which not only incur significant computational overhead during realignment but also lead to noticeable degradation in model utility. Contrary to this belief, we show that safety alignment can be fully recovered with only a single safety example, without sacrificing utility and at minimal cost. Remarkably, this recovery is effective regardless of the number of harmful examples used in fine-tuning or the size of the underlying model, and convergence is achieved within just a few epochs. Furthermore, we uncover the low-rank structure of the safety gradient, which explains why such efficient correction is possible. We validate our findings across five safety-aligned LLMs and multiple datasets, demonstrating the generality of our approach. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2601_01887 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Safety at One Shot: Patching Fine-Tuned LLMs with A Single Instance Zhang, Jiawen He, Lipeng Chen, Kejia Lou, Jian Liu, Jian Yang, Xiaohu Jia, Ruoxi Machine Learning Artificial Intelligence Fine-tuning safety-aligned large language models (LLMs) can substantially compromise their safety. Previous approaches require many safety samples or calibration sets, which not only incur significant computational overhead during realignment but also lead to noticeable degradation in model utility. Contrary to this belief, we show that safety alignment can be fully recovered with only a single safety example, without sacrificing utility and at minimal cost. Remarkably, this recovery is effective regardless of the number of harmful examples used in fine-tuning or the size of the underlying model, and convergence is achieved within just a few epochs. Furthermore, we uncover the low-rank structure of the safety gradient, which explains why such efficient correction is possible. We validate our findings across five safety-aligned LLMs and multiple datasets, demonstrating the generality of our approach. |
| title | Safety at One Shot: Patching Fine-Tuned LLMs with A Single Instance |
| topic | Machine Learning Artificial Intelligence |
| url | https://arxiv.org/abs/2601.01887 |