Safety at One Shot: Patching Fine-Tuned LLMs with A Single Instance

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.

Key Contributions

• Demonstrates that LLM safety alignment can be fully recovered using only a single safety example, regardless of the number of harmful fine-tuning examples or model size
• Uncovers the low-rank structure of the safety gradient, providing a theoretical explanation for why one-shot realignment is effective
• Validates the approach across five safety-aligned LLMs and multiple datasets with convergence in just a few epochs

🛡️ Threat Analysis

Details

Similar Papers