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

2026-01-05

arXiv.org

AI Summary

This work addresses the critical issue that fine-tuning safety-aligned large language models (LLMs) often leads to a significant degradation in their safety guarantees. Existing remediation approaches typically require extensive safe training samples or calibration sets, incurring high computational costs and compromising model utility. To overcome these limitations, we propose an efficient repair method that leverages only a single safe example to fully restore model safety within just a few training epochs, without sacrificing general performance. We further provide the first theoretical insight by revealing that safety-related gradients exhibit a low-rank structure, which underpins the efficacy of single-sample repair. Extensive experiments across five mainstream safety-aligned LLMs and multiple datasets demonstrate that our approach consistently and rapidly recovers safety across varying model scales and proportions of harmful examples.

Technology Category

• Natural Language Processing: Safety and Robustness
• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Large Vision Models

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Security and Privacy: Large-scale security measurements
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

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.

Problem

• safety alignment
• fine-tuning
• large language models
• computational overhead
• model utility

Innovation

• one-shot safety alignment
• low-rank gradient
• fine-tuned LLMs
• safety recovery
• minimal calibration