Enhancing Jailbreak Resistance in Large Language Models Using Model Merge

In recent years, large language models (LLMs) have demonstrated their utility across a wide range of fields. At the same time, however, concerns have been raised regarding security risks and potential vulnerabilities. Although fine-tuning (FT) has been widely adopted as a conventional counter-measure, the computational costs involved in retraining present a significant burden in practical settings. To address this challenge, the present study proposes leveraging model merge as a low-cost strategy for enhancing the security robustness of LLMs. Model merge is a technique that integrates the capabilities of multiple models to reinforce specific attributes. Here, we create models that have been FT to improve security, and we then combine them in an appropriate manner to efficiently construct LLMs with strengthened safeguards—a process analogous to applying security patches. Our method aims to reduce computational overhead while simultaneously increasing robustness, in comparison with traditional FT-based security measures. Through experiments, we validate the effectiveness of our proposed approach. The results indicate that employing model merge significantly enhances resistance to jailbreak attacks in large-scale language models, with minimal degradation in overall performance. We consider these findings to suggest new possibilities for improving security through model merge.