SafeMERGE is a novel framework that addresses the critical challenge of safety erosion when fine-tuning large language models for specific applications.

• Selectively merges layers between fine-tuned and safety-aligned models based on cosine similarity measures
• Preserves safety alignment while maintaining task performance
• Outperforms existing methods like PEFT and full fine-tuning in safety-utility balance
• Requires no additional training or special fine-tuning procedures

For medical applications, this approach enables safer deployment of specialized LLMs that maintain ethical guardrails while providing domain-specific expertise, reducing potential harm from biased or dangerous outputs in healthcare settings.

Original Paper: SafeMERGE: Preserving Safety Alignment in Fine-Tuned Large Language Models via Selective Layer-Wise Model Merging