This research addresses the critical challenge of alignment loss that occurs when fine-tuning large language models for specific applications.

• Identifies that fine-tuning can inadvertently increase harmful response rates by up to 30%
• Proposes Multi-view Parameter Fine-tuning (MPF) that selectively updates parameters to maintain alignment
• Demonstrates how to preserve ethical guardrails while optimizing for task performance
• Shows successful application across multiple model sizes and fine-tuning scenarios

For security professionals, this research provides practical techniques to ensure AI systems remain safe and aligned with human values even after customization for specific business needs.

Alleviating the Fear of Losing Alignment in LLM Fine-tuning