You Snooze You Lose: Neural Weight Translation Restores Safety Alignment in Fine-Tuned LLMs Without Losing Domain Knowledge

arXiv·medium signal

Third-party LoRA adapters often destroy base model safety alignment, but safety fine-tuning destroys the domain knowledge the adapter was built for. This paper proposes neural weight translation to restore safety guardrails without the opposing failure mode. Directly relevant to anyone deploying open-source models with community adapters — a growing attack surface as the LoRA ecosystem expands.