SafeGene introduces reusable adapters to preserve LLM safety during fine‑tuning
The paper presents SafeGene, a reusable safety‑adapter module that can be applied across tasks and model families to counteract safety degradation when open‑weight LLMs are fine‑tuned. By decoupling safety alignment from task updates via layer‑wise coefficient recalibration, SafeGene reduces harmful responses while maintaining downstream performance.