Publications

Combining Domain and Alignment Vectors Provides Better Knowledge-Safety Trade-offs in LLMs