VISA: Value Injection via Shielded Adaptation for Personalized LLM Alignment

ABSTRACT

Aligning large language models to individual user values without compromising the core safety parameters of the foundation model is notoriously difficult. This paper introduces VISA, a shielded adaptation method that injects personalized values into distinct, isolated adaptation layers. This allows users to heavily customize the agent's ethical and stylistic behavior while a static safety shield prevents catastrophic jailbreaks.