VISA: Personalized LLM Alignment via Value Injection and Shielded Adaptation

Researchers propose the closed-loop framework VISA, which achieves personalized LLM alignment through value injection and shielded adaptation, mitigating the alignment tax induced by task fine-tuning. Its architecture includes a high-precision value detector, a semantic-to-value converter, and a value rewriter; the rewriter is trained with GRPO using a composite reward that balances fine-grained value precision and semantic integrity. Experiments show that VISA can precisely control value expression while maintaining factual consistency and general capabilities, significantly outperforming standard fine-tuning and prompting baselines such as GPT-4o.