Adaptive Dual Distillation for On-Device Personalized Language Models: A Conceptual Framework and Feasibility Protocol

We propose a conceptual framework called ADDT that integrates general knowledge compression with local personalization in a single architecture executable on consumer devices. The framework integrates five components: knowledge distillation, preference distillation, style distillation, a retrieval-based memory layer, and continuous updates. We define digital twin in a narrow functional sense (an alignment tool, not an identity representation), and present UIS as an exploratory tool under validation. This paper is targeted at specialized workshops or as an arXiv preprint, not as a results paper for a main track.