Local LLMs as a plasticity layer for fixed cloud LLMs

Abstract The central role of local LLMs is not merely to approximate cloud models at a smaller scale, but also to provide a practical user-side implementation of persistent adaptation that is difficult to support universally within shared cloud services. The contribution is an architectural reframing of where user-specific adaptation should be implemented, rather than the proposal of a new component-level learning mechanism. This reframing shifts the design question from model capability to the placement of memory, preference updating, constraint formation, and continual adaptation in hybrid AI systems.