Personalizing large language models is essential for delivering AI assistance that matches each user's style, intent, and preferences. Conventional per‑user fine‑tuning yields strong results but incurs huge parameter and storage costs, making it hard to scale to massive user bases.
PLUME (Personalized Low‑Rank Adaptation through User Modulation and Shared Subspace) proposes low‑rank modulation within a shared task subspace. First, a global task subspace is learned from aggregated user data. Then, for each user only a small square matrix inside this subspace is trained while all other shared components remain fixed, yielding a tailored model.
To further cut redundancy, PLUME introduces cross‑layer shared parameters and adds rank‑1 residual terms to each layer. These mechanisms preserve expressiveness yet reduce per‑user extra parameters by over 95%.
Experiments on several personalized text‑generation benchmarks show that PLUME matches or surpasses strong baselines while dramatically lowering parameter overhead. The results confirm that shared‑subspace modulation with minimal residuals is both scalable and semantically grounded.
Review: PLUME’s combination of a global subspace and low‑rank user matrices offers an efficient, interpretable path to personalization, providing a practical solution for large‑scale LLM deployment.