PLUME: Parameter-Efficient Personalization of Large Language Models via Low-Rank User Modulation in Shared Subspaces

The researchers proposed the PLUME framework, aimed at achieving efficient personalization of large language models through low-rank user modulation in shared subspaces. The framework first learns global task subspaces from aggregated data, and then trains lightweight small matrices within those subspaces to perform personalized adaptation while keeping the shared components fixed. Additionally, the introduction of cross-layer shared parameters and rank-1 residual terms further reduces redundancy and maintains expressive power. In multiple personalized text generation benchmarks, PLUME’s performance was comparable to or better than that of strong baselines, proving that shared subspace modulation based on minimum residuals is a scalable and semantically grounded approach for LLM personalization.