Aplaud: Adaptive Personalized Low-Rank Decomposition for User-Specific LLM

To address challenges such as limited training data for individual users, scalability of model storage, and utilization of shared structures across problems, the research team proposed the Aplaud framework for personalized survey response prediction in user-specific large language models. This framework expands the adaptation process into two parts: freezing the shared low-rank basis and compacting user-specific modifications, supplemented by rank-one residuals for more refined personalization. By further decomposing the modification matrix into lower-rank forms to reduce parameter costs and mitigate overfitting, Aplaud achieves efficient and scalable personalized modeling. Empirical results show that this method outperforms state-of-the-art LoRA-based personalized LLM approaches in terms of generalization and inference efficiency.