This work tackles personalized survey‑response prediction with fine‑tuned large language models (LLMs). The task is challenged by scarce per‑user training data, the need for scalable model storage, and the desire to exploit shared structure across questionnaire items. To meet these demands we introduce Aplaud (Adaptive Personalized Low‑rank and User‑specific Nested Decomposition). Building on the LoRA paradigm, adaptation is split into a frozen shared low‑rank basis $B$, a compact user‑specific correction $C$, and an additional rank‑one residual $r$ for finer personalization. The correction matrix $C$ can itself be factorized into an even lower‑rank form, reducing per‑user parameter cost and mitigating overfitting. Empirical results show that Aplaud delivers efficient, scalable personalization across users while surpassing state‑of‑the‑art LoRA‑based personalized LLM methods in both generalization performance and inference efficiency.
Review