SPRInG: Continual LLM Personalization via Selective Parametric Adaptation and Retrieval-Interpolated Generation

TL;DR

This work tackles the challenge of continually personalizing LLM outputs as user preferences drift over time. It introduces SPRInG, a semi-parametric framework that combines drift-driven selective adaptation of a user-specific LoRA adapter with a residual replay buffer, and a gated retrieval-augmented generation strategy that fuses parametric and non-parametric knowledge at the logit level. Through training-time drift scoring and selective updates, plus inference-time relevance gating and logit interpolation, SPRInG achieves state-of-the-art performance on long-form personalized generation tasks, outperforming strong continual-learning baselines and retrieval-based methods. The approach demonstrates robust handling of noisy interaction streams, improved stability, and practical scalability, with implications for real-world, long-horizon personalized LLM deployments.

Abstract

Personalizing Large Language Models typically relies on static retrieval or one-time adaptation, assuming user preferences remain invariant over time. However, real-world interactions are dynamic, where user interests continuously evolve, posing a challenge for models to adapt to preference drift without catastrophic forgetting. Standard continual learning approaches often struggle in this context, as they indiscriminately update on noisy interaction streams, failing to distinguish genuine preference shifts from transient contexts. To address this, we introduce SPRInG, a novel semi-parametric framework designed for effective continual personalization. During training, SPRInG employs drift-driven selective adaptation, which utilizes a likelihood-based scoring function to identify high-novelty interactions. This allows the model to selectively update the user-specific adapter on drift signals while preserving hard-to-learn residuals in a replay buffer. During inference, we apply strict relevance gating and fuse parametric knowledge with retrieved history via logit interpolation. Experiments on the long-form personalized generation benchmark demonstrate that SPRInG outperforms existing baselines, validating its robustness for real-world continual personalization.

Figures (8)

• Figure 1: Comparison between static and continual personalization.(Top) Static personalization assumes time-invariant preferences, using a frozen adapter trained on a fixed history snapshot. (Bottom) Continual personalization models users as dynamic entities. The framework iteratively updates the adapter at sequential time steps ($t=1, 2, 3, \cdots$) to capture preference drifts (shifting colors) while maintaining historical context.
• Figure 2: An overview of SPRInG.Training stage: incoming interactions are scored by comparing likelihoods between the base and adapted models. High-scoring drift samples trigger a parametric update, while samples that remain poorly internalized are stored in the replay buffer through re-evaluation with respect to the updated model. Inference stage: for a new query, the model computes parametric logits ($p_{\texttt{par}}$) from the adapter and non-parametric logits ($p_{\texttt{ret}}$) via gated retrieval. The final output is generated through logit interpolation, dynamically weighting the two distributions to balance internal knowledge with retrieved context.
• Figure 3: Interaction scalability of SPRInG. Performance comparison on different interaction history lengths.
• Figure 4: Correlation between quality and novelty terms. The scatter plots show the relationship between the Quality Term and the Novelty Term. Spearman’s $\rho$ is $0.6647$ for Abstract Generation and $0.3204$ for Review Writing.
• Figure 5: Correlation between hardness and novelty terms. The scatter plots show the relationship between the Hardness Term and the Novelty Term. Spearman’s $\rho$ is $-0.2361$ for Abstract Generation and $0.3468$ for Review Writing.