Continual Learning for Sequential Personalization of Small Language Models: A Stability Monitoring Analysis

A new preprint examines how small language models behave when personalized sequentially, finding that standard task metrics can mask harmful degradation and that lightweight distributional diagnostics can reveal hidden instability patterns. The study, submitted to the arXiv preprint repository on 26 June 2026 by Thomas Da Silva Paula, investigates small language models (SLMs) in a continual learning setting [1]. SLMs are increasingly targeted for deployment on edge devices such as laptops, where they can enable private, low-latency applications [2]. Personalization on these devices requires models to adapt over time to evolving user- or task-specific data, a process that places them in a continual learning framework [2]. This adaptation carries the risk of catastrophic forgetting, in which learning new information degrades performance on previously learned tasks or broader model capabilities [2]. Recent benchmarks, including TRACE, have already demonstrated that continual fine-tuning can significantly erode the general abilities of aligned large language models [2]. The new work focuses specifically on sequential Low-Rank Adaptation, or LoRA, personalization of SLMs [1]. The researcher saved model checkpoints after each adaptation stage and evaluated them on current tasks, previously seen tasks, and a fixed reference set [2]. This checkpoint-level protocol allowed monitoring of task performance, forgetting, and reference set drift over time [2]. The paper reports that lightweight reference set distributional diagnostics can reveal model-specific instability patterns during sequential LoRA personalization [1]. Notably, the diagnostics exposed cases where task-level metrics alone hid harmful adaptation [2]. The submission, weighing 130 KB, appears on arXiv, an open-access repository of electronic preprints that is moderated but not peer-reviewed [1][8]. arXiv was founded in 1991 and by late 2024 was receiving about 24,000 articles per month [8]. The platform also hosts arXivLabs, a framework for community-built tools that appear on article record pages, including bibliographic explorers and code linkers [7][6]. The study's findings suggest new research avenues for monitoring SLM stability in continual learning settings [1].