Dynamic Outlier-Aware Personalization: Detecting and Leveraging Deviant User Behaviors in LLM Alignment

Most personalization approaches, such as the OPPU framework (Tan et al., 2024), focus on modeling user behavior shifts but typically treat outlier behaviors as either noise or anomalies to be smoothed over. However, these deviations may actually signal evolving needs or emerging user contexts. This research proposes a novel outlier-aware personalization module within LLMs—one that explicitly identifies significant behavioral deviations (via statistical or anomaly detection methods) and treats them as first-class signals for rapid, context-sensitive adaptation. By building atop personalized PEFT methods (Tan et al., 2024; Kim et al., 2025) but diverging in how deviations are handled, this approach could better capture users’ dynamic needs, particularly in high-stakes settings (e.g., mental health, learning, or crisis support). The impact would be more responsive, contextually aware LLMs and a new theoretical framework for understanding outlier behaviors as drivers of personalization, not just noise.

References:

1. Democratizing Large Language Models via Personalized Parameter-Efficient Fine-tuning. Zhaoxuan Tan, Qingkai Zeng, Yijun Tian, Zheyuan Liu, Bing Yin, Meng Jiang (2024). Conference on Empirical Methods in Natural Language Processing.
2. Personalized Language Models via Privacy-Preserving Evolutionary Model Merging. Kyuyoung Kim, Jinwoo Shin, Jaehyung Kim (2025). arXiv.org.