![[2602.12394] Synthetic Interaction Data for Scalable Personalization in Large Language Models](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.12394] Synthetic Interaction Data for Scalable Personalization in Large Language Models
Summary
The paper introduces PersonaGym, a framework for generating synthetic interaction data to enhance personalization in large language models (LLMs). It addresses the limitations of existing prompt optimization methods by modeling dynamic user preferences and providing a scalable...
Why It Matters
As large language models become integral to various applications, effective personalization is crucial for user satisfaction. This research addresses the challenges of data scarcity and user-specific preferences, offering a novel approach that could significantly improve LLM interactions in real-world scenarios.
Key Takeaways
- PersonaGym generates high-fidelity synthetic data for personalized user interactions.
- The framework models dynamic user preferences, enhancing the realism of interactions.
- Personalized Prompt Optimization (PPOpt) improves prompt effectiveness without altering LLMs.
- Extensive experiments show significant improvements in personalization quality and robustness.
- The research addresses critical gaps in existing personalization methods for LLMs.
Computer Science > Machine Learning arXiv:2602.12394 (cs) [Submitted on 12 Feb 2026] Title:Synthetic Interaction Data for Scalable Personalization in Large Language Models Authors:Yuchen Ma, Yue Huang, Wenjie Wang, Xiaonan Luo, Xiangliang Zhang, Stefan Feuerriegel View a PDF of the paper titled Synthetic Interaction Data for Scalable Personalization in Large Language Models, by Yuchen Ma and 5 other authors View PDF HTML (experimental) Abstract:Personalized prompting offers large opportunities for deploying large language models (LLMs) to diverse users, yet existing prompt optimization methods primarily focus on task-level optimization while largely overlooking user-specific preferences and latent constraints of individual users. This gap is primarily due to (i) the absence of high-quality, privacy-sensitive data that capture personalized user-LLM interactions at scale, and (ii) the lack of robust reward signals for individual preferences. To overcome existing data limitations, we introduce a high-fidelity synthetic data generation framework called PersonaGym. Unlike prior work that treats personalization as static persona-preference pairs, PersonaGym models a dynamic preference process via an agentic LLM system to simulate realistic preference behaviors and semantic-aware noise in order to generate personalized multi-turn interaction trajectories. Using PersonaGym, we release PersonaAtlas, a large-scale, high-quality, and diverse synthetic dataset of high-fidelity multi-t...
