stateful personalization modeling

This paper addresses the core challenge of synergistic evolution between Retrieval-Augmented Generation (RAG) and Large Language Model (LLM)-based agents in personalized AI. Methodologically, it proposes a three-stage personalized RAG framework—pre-retrieval, retrieval, and generation—and formally defines, for the first time, a dual-track personalization paradigm unifying RAG and agent architectures within a coherent analytical framework. It characterizes the transition from static retrieval augmentation to dynamic, agent-driven personalization, integrating user modeling, multi-step reasoning planning, LLM fine-tuning, and interpretable evaluation design. The work systematically surveys state-of-the-art approaches, catalogs mainstream datasets and evaluation benchmarks, and identifies key open challenges—including interpretability, long-term memory, and privacy preservation. Concurrently, it maintains an authoritative open-source knowledge repository, providing both theoretical foundations and practical guidance for advancing personalized AI research.

This study addresses the lack of systematic theoretical frameworks accounting for the pronounced individual differences observed in user interactions with AI chatbots. It pioneers the extension of Bronfenbrenner’s bioecological systems theory into the domain of personalized artificial intelligence, proposing a user-centered heuristic analytical framework that integrates multilevel contextual factors. At its core, the framework conceptualizes the reciprocal, iterative, and co-adaptive exchanges between users and memory-endowed AI agents as a dynamic developmental process. Through theoretical modeling, human–AI interaction analysis, and contextual systems mapping, the project formulates a structured explanatory model of variability, elucidating the mechanisms underlying the co-evolution of human–AI relationships. This approach offers both theoretical grounding and practical insights for the design and study of adaptive, context-aware AI systems.

研究通过对比非个性化与两种个性化语言模型在五天内对992名参与者的影响，探讨了个性化语言模型对用户自我认知及人际交往的长期效应。

This work addresses the challenge of effectively leveraging long-term first-person perspective data to deliver personalized assistive services by proposing EgoSelf, a graph-structured episodic memory system. EgoSelf employs graph neural networks to model the temporal and semantic relationships inherent in users’ historical behaviors, thereby constructing personalized behavioral profiles. The system is trained end-to-end with the objective of predicting future interaction tasks. As the first framework to systematically learn user-specific representations from extended first-person data, EgoSelf significantly outperforms existing methods across multiple personalized assistance benchmarks. The implementation has been made publicly available.

This study addresses critical limitations in existing user modeling approaches within model-driven engineering, including fragmentation, incomplete dimensional coverage, and weak dynamic evolution capabilities, compounded by a lack of effective tool support. To overcome these challenges, the authors propose a low-code-driven unified user modeling framework that integrates multidimensional domain knowledge to construct reusable user models. Leveraging machine learning, the framework enables dynamic, incremental updates and automatic adaptation of user profiles. Grounded in a systematic literature review, user behavior analysis, and automated pipeline technologies, this work not only identifies current modeling shortcomings but also outlines a technical roadmap for developing dynamic, comprehensive, and automated personalized dialogue systems, thereby establishing a novel paradigm for efficient personalization in intelligent agents.

This study investigates how implicit model updates in AI systems affect users’ perception, cognition, and behavior—specifically in historical figure face recognition—where such updates often occur without user awareness or consent. Method: We employed a dual-method design: (1) an A/B online experiment and (2) a longitudinal field diary study, complemented by behavioral logging, qualitative interviews, and comparative analysis using a FaceNet variant. Contribution/Results: We first empirically identify a significant perceptual blind spot: 62% of users failed to detect minor model updates. Unaware of these changes, users developed biased “folk theories” about system behavior, leading to diminished trust, misaligned interaction strategies, and reduced task completion rates. Based on these findings, we propose a transparency framework for human-AI co-evolution, supporting sustainable, explainable, and trustworthy AI updates. The framework advances theoretical understanding and provides actionable design principles for long-term AI system usability and update governance.

This project addresses the limitation of single-model analyses in capturing system-level privacy risks within personalized AI by departing from traditional component-level paradigms to establish privacy as an emergent system property. Through systematic architectural analysis and privacy risk modeling, it identifies four distinct risk channels and constructs a multidimensional evaluation framework encompassing interaction trajectories, information flows, indirect leakage, and utility-privacy trade-offs. Ultimately, this work delivers actionable, systematized privacy auditing standards and an assessment methodology that effectively bridges critical gaps in existing audit approaches at the system level.

This study addresses the failure of AI model iterations to preserve established user value, resulting in ineffective technological substitution. Using Keep4o as a case study, we analyzed 60,000 social media posts through systematic coding and large language model-assisted analysis. The findings reveal users’ profound attachment to interactional and relational values, demonstrating that technical upgrades do not equate to effective replacement. This work provides the first empirical evidence from the user perspective of value continuity dilemmas during model iteration. Consequently, we propose integrating user experience assessment into model lifecycle governance frameworks. This approach offers both theoretical grounding and practical pathways for balancing technological advancement with the preservation of user value, ensuring that AI evolution remains aligned with human-centric needs rather than purely technical metrics.

Existing black-box auditing methods struggle to disentangle user attributes from behavior while preserving behavioral authenticity, limiting causal understanding of personalization algorithms. This study introduces the first large-scale algorithmic audit using generative AI agents: leveraging real survey data, it constructs synthetic accounts with fixed personas that autonomously interact on the X platform, enabling counterfactual experiments through controlled manipulation of visible user attributes. This approach effectively decouples attributes from behavior, supporting fine-grained, scalable, and behaviorally authentic analysis. An empirical deployment of 1,120 agents reveals that the platform’s algorithms significantly amplify toxic, polarizing, and right-leaning political content, with amplification effects varying by user ideology; furthermore, the influence of demographic signals exhibits substantial heterogeneity across subpopulations.

本文提出了一种名为UMPeek的黑盒攻击方法，通过假设引导的自适应探测来推断隐藏的用户模型，从而揭示了即使在源记录和后端状态不可访问的情况下，个性化LLM代理仍可能泄露隐私信息的问题。

研究通过实验探讨了基于风格和情境的个性化在AI辅助决策中的影响，发现情境个性化更显著改变用户决策行为。