personalization

This work addresses the persistent gap between offline evaluation and online performance in recommender systems, which stems largely from existing large language model (LLM)-based user simulators neglecting critical contextual factors such as time, location, and user intent. To bridge this gap, the authors propose ContextSim, a novel framework that integrates real-life contextual dynamics into LLM-powered user agents for the first time. ContextSim employs a life-simulation module to generate daily scenarios enriched with temporal, spatial, and motivational cues, and enforces both internal chain-of-thought reasoning and behavioral–trajectory consistency constraints to produce context-aware agents that more faithfully emulate human interactions. Experimental results demonstrate that interactions synthesized by ContextSim closely mirror real user behavior, yielding offline A/B test outcomes highly correlated with live metrics; recommendation strategies optimized using this framework significantly enhance actual user engagement.

This study challenges the conventional explicit/implicit feedback dichotomy by introducing the novel concept of “intentional implicit feedback”—user-initiated, strategic implicit behaviors (e.g., deliberate skipping, repeated rewatching, cross-category searching) undertaken to optimize recommendations, yet not explicitly solicited or designed by the platform. Through in-depth interviews and thematic coding with 34 active users of platforms including Xiaohongshu and Douyin, the research identifies a purpose-driven taxonomy of such feedback behaviors. Results demonstrate that intentional implicit feedback significantly enhances recommendation diversity (+37%) and perceived relevance (self-reported improvement of 2.1 points on a 5-point scale). The findings advance theoretical understanding of human–algorithm co-adaptation in recommender systems and provide empirical grounding for designing intention-aware feedback interfaces that better capture user agency and strategic engagement.

Traditional recommender systems heavily rely on implicit signals such as clicks, often neglecting the rich semantic information embedded in explicit contextual feedback like reviews and ratings, which can lead to preference misalignment and filter bubbles. This work is the first to systematically underscore the pivotal role of explicit feedback in large language model (LLM)-driven recommendation and proposes a novel heterogeneous information modeling framework that deeply integrates user-generated textual signals—such as reviews—into a scalable recommendation architecture. The contributions include a new framework, a dedicated evaluation benchmark, and tailored metrics, collectively enhancing recommendation accuracy, diversity, and explainability. This study establishes a new paradigm for next-generation recommender systems that are both transparent and highly personalized.

Existing evaluations of personalized AI agents for information retrieval rely heavily on static benchmarks, failing to capture the temporal evolution of user needs and long-term interactive adaptability. To address this, we propose a novel dynamic, interaction-aware paradigm for longitudinal evaluation: (1) a time-varying preference modeling-based persona simulator; (2) a reference-interview-driven protocol for structured preference elicitation; and (3) cross-session behavioral adaptability metrics. Our method integrates large language model–driven user simulation, preference elicitation, and longitudinal interaction analysis. We conduct empirical evaluation in the e-commerce search setting using the PersonalWAB dataset. This work is the first to systematically define and empirically validate an evaluation framework for personalized agents’ *sustained adaptability*—a core capability for long-term user alignment. It establishes a reproducible, scalable theoretical and practical foundation for user-centered, longitudinal interaction optimization.

This work addresses the insufficient collaboration and cumulative bias arising from disjoint modeling of recommender and user agents in recommendation systems. We propose the first collaborative optimization framework explicitly designed for a dual-agent closed-loop feedback paradigm. Methodologically, we establish a bidirectional iterative feedback mechanism: the recommender agent generates recommendations and observes responses from the user agent, while the user agent dynamically refines its preference representation based on feedback; both agents co-evolve via an LLM-driven, interpretable interaction protocol. Our key contribution is the first formalization of the recommender–user dual-agent closed-loop feedback process, jointly optimizing recommendation quality and mitigating bias—without exacerbating popularity or positional biases. On three benchmark datasets, our approach achieves average improvements of 11.52% in recommendation accuracy over a recommender-only baseline and 21.12% over a user-only baseline, significantly enhancing fidelity in user behavior modeling.

Existing agent evaluation frameworks overlook experience-driven adaptive learning and reasoning capabilities in dynamic environments, particularly in multi-turn natural language dialogue for product recommendation. Method: We introduce BELA, the first benchmark for context-aware experiential learning, integrating real-world Amazon product data, structured user profiles, and a large language model–driven user simulator to systematically assess agents’ active exploration, continual learning, and adaptive decision-making. Contribution/Results: Experiments reveal that state-of-the-art large language models fail to improve performance across dialogue turns, exposing fundamental limitations in contextual accumulation, preference evolution modeling, and policy iteration. BELA establishes a reproducible, scalable paradigm for evaluating and advancing long-term agent adaptability in interactive, evolving settings.

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 work addresses the limitation of existing large language model (LLM) personalization approaches, which typically employ flat behavioral modeling and overlook the underlying structural complexity of user behavior. To overcome this, the study introduces Bourdieu’s theory of practice into LLM personalization for the first time, proposing the PHF framework that hierarchically models user behavior through three interrelated dimensions: practice, habitus, and field. This enables more structured and interpretable personalization. The framework implements a lightweight, model-agnostic PHF_Compass module that learns hierarchical behavioral representations while keeping the base LLM frozen. Evaluated on the LaMP benchmark, PHF consistently enhances performance across multiple tasks, demonstrating both the generalizability and interpretability of the learned behavioral structures.

This work addresses the degradation in recommendation performance for cold-start users in new scenarios, which stems from sparse user behavior, low engagement, and model instability. To tackle this challenge, the authors propose an end-to-end scene-aware recommendation framework that jointly designs feature engineering, model architecture, and a stable online updating mechanism. By leveraging cross-scenario feature extraction and knowledge transfer, the framework enables effective modeling of new users in novel contexts. Evaluated on a billion-scale user product migration task at Microsoft, the approach demonstrates significant improvements over existing methods in both offline and online experiments, substantially enhancing the accuracy and robustness of cold-start recommendations.

To address insufficient user intent representation in implicit-feedback recommendation and semantic biases introduced by large language models (LLMs) in profile generation, this paper proposes ProEx—a unified framework. Methodologically, ProEx introduces: (i) a chain-of-thought prompting-based mechanism for multi-perspective user/item profiling to enhance intent coverage; (ii) joint semantic vector extrapolation and environment-invariant learning to disentangle intrinsic preference features from environmental noise; and (iii) collaborative discriminative and generative modeling to improve robustness. Extensive experiments on three public benchmarks demonstrate that ProEx consistently outperforms six state-of-the-art baselines, achieving average improvements of 12.7% in Recall@10 and 9.3% in NDCG@10. The framework exhibits both superior accuracy and stability, validating its effectiveness in mitigating semantic drift and environmental confounding in implicit-feedback settings.