User Modeling in the Era of Large Language Models: Current Research and Future Directions

• User modeling extracts insights from user behavior to adapt systems to individual needs across e-commerce, entertainment, and social networks.Source: §1 · p. 1
• User data mainly comprise UGC text and interaction graphs, motivating distinct text-based and graph-based modeling approaches.Source: §1 · pp. 1–2
• Text-based modeling progressed from BoW and topic models to embeddings, pretrained language models, and LLMs with stronger semantic and reasoning abilities.Source: §1 · pp. 2–3
• Graph-based modeling evolved from random walks and matrix factorization to Node2Vec and GNNs, which support recommendation, profiling, behavior prediction, and suspicious-user detection.Source: §1 · p. 3
• LLMs unify many text-to-label tasks through prompting, provide reasoning signals for graph prediction, and expand sparse user or item information with parametric knowledge.Source: §1 · p. 3
• The survey reviews LLM advantages, methodological taxonomies, application-specific techniques, and challenges to provide a comprehensive LLM-UM reference.Source: §1 · p. 4

2.1 User Modeling

• User modeling divides into text-based methods for UGC and graph-based methods for user-user or user-item interactions.Source: §2.1 · p. 4
• Text-based methods progressed from BoW and topic models through Word2Vec and BERT4Rec to LLMs for profiling, recommendation, personalization, and misbehavior detection.Source: §2.1 · pp. 5–6
• Graph-based methods progressed from PageRank, collaborative filtering, and matrix factorization to Node2Vec and message-passing GNNs such as GCN and GAT.Source: §2.1 · p. 6

2.2 Large Language Model

• LLMs are Transformer-based models with billions of parameters, commonly using encoder-only, decoder-only, or encoder-decoder designs.Source: §2.2 · pp. 6–7
• Scaling produces emergent in-context learning, instruction following, and step-by-step reasoning, alongside tool use, graph understanding, agency, and optimization capabilities.Source: §2.2 · p. 7
• Parameter-efficient fine-tuning adapts LLMs with fewer trainable parameters through methods including prefix tuning, prompt tuning, LoRA, and QLoRA.Source: §2.2 · p. 7
• LLMs are available as open-source or API-based models, differing in whether users can access model weights and run models locally.Source: §2.2 · p. 7
• Persistent limitations include hallucination, bias, high inference latency, limited context, outdated knowledge, misalignment, brittle evaluation, and weak structural understanding.Source: §2.2 · p. 8

• LLMs can model user interests, characteristics, and intentions from behavior history and UGC for recommendation and profiling.Source: §3 · p. 8
• Zero-shot LLMs lag behind traditional recommendation models for rating prediction when user interaction data are absent.Source: §3 · p. 9
• LLMs achieve impressive zero-shot results in personality recognition from UGC history.Source: §3 · p. 9
• Zero-shot prompting for hate speech detection can achieve performance comparable to or better than fine-tuned models.Source: §3 · p. 9
• Collectively, these studies demonstrate LLM capabilities for modeling, understanding, and reasoning over UGC and user behavior.Source: §3 · p. 9

4 Taxonomies

• The survey organizes LLM-UM with taxonomies based on methodology and downstream applications.Source: §4 · p. 9

4.1 Approach Perspective

• The approach taxonomy includes predictors, enhancers, controllers, and evaluators, based on the functionality assigned to LLMs.Source: §4.1 · p. 9
• LLMs-as-Predictors directly generate answers, whereas enhancers augment systems, controllers automate pipeline operations, and evaluators assess generated content or conversations.Source: §4.1 · p. 9

4.2 Application Perspective

• Personalization applications include profiling, recommendation, assistants, dialogue systems, and education or healthcare applications.Source: §4.2 · p. 10
• Suspiciousness detection applications cover fraud, discrimination, misinformation, and LLM-generated text detection.Source: §4.2 · p. 10

• LLM-UM categorizes LLM roles as predictors, enhancers, controllers, and evaluators.Source: §5 · p. 10
• Predictors generate predictions directly, enhancers augment user-modeling systems, controllers automate the modeling pipeline, and evaluators assess system performance.Source: §5 · p. 10
• The user model remains knowledge and patterns discovered from UGC and user-user or user-item interaction networks.Source: §5 · p. 10

5.1 LLMs-as-Predictors

• LLMs-as-Predictors directly make downstream predictions through generative reasoners, simulators or agents, classifiers, scorers, explainers, and chatbots.Source: §5.1 · p. 10

5.1.1 Common Generative Reasoner

• Generative reasoners use non-tuned prompting or fine-tuning, with PEFT methods supporting domain adaptation and retrieval-enhanced combinations improving performance.Source: §5.1.1 · pp. 10–12

5.1.2 Simulator/Agent

• Simulator and agent approaches imitate user behavior or perform user-modeling tasks through planning, memory, communication, and external tools.Source: §5.1.2 · p. 12

5.1.3 Classifier/Detector

• Classifier and detector approaches analyze UGC for stance, hate speech, suspicious behavior, mental health risks, intent, sentiment, and other social or health signals.Source: §5.1.3 · pp. 12, 14

5.1.4 Scoring Function

• Scoring functions rate or rank items and responses from user behavior history, supporting point-wise, pair-wise, and list-wise prediction.Source: §5.1.4 · p. 14

5.1.5 Explainer

• Explainers generate human-understandable rationales for recommendations, classifications, financial data, learning outputs, and social content.Source: §5.1.5 · p. 14

5.1.6 Chatbot

• Chatbots use user profiles, behavior history, and interaction to provide personalized recommendation, advisory, emotional-support, medical, and content-generation services.Source: §5.1.6 · pp. 14–15

5.2 LLMs-as-Enhancers

• LLMs-as-Enhancers use language models as plug-in modules for profiling, feature encoding, knowledge augmentation, and data generation.Source: §5.2 · p. 15

5.2.1 Profiler

• LLM-generated profiles summarize users’ characteristics, preferences, interests, and locations from heterogeneous behavior histories for downstream recommendation and prediction.Source: §5.2.1 · pp. 15–16

5.2.2 Feature Encoder

• LLM-encoded embeddings inject semantic understanding and open-world knowledge from user, item, query, job, and news text into downstream task-specific models.Source: §5.2.2 · pp. 16–17

5.2.3 Knowledge Augmenter

• Knowledge augmentation uses LLMs to add factual, contextual, rephrased, or synthetic information to recommendations, dialogue, classification, and patient-trial matching.Source: §5.2.3 · p. 17

5.2.4 Data Generator

• Figure 6 distinguishes controllers that manage the LLM-UM pipeline from evaluators that assess UM-system outputs.Source: Figure 6 · p. 18
• Training on LLM-synthesized data or weak labels lets smaller UM models inherit user-understanding capabilities, especially under low-resource or privacy-constrained conditions.Source: §5.2.4 · pp. 17–18

5.3 LLMs-as-Controllers

• LLMs-as-Controllers let language models manage pipeline operations and coordinate expert models or external tools.Source: §5.3 · pp. 18–19
• Controller systems support dialogue management, context tracking, recommendation-tool selection, conversational repair, and conditional job-post summarization.Source: §5.3 · pp. 18–19

5.4 LLMs-as-Evaluators

• LLMs-as-Evaluators assess customized outputs in dialogue, conversational recommendation, question generation, and job recommendation tasks.Source: §5.4 · p. 19
• LLM-based evaluation can approximate human preferences and provide scalable assessment for complex conversational generations where user studies are expensive or biased.Source: §5.4 · p. 19
• Personalized prompting has vanilla, retrieval-augmented, and profile-augmented forms for encoding user behavior and preferences into LLM inputs.Source: §5.4 · p. 19
• Longer behavior histories can improve personalization, whereas retrieval selects relevant history when context length is limited.Source: §5.4 · p. 19
• Profile-augmented prompts summarize task-specific preferences, such as positive and negative ratings or research interests, alongside retrieved behavior history.Source: §5.4 · p. 20

6 Applications of LLM-UM

• LLM-UM applications are categorized into personalization and suspiciousness detection.Source: §6 · p. 20

6.1 Personalization

• LLM-based personalization spans user profiling, recommendation, assistance, dialogue, education, and healthcare applications.Source: §6.1 · p. 20

6.1.1 User Profiling

• LLMs profile users by detecting stances and sentiments, modeling personality, and mining interpretable interest journeys from user-generated content.Source: §6.1.1 · p. 20

6.1.2 Personalized Recommendation

• Personalized recommendation includes top-k item prediction, rating prediction, and conversational recommendation using prompts, behavior histories, fine-tuning, and external reasoning tools.Source: §6.1.2 · p. 21

6.1.3 Personalized Assistance

• Personalized assistance applies LLMs to programming, task-oriented dialogue, citation and news generation, legal advice, investment advice, and creative writing.Source: §6.1.3 · p. 21

6.1.4 Personalized Dialogue System

• Personalized dialogue systems retrieve user history, detect personas, guide task completion, and generate customized responses while addressing factuality concerns.Source: §6.1.4 · p. 22

6.1.5 Personalized Education

• In education, LLMs adapt activities to students’ culture, age, and education level, support teaching and assessment, and approach human-tutor performance in several programming scenarios.Source: §6.1.5 · p. 22

6.1.6 Personalized Healthcare

• In healthcare, LLMs support physiological inference, clinical understanding, disease-risk estimation, medication planning, diagnosis, and case-specific medical advice.Source: §6.1.6 · p. 22

6.2 Suspiciousness Detection

• Suspiciousness detection uses user modeling to identify potentially anomalous, illegal, harmful, or malicious behavior.Source: §6.2 · p. 23

6.2.1 Fraud Detection

• Fraud-detection studies address clickbait, spam, fake profiles, fraudulent reviews, and network intrusion through prompting, fine-tuning, synthetic data, and model explanations.Source: §6.2.1 · p. 23

6.2.2 Discrimination Detection

• LLMs support discrimination detection for sexist, racist, multilingual, emoji-based, and abusive content through prompting, augmentation, explanation generation, and fine-tuning.Source: §6.2.2 · p. 23

6.2.3 Misinformation Detection

• Misinformation research uses fine-tuned models, credibility signals, weak supervision, adversarial training, and extra information, while LLM-generated misinformation can be harder to detect.Source: §6.2.3 · p. 23

6.2.4 LLM-Generated Text Detection

• LLM-generated text detection is motivated by misuse and evasion of plagiarism checks, with approaches including similarity networks, edit distance, coherence-based contrastive learning, and detector prompting.Source: §6.2.4 · p. 24

7.1 Hallucination Mitigation

• Hallucinations can be factually incorrect, ungrounded in user input, dangerous in high-stakes applications, or caused by misunderstanding user intent.Source: §7.1 · p. 24

7.2 Privacy and Security

• LLM-UM must balance personalization with privacy because training-data extraction, prompt attacks, jailbreaking, and changing privacy laws can expose sensitive information.Source: §7.2 · pp. 24–25

7.3 Complex Structural Data Understanding

• User interactions form heterogeneous, temporal, text-rich graphs whose lack of sequential representation, mismatch with language, and structural variability complicate LLM reasoning.Source: §7.3 · p. 25

7.4 Comprehensive Benchmark and Evaluation Criteria

• LLM-UM evaluation needs harder AGI-level tasks, intermediate pipeline assessment, and dynamic benchmarks that reflect evolving user data rather than static datasets alone.Source: §7.4 · pp. 25–26

7.5 Trustworthy User Modeling

• Trustworthy LLM-UM requires transparent reasoning, reliable and consistent outputs, and user control over data use and system decisions.Source: §7.5 · p. 26

7.6 Fairness in LLM-UM

• Fairness challenges arise when LLMs reproduce training-data biases, requiring bias measurement, mitigation, explainability, and ethical or regulatory work.Source: §7.6 · p. 26

7.7 Efficient and Effective Domain Adaptation

• Fine-tuning and deployment are computationally expensive at scale, while sparse and noisy user data can cause overfitting and poor generalization.Source: §7.7 · pp. 26–27

7.8 Personalized LLM-UM Deployment

• Personalized LLM deployment faces limited prompt context, costly preference alignment, unequal local compute, and privacy risks from reconstructing data encoded in embeddings or parameters.Source: §7.8 · p. 27

• The survey categorizes LLM-UM approaches, applications, challenges, and future directions to support researchers and practitioners.Source: §8 · p. 27