Source-linked AI summary
A Survey of Personalized Large Language Models: Progress and Future Directions
Jiahong Liu, Zexuan Qiu, Zhongyang Li, Quanyu Dai, Wenhao Yu, Jieming Zhu, Minda Hu, Menglin Yang, Tat-Seng Chua, Irwin King
TL;DR
LLMs perform strongly on general tasks but struggle to tailor responses to individual users, creating a need for systematic PLLM research. This survey organizes PLLM methods across prompting, adaptation, and alignment, reviews their evaluation and trade-offs, and identifies future directions. It concludes that prompting is efficient and adaptable for explicit, retrieval-oriented personalization, while deeper personalization introduces efficiency challenges.
Problem
General-purpose LLMs lack user-specific understanding, although personalized responses are important across conversational agents, recommendation, education, healthcare, and finance.
Method
The survey systematically organizes PLLM research through a three-level taxonomy covering prompting, model adaptation, and objective alignment, while reviewing methods and evaluations.
Results
Prompting-based methods are efficient and adaptable, particularly for explicit queries requiring concrete information retrieved from user data, while adaptation methods provide deeper personalization at higher computational cost.
Takeaways & Limitations
PLLM development should balance memory-based personalization capabilities—remember, adapt, and evolve—with trade-offs among performance, privacy, and efficiency.
Takeaways & Limitations
Personalization methods face a scalability boundary: deeper personalization can conflict with lightweight deployment because continual updating adds synchronization, latency, storage, and energy costs.
Abstract
from arXiv · showhide
Large Language Models (LLMs) excel in handling general knowledge tasks, yet they struggle with user-specific personalization, such as understanding individual emotions, writing styles, and preferences. Personalized Large Language Models (PLLMs) tackle these challenges by leveraging individual user data, such as user profiles, historical dialogues, content, and interactions, to deliver responses that are contextually relevant and tailored to each user's specific needs. This is a highly valuable research topic, as PLLMs can significantly enhance user satisfaction and have broad applications in conversational agents, recommendation systems, emotion recognition, medical assistants, and more. This survey reviews recent advancements in PLLMs from three technical perspectives: prompting for personalized context (input level), finetuning for personalized adapters (model level), and alignment for personalized preferences (objective level). To provide deeper insights, we also discuss current limitations and outline several promising directions for future research. Updated information about this survey can be found at the https://github.com/JiahongLiu21/Awesome-Personalized-Large-Language-Models.
1 Introduction
PLLMs address the limitations of general-purpose LLMs when user-specific understanding is crucial. This survey systematically formulates personalization scenarios, organizes methods across three technical levels, reviews their applicability, and summarizes benchmarks and future directions.
- Motivation: PLLMs are needed because general-purpose LLMs lack user-specific understanding for personalized interactions across domains such as recommendation, education, healthcare, and finance.Conversational agents, for example, must adapt to users’ preferred tones and incorporate past interactions.
- Survey scope: The survey addresses the lack of a systematic review by consolidating recent PLLM advancements and organizing their methodologies and future directions.
- Contributions: It formulates personalization scenarios using query types and personalized-data characteristics, establishing a unified framework for distinguishing techniques.
- Contributions: Its taxonomy classifies PLLM approaches into input-level prompting, model-level adaptation, and objective-level alignment.
- Contributions: The survey reviews state-of-the-art methods, analyzes fine-grained differences and scenario applicability, summarizes benchmarks and metrics, and outlines future directions.Highlighted directions include multimodal personalization, edge computing, lifelong updating, and trustworthiness.
2 Preliminary
PLLMs address the gap between general-purpose LLMs and user-specific needs by injecting personalized data through prompting, adaptation, or alignment. This section defines personalized data and query types, and frames personalization across input, model, and objective levels.
- Motivation: LLMs are powerful general-purpose models, but typically lack user-specific information and therefore cannot reliably tailor outputs to individual tastes and expectations.
- Motivation: PLLMs use personalized data to generate responses aligned with a user’s style and expectations, enabling different answers to the same query for different users.
- Personalized Data: Personalized data includes user profiles, relationships, historical dialogues, historical content, interactions, and predefined human preferences.
- Query Types: Queries are categorized as explicit extraction, implicit abstraction, and implicit generalization according to how responses use personalized data and external knowledge.Extraction retrieves facts or aggregates from personalized data; abstraction summarizes long-term tendencies; generalization dynamically combines user data with the current query.
- Downstream Tasks: Personalized LLMs support generation, recommendation, and classification tasks by incorporating user-specific data into response production and prediction.Recommendation uses historical interactions to suggest items and may provide explanations, while classification labels entities according to user preferences.
- Proposed Taxonomy: The survey organizes PLLM methods into three technical levels: input-level personalized prompting, model-level adaptation, and objective-level alignment.The framework represents personalization as an operator that injects personalized data into a base LLM and may optimize tunable parameters or additional modules.
3 Personalized Prompting
Personalized prompting injects user context into LLMs through profile summaries, retrieved memories, soft embeddings, or contrastive signals. These methods offer efficient, adaptable personalization, but differ in information richness, model access, controllability, and suitability for explicit versus implicit queries.
- 3 Personalized Prompting: Personalized prompting extracts relevant context from user data and fuses it with a query before generation.The survey groups these methods into profile-augmented, retrieval-augmented, soft-fused, and contrastive prompting.
- 3.1 Profile-Augmented Prompting: Profile-augmented prompting summarizes user preferences or history into natural-language tokens that augment the LLM input.Summarizers may be frozen or tuned; tuned approaches address input noise and relevance problems.
- 3.2 Retrieval-Augmented Prompting: Retrieval-augmented prompting selects relevant records from user data, often using sparse or dense retrievers and external memory.Memory construction can be non-parametric, storing tokenized feedback, or parametric, encoding user data into learnable representations.
- 3.3 Soft-Fused Prompting: Soft-fused prompting compresses personalized data into embeddings generated by a user feature encoder rather than discrete profile tokens.The embeddings can enter through input prefixes, cross-attention, or output-logit adjustment.
- 3.4 Contrastive Prompting: Contrastive prompting compares outputs or hidden representations with and without personal context to extract or steer personalization factors.CoS adjusts the personalization degree through scaling factor α without fine-tuning, while related methods steer decoding or hidden states.
- 3.5 Discussions: Prompting methods trade off efficiency, information coverage, semantic nuance, interpretability, model access, and sensitivity to irrelevant data or α.Profile compression risks information loss; retrieval can retrieve irrelevant records; soft prompting is limited to black-box models; contrastive prompting is hyperparameter-sensitive.
- 3.5 Discussions: Prompting-based methods are efficient and adaptable, particularly for explicit factual queries that require retrieving concrete information from user data.Retrieval-augmented prompting can ground responses in retrieved evidence, while prompting generally uses minimal computational overhead.
- 3.5 Discussions: Prompting has limited deeper personalization analysis and often underperforms adaptation or alignment methods on implicit generation queries.Its predefined prompt structures constrain access to global knowledge and the analysis of user preferences.
4 Personalized Adaptation
Personalized adaptation methods tailor LLM parameters or parameter-efficient modules to user needs, balancing personalization depth, efficiency, privacy, and scalability. Key challenges include overfitting, cold starts, and the performance–privacy–efficiency trade-off.
- Shared adaptation: Single shared PEFT modules improve efficiency but can limit personalization depth and require centralized handling of user data.Centralized approaches raise concerns about storing, using, and protecting personal data.
- User-specific adaptation: Per-user PEFT modules strengthen personalization and privacy but may exceed the fine-tuning resources available on user devices.This creates deployment difficulties in resource-limited environments.
- Collaborative adaptation: Collaborative and federated strategies address storage and computation burdens by sharing knowledge across users, but introduce coordination and privacy challenges.Federated learning can outperform purely local fine-tuning, while parameter sharing may create privacy leakage or edge–cloud efficiency costs.
- Adaptation methods: PEFT methods such as LoRA and prefix-tuning reduce computational and memory costs while preserving deep user-specific adaptation.They tailor models to user needs while leveraging pretrained knowledge.
- Challenges: Limited or noisy user data can cause overfitting, while sparse or unseen data makes adaptation to new users difficult.These issues constrain generalization and robustness in personalized models.
- Challenges: Balancing performance, privacy, and efficiency requires complementary strategies including federated learning, synthetic data, and hybrid prompting–adaptation pipelines.The proposed direction targets efficiency, personalization depth, and reliability together.
5 Personalized Alignment
Personalized alignment adapts LLM outputs to individual preferences through reward modeling and policy optimization at training or decoding time. The survey organizes alignment methods, datasets, and benchmarks while identifying limited preference representations as a major boundary.
- Motivation and data: Personalized alignment extends generic preference alignment to user-specific differences in language style, knowledge depth, and values.The central challenge is constructing high-quality data that captures individual preferences.
- Motivation and data: Alignment research constructs preference datasets through simulated interactions, user personas, and feedback from participants across countries and LLMs.One dataset uses 3,310 user personas, while PRISM maps 1,500 participants from 75 countries to feedback on 21 LLMs.
- Training-time alignment: Multi-objective RLHF trains separate reward models for preference dimensions and combines their rewards to guide policy optimization.This is the training-time personalization paradigm shown in Figure 7(a).
- Decoding-time alignment: Decoding-time personalization combines trained policy models through personalized weight merging or model ensembling during inference.These strategies are contrasted with training-time personalization in Table 4.
- Alignment methods: Personalized reward modeling can generate token-level rewards that dynamically guide base-model decoding toward individual preferences.PAD uses this strategy to adapt predictions during generation.
- Limitations and benchmarks: Current approaches often rely on predefined dimensions and small preference sets, limiting their ability to capture diverse real-world personalization needs.More realistic methods must infer preferences from interaction histories when preference vectors are unknown.
6 Metric and Benchmark
The survey organizes personalization evaluation by task type, data, and benchmark, combining conventional, human, LLM-based, and personalization-specific metrics. It emphasizes that current benchmarks remain incomplete for evolving, cross-domain, multimodal, and scarce-data settings.
- Evaluation framework: Evaluation coverage is organized around task types, commonly used metrics, and benchmarks for personalization approaches.The survey first summarizes metrics and then reviews benchmarks and their metric usage.
- Classification metrics: Accuracy, F1, and MCC are used for categorical classification, while MAE and RMSE suit ordinal classification because they reflect error magnitude.MCC is more robust under class imbalance, and RMSE penalizes larger deviations more heavily.
- Generation metrics: BLEU, ROUGE-1, ROUGE-L, and METEOR measure lexical overlap, while human and LLM judges provide broader response-quality assessments.Human evaluation is described as reliable but costly and time-intensive; LLM evaluation is scalable and context-aware.
- Personalization metrics: Personalization-specific metrics address a gap left by standard quality measures, but the personalization-accuracy paradox can reward responsiveness despite poor accuracy.PerSEval addresses this by combining responsiveness with accuracy-based penalties.
- Benchmark scope: Existing benchmark summaries exclude some memory-retrieval tasks because the survey focuses on injecting and using personalized data for personalized performance.The excluded examples include PerLTQA, LongMemEval, and IMPLEXCONV memory-related testing tasks.
- Benchmark gaps: Future benchmarks should test lifelong updates, cross-domain adaptation, cross-modal personalization, and generalization from scarce, noisy, or imbalanced data.These settings target conflict resolution, domain transfer, multimodal inputs, inconsistency detection, refusal, and limited-data generalization.
7 Vision and Future Directions
The survey’s vision for PLLMs is that they remember user knowledge, adapt through abstraction and generalization, and evolve through continual learning. Achieving this vision requires balancing efficacy, efficiency, privacy, and trustworthiness while addressing complex data, edge deployment, updating, and fairness challenges.
- Vision: PLLMs are envisioned to remember short- and long-term user knowledge, adapt preferences through abstraction, and evolve without forgetting.Evolution requires continual updating as users’ goals and behavior change.
- Efficacy: Prompting is effective and efficient for extracting personalized information, whereas adaptation and alignment become more central for abstraction and generalization.Deeper personalization generally requires higher computational and data costs.
- Efficacy: Lifelong evolution remains under-explored because models struggle to update and grow without catastrophic forgetting.This limits performance at the continual-learning stage.
- Efficiency: Efficiency declines as methods move from prompting toward adaptation and alignment, and continual updating adds synchronization, latency, storage, and energy costs.The resulting tension is between personalization depth and scalable lightweight deployment.
- Privacy and trade-offs: Per-user PEFT improves privacy but can reduce performance, while collaborative approaches improve performance yet increase privacy-leakage risks.Parameter sharing offers a compromise but does not remove the broader trade-off.
- Future directions: Future research must address complex multimodal user data, resource-constrained edge updates, edge–cloud synchronization, and efficient parameter updates after base-model changes.Suggested directions include personalized small models, incremental learning, transfer learning, and continual learning.
- Trustworthiness and applications: Privacy-preserving methods, interpretability, fairness, and bias mitigation remain necessary for trustworthy PLLM adoption across applications.Personalization can amplify stereotypes or raise privacy concerns in domains such as healthcare and education.
8 Conclusions
The survey synthesizes PLLM research around personalized responses to individual user data and organizes methods into three technical perspectives. It also identifies memory-centric capabilities, trade-offs, domain challenges, limitations, and future directions.
- The survey categorizes PLLM methods into personalized prompting, personalized adaptation, and personalized alignment at the input, model, and objective levels.
- It proposes a memory-centric PLLM vision built around the capabilities to remember, adapt, and evolve.
- The survey analyzes trade-offs among performance, privacy, and efficiency in memory-centric PLLMs.
- It examines domain-specific challenges and outlines future directions for PLLM research and practical deployment.
A Notation Summary
The notation summary consolidates the symbols used throughout the survey and groups them by their roles in the paper’s formal framework.
- Table 8 groups the survey’s notation into spaces, functions, models, operations, and variables.