PaperlayerView exact revision on arXiv

Source-linked AI summary

Personalization of Large Language Models: A Survey

Zhehao Zhang, Ryan A. Rossi, Branislav Kveton, Yijia Shao, Diyi Yang, Hamed Zamani, Franck Dernoncourt, Joe Barrow, Tong Yu, Sungchul Kim, Ruiyi Zhang, Jiuxiang Gu, Tyler Derr, Hongjie Chen, Junda Wu, Xiang Chen, Zichao Wang, Subrata Mitra, Nedim Lipka, Nesreen Ahmed, Yu Wang

arXiv:2411.00027v3cs.CL

TL;DR

Personalized LLM research spans user-tailored text generation and personalization for downstream tasks, while the paper frames adaptation to user needs as important for user-focused applications. The survey unifies these directions and organizes the field’s concepts, methods, evaluation, and applications; it identifies open research challenges including privacy, bias, and cold-start.

  • Problem

    LLMs are increasingly adapted to user-specific contexts and preferences for user-focused interactions across applications.

    Source: §1 · p. 2

  • Method

    The survey formalizes personalized LLM foundations and organizes usage, granularity, techniques, evaluation, datasets, applications, and open problems into taxonomies.

    Source: §1 · pp. 2–3

  • Results

    The survey unifies direct personalized text generation and downstream-task personalization, and describes trade-offs among user-, persona-, and global-preference approaches.

    Source: §1 · p. 3, §10 · p. 41

  • Takeaways & Limitations

    The synthesis highlights granularity trade-offs and identifies cold-start, bias, and privacy as future research priorities.

    Source: §1 · p. 3, §10 · p. 41

    LLM-based evaluators can exhibit positional bias, be vulnerable to adversarial inputs, and require manual prompting that limits efficiency and flexibility.

    Source: §9.1 · p. 38

Abstract

from arXiv · show

Personalization of Large Language Models (LLMs) has recently become increasingly important with a wide range of applications. Despite the importance and recent progress, most existing works on personalized LLMs have focused either entirely on (a) personalized text generation or (b) leveraging LLMs for personalization-related downstream applications, such as recommendation systems. In this work, we bridge the gap between these two separate main directions for the first time by introducing a taxonomy for personalized LLM usage and summarizing the key differences and challenges. We provide a formalization of the foundations of personalized LLMs that consolidates and expands notions of personalization of LLMs, defining and discussing novel facets of personalization, usage, and desiderata of personalized LLMs. We then unify the literature across these diverse fields and usage scenarios by proposing systematic taxonomies for the granularity of personalization, personalization techniques, datasets, evaluation methods, and applications of personalized LLMs. Finally, we highlight challenges and important open problems that remain to be addressed. By unifying and surveying recent research using the proposed taxonomies, we aim to provide a clear guide to the existing literature and different facets of personalization in LLMs, empowering both researchers and practitioners.

1 Introduction

Personalized LLM research has largely developed along separate tracks: generating text tailored to users and using personalization for downstream tasks. This survey connects those tracks with a framework spanning foundations, granularity, techniques, evaluation, datasets, applications, and open problems.

1 Introduction

  • The survey unifies direct personalized text generation and downstream-task personalization in a taxonomy of two main uses.The introduction also presents formal foundations and taxonomies for personalization granularity, techniques, evaluation, datasets, and applications.Source: §1 · pp. 2–3, Figure 1 · p. 4
  • User-level personalization is the finest-grained approach but needs sufficient user data; persona-level personalization can serve users with limited data, while global preference personalization is not user-specific.The survey discusses trade-offs among the three levels.Source: §1 · p. 3
  • The survey identifies improved benchmarks and metrics, cold-start with sparse user data, stereotypes and bias, privacy, and multimodal integration as open challenges.It also calls for more diverse and representative datasets.Source: §1 · p. 3
  • The article proceeds from a unifying view and theoretical foundations to granularity, techniques, evaluation, datasets, applications, and future research directions.Source: §1 · p. 3

2 Unifying Personalized LLMs

The survey distinguishes personalized text generation, evaluated as generated content, from downstream-task personalization, where user-specific information supports another task. It presents these as related approaches with different evaluation targets and notes challenges in evaluating personalized text.

2 Unifying Personalized LLMs

  • Direct personalized text generation integrates user-specific information through an adaptation function, then forms a personalized prompt for an LLM.The adaptation function can use a retrieval module and user documents, attributes, interaction histories, or preferences.Source: §2 · p. 5
  • Direct approaches assess generated text against user-specific references, whereas downstream approaches assess task outcomes such as recommendation performance.References for direct evaluation may be user-written text, preferences, reward-model outputs, or user judgments.Source: §2 · pp. 4–5
  • Scarce high-quality user-written labels make direct evaluation of personalized text challenging, and indirect approaches typically do not evaluate the generated text itself.Indirect methods instead focus on whether intermediate text or embeddings improve downstream performance, though their generated information can be difficult to interpret.Source: §2 · p. 5
  • Downstream personalization can use personalized text or embeddings alongside task-specific information as input to a task model.The model produces predictions such as inferred ratings or scores.Source: §2 · p. 5
  • The two approaches share mechanisms for using user data and constructing personalized prompts or embeddings, but differ in datasets and evaluation methods.The survey describes them as potentially complementary, with advances in either direction able to inform the other.Source: §2 · p. 5

3 Foundations of Personalized LLMs

Section 3 establishes foundations for personalized LLMs through formalization and taxonomies of generation and personalization criteria. It distinguishes preference-driven personalization by target-group size from general task fine-tuning and covers relevant model, prompt, and data foundations.

3 Foundations of Personalized LLMs

  • The section addresses the lack of a comprehensive theoretical framework with formal foundations, a formalization of personalized generation, and a taxonomy of personalization criteria.It establishes principles, definitions, and structures for formalizing and evaluating personalization in LLMs.Source: §3 · p. 6

3.1 Preliminaries

  • An LLM is described as a billion-parameter-or-larger Transformer mapping text inputs to output sequences, while prompts guide generation and may include text, image, audio, or video.The section also outlines pretraining, instruction tuning, RLHF, and task-specific fine-tuning.Source: §3.1 · pp. 6–7

3.3 Personalization Data

  • Personalization data include user documents, relatively stable attributes, and dynamic interactions; the example pairs user information with a movie's title, year, and genre.The listed interaction data include clicks, views, and purchases.Source: §3.2 · p. 7, §3.3 · p. 8

3.2 Formulation of Personalization

  • Personalization targets individuals, groups, or the general public, and fine-tuning counts as personalization only when it adapts to user or group preferences.General-corpus fine-tuning for overall task performance is not personalized.Source: §3.2 · p. 8

Interaction History

The example logs one user's interaction with a movie, including a rating and timestamp.

Interaction History

User-written Text

The example records a user's written review of Movie 24.

User-written Text

Pair-Wise Human Preferences

This section surveys user-specific data for personalization and formalizes how personalized generations differ from general high-quality outputs, alongside taxonomies of personalization criteria, granularity, techniques, evaluation, and datasets. It distinguishes datasets for direct evaluation of personalized text generation from those supporting indirect evaluation through downstream applications.

3.3.1 Static Attributes; 3.3.2 Interaction History

  • Static attributes support long-term targeting but can become outdated and raise privacy concerns; interaction histories provide timely context but are complex to process and may not reflect current preferences.Static attributes include user and item information, while interaction histories include past activities and engagement with items or LLM outputs.Source: §3.3.1 · pp. 8–9, §3.3.2 · p. 9

3.3.3 User-Written Text

  • User-written reviews, dialogues, social posts, and few-shot exemplars reveal preferences and sentiment, but unstructured, noisy text is difficult to evaluate and costly to annotate.Examples include Amazon Review Data’s 233.1 million reviews and persona-based ConvAI2 dialogues (Dinan et al., 2020).Source: §3.3.3 · pp. 9–10

3.3.4 Pair-Wise Human Preferences

  • Pair-wise preferences provide direct, specific guidance about desired responses, but current alignment work often targets aggregate preferences rather than diverse individual perspectives (Jang et al., 2023).Users can express preferences such as whether an answer should suit a layperson or an expert; designing effective personalized alignment remains challenging.Source: §3.3.4 · p. 10

3.4 Space of Personalized Generations

  • Personalized generation targets a small subset of high-quality responses that also match an individual user’s preferences, making both generation and its evaluation challenging.The space of user-specific generations may be much smaller than the space of all possible generations, and different users’ personalized high-quality spaces may be disjoint.Source: Figure 3 · p. 10, §3.4 · p. 11

3.5 Personalization Criterion Taxonomy

  • The proposed evaluation-criteria taxonomy covers tone and style, relevance, and accuracy, including factual correctness and accurate, up-to-date user data.Relevance includes both alignment with user interests and fit to the immediate context.Source: §3.5 · p. 12, Figure 4 · p. 12

3.6.1 Taxonomy of Personalization Granularity of LLMs; 3.6.2 Taxonomy of Personalization Techniques for LLMs

  • The granularity taxonomy distinguishes user-level, persona-level, and global-preference personalization, while the techniques taxonomy covers retrieval-augmented generation, prompting, representation learning, and reinforcement learning from human feedback.These levels range from one user’s preferences and behavior to shared persona traits and general norms; the techniques differ in how user information is incorporated.Source: §3.6.1 · p. 13, §3.6.2 · p. 13

3.6.3 Taxonomy of Evaluation Methodologies for Personalized LLMs; 3.6.4 Taxonomy of Datasets for Personalized LLMs

  • The evaluation taxonomy separates intrinsic assessment of personalized text from extrinsic assessment through downstream applications; the dataset taxonomy distinguishes data containing user-written ground-truth text.The dataset distinction is intended to clarify datasets’ roles in directly or indirectly training and evaluating personalized LLMs.Source: §3.6.3 · p. 14, §3.6.4 · p. 14

3.6.4 Taxonomy of Datasets for Personalized LLMs

  • Datasets with user-specific ground-truth text enable direct evaluation of personalized text generation.Source: §3.6.4 · p. 14
  • Datasets without user-specific ground-truth text are used for indirect evaluation via downstream applications.Source: §3.6.4 · p. 14

4 Personalization Granularity of LLMs

Personalization granularity ranges from individual users to personas and global preferences, with distinct trade-offs in precision, scalability, and data needs. The survey also discusses hybrid approaches and integrated datasets as ways to balance these levels.

4 Personalization Granularity of LLMs

  • Granularity distinguishes personalization for individual users, groups sharing characteristics, and the general public.The level determines how finely or broadly responses are tailored.Source: §4 · p. 15

4.1 User-level Personalization

  • User-level personalization uses an individual’s profile and history for fine-grained recommendations, but sparse interaction histories create cold-start challenges.Some studies remove users with minimal histories during preprocessing, potentially omitting insights from these underrepresented interactions.Source: §4.1 · p. 16

4.2 Persona-level Personalization

  • Persona-level personalization targets groups through shared traits, making it useful when group characteristics matter more than individual attributes.It is less granular than user-level personalization, and extracting representative traits can require substantial domain knowledge.Source: §4.2 · pp. 16–17

4.3 Global Preference Personalization

  • Global preference personalization aligns models with broadly accepted norms and values, but aggregated preferences can introduce noise and lack fine-grained specificity.The survey includes this level for completeness while noting it falls outside its primary scope.Source: §4.3 · p. 17

4.4 Discussion

  • The levels trade off precision, scalability, and experience richness; the survey proposes hybrid transitions and integrated datasets to combine them.One proposed hierarchy uses user-level personalization for frequent users, persona-level for occasional users, and global preferences for new users.Source: §4.4 · pp. 17–18

5 Taxonomy of Personalization Techniques for LLMs

The survey classifies personalization techniques by how they use user information, including retrieval, prompting, representation learning, and RLHF. Its retrieval discussion contrasts efficient lexical methods with denser methods that tend to perform better downstream but require more training.

5 Taxonomy of Personalization Techniques for LLMs

  • The taxonomy groups personalization techniques by how user information is used, covering RAG, prompting, representation learning, and RLHF.The approaches are mostly orthogonal and can coexist in a system.Source: §5 · p. 18

5.1 Personalization via Retrieval-Augmented Generation

  • RAG retrieves external information and incorporates it into the input so an LLM can generate grounded, personalized outputs.Large user profiles can serve as external knowledge bases when they do not fit within the model’s context.Source: §5.1 · pp. 18–19

5.1.1 Sparse Retrieval

  • Sparse retrieval uses frequency-based vectors and exact term overlap, offering efficiency but struggling with semantic relationships in complex personalization tasks.BM25 and TF-IDF are common sparse retrievers; one surveyed BM25 approach reduced retrieved data volume by 75% while improving LaMP performance.Source: §5.1.1 · pp. 20–21, §5.1 · p. 20

5.1.2 Dense Retrieval

  • Dense retrievers use neural embeddings for similarity search and tend to outperform sparse retrievers on downstream personalization tasks, but require more costly training.Effective training data, loss functions, and ways to use LLMs in retriever training remain open challenges.Source: §5.1 · p. 20, §5.1.2 · p. 21

5.2 Personalization via Prompting

The survey organizes prompt-based personalization around contextual histories, personas, profile augmentation, and prompt refinement. These strategies vary in how they represent user information and face challenges including data scale, persona risks, and dependence on hand-crafted prompts.

5.2 Personalization via Prompting

  • Prompt-based personalization includes contextual, persona-based, profile-augmented, and refined prompts.The categories describe different ways of incorporating or improving user information in prompts.Source: §5.2 · p. 22

5.2.1 Contextual Prompting

  • Contextual prompting places user history and candidate information in the prompt for downstream tasks, but large unstructured histories can challenge scalability and interpretation.Studies report mixed recommendation results; some find LLMs comparable to or better than traditional systems after fine-tuning with minimal user interaction data.Source: §5.2.1 · p. 22, §5.2 · p. 22
  • Contextual prompting can improve explainability without necessarily producing substantial gains over traditional non-LLM methods.Reported results vary by task, including strong explanation and summary generation but weaker sequential and direct recommendations in one study.Source: §5.2.1 · p. 22

5.2.2 Persona-based Prompting

  • Persona-based prompting asks an LLM to role-play a specified persona, which can alter response characteristics to match prompt requirements.Personas may represent demographic segments, established characters, or individuals described through behavioral and preference data.Source: §5.2.2 · pp. 22–23, §5.2 · p. 22
  • Persona prompting may produce character hallucination, biases, toxicity, jailbreaking, ecological fallacy, or caricatures.The survey notes these risks alongside the potential for persona-based role-playing to reflect certain personalities.Source: §5.2.2 · p. 23

5.2.3 Profile-Augmented Prompting

  • Profile augmentation addresses oversized, irrelevant, incomplete, or sparse user profiles by eliciting LLM knowledge to augment or distill profile information.Examples include summaries of browsing interests, enriched item descriptions, and hierarchical personas refined with information from similar users.Source: §5.2 · p. 22, §5.2.3 · p. 23

5.2.4 Prompt Refinement

  • Prompt refinement addresses the cost and trial-and-error burden of hand-crafted prompts by training or iteratively optimizing prompts with user feedback.FERMI uses feedback scores, a memory bank of high-scoring prompts and contexts, and prompt selection for each test query.Source: §5.2 · p. 22, §5.2.4 · pp. 23–24

5.3 Personalization via Representation Learning

Representation-learning approaches personalize LLMs by using user information to adapt model parameters or representations. The section surveys full-parameter fine-tuning, parameter-efficient fine-tuning, and embedding learning.

5.3 Personalization via Representation Learning

  • Representation-learning methods span full-parameter updates, selective parameter updates, and learned embeddings for encoding user information.The survey groups these approaches by how they use user information to personalize LLMs.Source: §5.3 · p. 24

5.3.1 Full-Parameter Fine-tuning

  • Full-parameter fine-tuning updates all LLM parameters on targeted data, and can improve personalization-task performance over prompting when parameters are accessible and costs acceptable.Examples include instruction-tuning datasets for recommendation and training a model to distinguish documents by author.Source: §5.3 · p. 24, §5.3.1 · p. 24

5.3.3 Embedding Learning

  • Embedding learning represents text and user information as vectors, with user-specific embeddings derived from interactions and integrated into the model to personalize responses.Examples use embeddings to capture preferences and behavior, including input-aware embeddings that guide a fixed LLM without changing its parameters.Source: §5.3 · p. 24, §5.3.3 · pp. 25–26

5.3.2 Parameter-Efficient Fine-tuning

  • Parameter-efficient fine-tuning adapts an LLM by changing only a small parameter subset or adding parameters while keeping most original parameters fixed.Methods include personalized PEFT modules and collaboratively assembled PEFT components.Source: §5.3 · p. 24, §5.3.2 · p. 25

5.4 Personalization via Reinforcement Learning from Human Feedback (RLHF)

The section reviews using reinforcement learning from human feedback (RLHF) to align LLMs with personalized preferences, which can differ across users. It describes approaches that personalize feedback, rewards, and alignment during training or inference.

5.4 Personalization via Reinforcement Learning from Human Feedback (RLHF)

  • RLHF optimizes an LLM policy using preferential human feedback on its outputs, with earlier methods learning a proxy reward model and modern methods optimizing directly from feedback.The section situates RLHF within alignment to human preferences and values.Source: §5.4 · p. 26
  • Personalized alignment addresses differing user preferences for the same prompt, including preferences that vary across dimensions.Approaches include decomposing preferences into objectives, personalized reward models, and adapting alignment during decoding.Source: §5.4 · p. 26
  • Personalized RLHF methods include inference-time personalized rewards, joint user-specific model learning, preference aggregation, and user-specific latent-variable models.Demonstrations can also provide an alternative to pairwise preferences for fine-grained alignment with limited user input.Source: §5.4 · pp. 26–27

5.5 Discussion

The discussion compares personalization methods by effectiveness, resource demands, and application constraints. It concludes that no single approach is optimal across all tasks.

5.5 Discussion

  • Dense retrieval may incur indexing overhead and GPU memory use, while deriving personas or hierarchical prompts can add inference latency.The discussion identifies infrastructure, data availability, latency, and regulatory considerations as factors shaping method choice.Source: §5.5 · pp. 27–28
  • Retrieval-based methods generally excel when tasks require explicit grounding, while dense retrievers are particularly adept at complex abstractions.The discussion draws on LamP and longLamP benchmarks.Source: §5.5 · p. 27
  • Prompt-based personalization is flexible and lightweight with moderate user data, but performance degrades as context windows become saturated.RAG can suit large knowledge bases or dynamic profiles, whereas prompting suits systems using brief interaction histories without retraining.Source: §5.5 · pp. 27–28
  • Fine-tuning or user embeddings can yield deeper personalization gains, but sparse per-user data risks overfitting and training requires substantial computation.Representation learning is suggested when data and computational capacity are abundant and nuanced personalization matters.Source: §5.5 · pp. 27–28

6 Taxonomy of Evaluation Metrics for Personalized LLMs

The evaluation taxonomy separates direct assessment of personalized text from assessment of its utility in downstream applications. It surveys metrics and evaluation challenges for both approaches.

6 Taxonomy of Evaluation Metrics for Personalized LLMs

  • Intrinsic evaluation assesses personalized text directly, while extrinsic evaluation measures its utility through downstream-application outcomes.The taxonomy distinguishes text-level evaluation from application-specific evaluation.Source: §6 · p. 28

6.1 Intrinsic Evaluation

  • Intrinsic evaluation commonly uses reference-based text metrics, while EGISES measures preference responsiveness and PerSEval also accounts for accuracy and consistency.The survey notes that EGISES captures personalization independently of accuracy, whereas PerSEval adds an accuracy penalty.Source: §6.1 · p. 29, §6 · p. 29
  • LLM judges are not yet considered ready to replace human judges, and the survey reports no framework specifically designed to evaluate personalization systematically.Human evaluations remain essential for checking whether outputs meet user expectations, despite their practical cost.Source: §6.1 · pp. 29–30

6.2 Extrinsic Evaluation

  • Extrinsic evaluation uses task-specific measures, including Recall, Precision, and NDCG for recommendation and classification metrics such as Accuracy and F1 Score.Most such metrics do not directly assess personalization, though improvements may implicitly reflect better use of user information.Source: §6 · p. 29, §6.2 · p. 30

7 Taxonomy of Datasets for Personalized LLMs

The survey distinguishes datasets for direct evaluation of personalized text from datasets that evaluate generated text through downstream applications. Ground-truth user text is relatively rare, while indirect-evaluation datasets are more common.

7.1 Personalized Datasets with Ground-Truth Text

  • Ground-truth datasets support direct evaluation and include short-text tasks such as titles and long-text tasks such as reviews.Short outputs may be 9–10 words, whereas long-text benchmarks may target hundreds or thousands of words.Source: §7 · p. 30, §7.1 · pp. 31–32, Table 5 · p. 32

7.2 Personalized Datasets without Ground-Truth Text

  • Datasets without user-written ground truth are more common and support indirect evaluation through recommendation, classification, dialogue, and question answering.Such evaluation shows whether generated text helps a downstream application, but not whether the text is meaningful or relevant to the user.Source: §7 · pp. 30–31, §7.2 · p. 33

7 Taxonomy of Datasets for Personalized LLMs

  • Table 5 organizes datasets by task and records prior-activity filtering, user-generated text, numerical and categorical attributes, and evaluation metrics.It distinguishes user-written text from other text, such as item descriptions.Source: §7 · p. 31, Table 5 · p. 32

7.2 Personalized Datasets without Ground-Truth Text

  • Indirect-evaluation datasets can include user attributes and interaction data, while item descriptions may contribute to techniques without being user-specific.Generated text or embeddings can augment downstream models, but improved task performance alone does not establish user-specific relevance.Source: §7 · p. 31, §7.2 · p. 33
  • PRISM Alignment and Empathic Conversations contain rich user-specific information and may support personalized evaluation where user-written ground-truth text is unavailable.The survey identifies reward models based on this information as a largely unexplored possibility.Source: §7.2 · p. 33

8 Applications of Personalized LLMs

The survey describes personalized LLM applications across education, healthcare, and other domains, as well as recommendation and search. It highlights both specific systems and challenges reported or anticipated in these settings.

8.1.1 Education

  • Educational applications include tailored feedback and materials, while reported challenges include limited conversation history, role shifts, engagement, privacy, and fair access.EduChat offers customized support for Socratic teaching, emotional counseling, and essay assessment.Source: §8.1.1 · pp. 33–34

8.1.2 Healthcare

  • Healthcare approaches include multimodal disease-risk prediction, conversational agents using external knowledge and tools, memory-based assistants, and a RAG pipeline for disease prediction.MaLP combines dual-process enhanced memory with PEFT to provide personalized responses while maintaining low resource consumption.Source: §8.1.2 · p. 34

8.1.3 Other Domains

  • Beyond education and healthcare, examples include personalized financial advice, legal tools and role-based support, and code generation adapted to individual coding styles.MPCODER combines explicit and implicit style learning with a multi-user style adapter; its authors introduce Coding Style Score.Source: §8.1.3 · pp. 34–35

8.2 Recommendation

  • Recommendation approaches augment traditional systems with LLM embeddings or generated features, or use LLMs directly as recommenders.The survey reports few-shot and zero-shot performance and enhanced explainability, while noting privacy, cost, and latency challenges at scale.Source: §8.2 · p. 36

8.3 Search

  • Personalized search can use search histories and interaction records, but LLM-powered search has been associated with more biased querying and opinion polarization than conventional web search.The cited finding is especially noted when the LLM reinforces users’ existing views.Source: §8.3 · p. 36

9 Open Problems & Challenges

The survey identifies unresolved challenges in evaluation, sparse-data personalization, bias, privacy, and multimodal systems. It calls for more comprehensive evaluation and methods that address the constraints of each area.

9.1 Comprehensive Evaluation: Benchmarks, Automatic Metrics, LLM-as-a-Judge, and Beyond

  • Personalization benchmarks often emphasize downstream predictions, while LaMP covers only text classification and short, single-turn generation rather than realistic interactions.The survey also notes the lack of a comprehensive quantitative personalization metric and the difficulty of comparisons across inconsistent datasets and tasks.Source: §9.1 · pp. 37–38

9.2 Cold-start Problem

  • Cold-start adaptation remains underexplored because sparse user histories challenge fine-tuning and many datasets filter out users with fewer than five interactions.Persona-DB uses hierarchical persona construction and collaborative refinement to draw on richer users’ histories.Source: §9.2 · pp. 38–39

9.3 Stereotype & Bias Issues of Personalization

  • Personalization can reinforce societal biases, and studies report demographic disparities in educational content and restricted recommendation options for minority users.The survey also notes that eliminating personalization bias remains an open challenge.Source: §9.3 · p. 39

9.4 Privacy Issues

  • Personalization and privacy can conflict: LLMs may infer personal attributes from unstructured text, while work specifically addressing personalization-related privacy remains limited.The survey calls for defining task- and user-dependent privacy boundaries and preventing explicit and implicit leaks across data processing, training, and retrieval.Source: §9.4 · p. 40

9.5 Multimodality

  • Multimodal personalization must align heterogeneous inputs, handle sparse or noisy data, and capture fine-grained cues such as color, texture, or rhythm.Task-specific challenges include balancing fidelity and diversity in personalized image generation and improving alignment and scalability across applications.Source: §9.5 · pp. 40–41

10 Conclusion

The survey unifies direct personalized text generation and downstream applications through taxonomies and foundations, then synthesizes techniques, evaluation, and open challenges. It highlights trade-offs among personalization granularity and the need for benchmarks and methods addressing cold starts, bias, privacy, and multimodal systems.

10 Conclusion

  • The survey bridges direct personalized text generation and downstream applications with a unified taxonomy and formalized foundations.Its review also spans personalization techniques, evaluation methodologies, and datasets.Source: §10 · p. 41
  • User-level, persona-level, and global preference alignment involve trade-offs, motivating future hybrid systems that adapt to user needs and data availability.Source: §10 · p. 41
  • The survey examines retrieval-augmented generation, prompting, representation learning, and personalized-feedback learning, while noting the need for more robust and nuanced methods.Source: §10 · p. 41
  • Evaluation methodologies and datasets reveal a critical need for benchmarks and metrics designed specifically to assess personalized LLM outputs.Source: §10 · p. 41
  • Open challenges include low-resource cold starts, stereotypes and bias, user privacy, and extending personalization to multimodal systems.The survey frames its foundations, techniques, and challenges as relevant to developing more effective, fair, and socially responsible systems for diverse user needs.Source: §10 · pp. 41–42
Loading 2411.00027v3…