Personalization of Large Language Models: A Survey

TL;DR

The paper provides a unified framework for personalization of large language models by bridging direct personalized text generation with downstream task personalization through formal foundations and comprehensive taxonomies. It introduces three levels of personalization granularity, a taxonomy of techniques (RAG, prompting, representation learning, RLHF), and explicit evaluation and dataset taxonomies to standardize research and practice. The work further surveys applications across education, healthcare, finance, law, coding, and search, and discusses open problems including benchmarks, cold-start, bias, privacy, and multimodality. Together, these contributions offer a structured, multi-faceted roadmap for designing, evaluating, and deploying personalized LLMs in diverse domains. The ultimate goal is to enable integrated, scalable, and ethically responsible personalized AI that adapts to individual and group user needs while maintaining safety and fairness.

Abstract

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.

Figures (8)

• Figure 1: Taxonomy for Personalized LLM Usage. To bridge the gap in the existing literature on personalized LLMs, we propose the intuitive taxonomy outlined above, which categorizes work into two main areas. The first focuses on studying the ➊personalized text generated directly, while the second emphasizes using personalized information as intermediate steps or implicitly as embeddings to improve the quality of a ➋downstream task such as recommendation systems. See Section \ref{['sec:bridging-the-gap']} for a detailed discussion. An example of an adaptation function $A$ is a retrieval module. Please note that $y$ here can represent user-written text if available, or alternatively, user preferences and separate reward models that reflect user judgments.
• Figure 2: Overview of Personalization Data. This figure presents an overview of the various types of user-specific data used in downstream personalization tasks. It categorizes the data into three primary formats: (i) Static Attributes, which include demographic information and item metadata that remain relatively constant over time; (ii) Interaction History, capturing dynamic user behaviors and preferences through previous activities and engagement data; (iii) User-Written Text, encompassing reviews, dialogues, and social media posts that provide rich insights into user sentiment and preferences; and (iv) Pair-Wise Human Preferences, explicit feedback or annotations that guide the system to align with individual user needs.
• Figure 3: Space of Personalized Generations. We characterize the space of generations for query $x$, including the space of all possible generations $\mathcal{S}(x)$, the space of all high quality generations $\mathcal{S}_h(x)$, and finally, the space of user-specific high-quality personalized generations $\mathcal{S}_i(x)$ for user $i$. Intuitively, given two users $i$ and $j$, the space of high-quality personalized generations for each user may be completely disjoint.
• Figure 4: Dimensions in Personalized Criterion. We propose a framework that expands the dimensions of personalization criteria LLMs along three aspects: (i) Tone and Style, which includes writing style and tone preferences to match the user-written text; (ii) Relevance, encompassing content relevance to user interests and contextual relevance for specific situations; and (iii) Accuracy, which ensures both factual correctness and accurate representation of user data. These aspects interact to form a comprehensive taxonomy, addressing the multi-faceted nature of effective personalization in LLM-generated text.
• Figure 5: Examples of Personalization Tasks and Data.

Theorems & Definitions (21)

• Definition 1: Large Language Model
• Definition 2: Downstream Tasks
• Definition 3: Prompt
• Definition 4: Personalization
• Definition 5: User Preferences
• Definition 6: Personalized Large Language Model
• Definition 7: User Documents
• Definition 8: User Attributes
• Definition 9: User Interactions
• Definition 10: Alignment