A Survey of Personalized Large Language Models: Progress and Future Directions

TL;DR

<3-5 sentence high-level summary> The paper identifies the gap between general LLM capabilities and the need for user-specific personalization, proposing PLLMs as a solution. It introduces a three-level taxonomy—Personalized Prompting (input level), Personalization Adaptation (model level), and Personalization Alignment (objective level)—and systematically reviews methods, data types, benchmarks, and evaluation metrics. It analyzes the state-of-the-art, discusses limitations, and outlines future directions such as lifelong updating, multimodal data, edge computing, and trust-aware personalization. The authors advocate memory-centric PLLMs that remember, adapt, and evolve while balancing efficacy, efficiency, and privacy for practical deployment.

Abstract

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.

Figures (8)

• Figure 1: Comparison of General LLM vs. Personalized LLM. Different users have different preferences. General LLMs fail to satisfy diverse needs with a one-size-fits-all approach, demonstrating the necessity for personalized LLMs.
• Figure 2: Illustration of PLLM techniques for generating personalized responses through three levels: prompting (input level, \ref{['sec: personalized prompting']}), adaptation (model level, \ref{['sec: personalized adatation']}), and alignment (objective level, \ref{['sec: personalized alignment']}). \ref{['para:personalized data']}
• Figure 3: A taxonomy of PLLMs with representative examples.
• Figure 4: Examples of the personalized data and query type. The human brain regions (hippocampus as spatiotemporal memory integrator and experience simulator, angular gyrus as conceptual processing hub, and default mode network (including medial prefrontal cortex) as creative ideation and remote association hub) serve as analogical references to understand these query types, though the correspondence may not be strictly one-to-one.
• Figure 5: The illustration of personalized prompting approaches: (a) Profile-Augmented Prompting summarizes personalized data into tokens and concatenates with queries; (b) Retrieval-Augmented Prompting retrieves relevant records from memory and combines with queries; (c) Soft-Fused Prompting encodes personalized data into embeddings integrated via input prefix, cross-attention, or output logits; (d) Contrastive Prompting compares model outputs with and without personalized information to extract personalization factors.