A Survey on Large Language Models for Recommendation Systems

Abstract

Large language models (LLMs) offer new approaches to overcoming the limitations of recommendation systems in representation, user modeling, and decision-making. Accordingly, recommendation systems are evolving from paradigms primarily based on behavioral statistics and discriminative learning toward new paradigms featuring semantic understanding, reasoning-based decision-making, and autonomous interaction. This paper systematically reviews recent advances in LLM-empowered recommendation systems. The principal advantages of LLMs in semantic modeling, knowledge transfer, and generative reasoning are analyzed. Then, the representative technical paradigms including representation-enhanced recommendation, user-interest and context enhancement, prompt learning and task reformulation, and generative recommendation are subsequently reviewed and summarized. Finally, the paper analyzes major challenges related to hallucination, inference efficiency, privacy and security, fairness and trustworthiness, and recommendation agents, and outlines future research directions.