How Can Recommender Systems Benefit from Large Language Models: A Survey

Analysis

Why This Paper Matters

Recommender systems are central to managing information overload in online services, yet conventional recommendation models (CRMs) often lack open-domain world knowledge and struggle to infer users' underlying preferences and motivations. Large language models (LLMs) have demonstrated impressive general intelligence, including extensive open-world knowledge, logical and commonsense reasoning, and comprehension of human culture and society. This survey is significant because it systematically explores how LLMs can compensate for the limitations of CRMs, providing a timely and comprehensive overview of a rapidly growing research direction.

The paper's bird's-eye view from the perspective of the whole recommendation pipeline makes it particularly valuable. By organizing existing research along two orthogonal aspects—where to adapt LLMs to recommender systems and how to adapt them—the survey offers a clear and actionable framework for researchers and practitioners. This structure helps clarify the roles LLMs can play at different stages, from feature engineering to pipeline control, and the training and inference strategies that determine their integration.

Moreover, the survey highlights critical challenges in efficiency, effectiveness, and ethics, which are essential considerations for deploying LLM-enhanced recommender systems in real-world applications. By discussing future prospects, it also sets a research agenda that can guide subsequent work.

Technical Contributions

• Where to adapt LLMs: The survey identifies five roles for LLMs in the recommendation pipeline: feature engineering, feature encoder, scoring/ranking function, user interaction, and pipeline controller. This taxonomy provides a comprehensive map of integration points.
• How to adapt LLMs: It introduces a fine-grained taxonomy based on two criteria: whether to tune the LLM during training and whether to involve CRM for inference. This yields four categories of training and inference strategies, enabling systematic comparison.
• Detailed analysis and development paths: For each dimension, the paper provides detailed analysis and outlines general development paths, helping researchers understand the evolution and trade-offs of different approaches.
• Challenge identification: The survey highlights key challenges from three aspects—efficiency, effectiveness, and ethics—offering a balanced view of the opportunities and risks of LLM-enhanced recommender systems.

Results

As a survey paper, it does not present new experimental results or quantitative comparisons. Instead, it synthesizes findings from existing literature, categorizes methods, and discusses their strengths and weaknesses qualitatively. The paper's contribution lies in its organizational framework and analysis rather than empirical metrics.

Significance

The survey's significance lies in providing a structured foundation for the emerging field of LLM-enhanced recommender systems. By mapping the landscape and identifying challenges, it helps researchers navigate the rapidly expanding literature and design more effective and ethical systems. It also underscores the potential of LLMs to transform recommendation by leveraging their common knowledge and reasoning capabilities, while cautioning about efficiency and ethical concerns. This work is likely to influence both academic research and industrial practice in personalized recommendation.