A Comprehensive Review on Harnessing Large Language Models to Overcome Recommender System Challenges

TL;DR

This work surveys how large language models (LLMs) can address core bottlenecks in contemporary recommender systems, including data sparsity, cold-start, and dynamic user behavior, by enabling unified, language-native reasoning across candidate retrieval, ranking, and dialogue. It organizes a broad design space around prompt-driven retrieval, retrieval-augmented generation (RAG), zero-shot personalization, and multimodal representation learning, illustrating how LLMs can bootstrap representations, reason over content, and ground recommendations in external knowledge. The paper presents a taxonomy of data-centric, modeling/algorithmic, evaluation, and privacy/regulatory challenges, and proposes a framework of trade-offs between accuracy, scalability, and real-time performance. It argues that LLMs are not merely auxiliary components but foundational enablers for adaptive, semantically rich, and user-centric recommender systems, while candidly acknowledging latency, hallucination, drift, and governance concerns that must be overcome for production deployment.

Abstract

Recommender systems have traditionally followed modular architectures comprising candidate generation, multi-stage ranking, and re-ranking, each trained separately with supervised objectives and hand-engineered features. While effective in many domains, such systems face persistent challenges including sparse and noisy interaction data, cold-start problems, limited personalization depth, and inadequate semantic understanding of user and item content. The recent emergence of Large Language Models (LLMs) offers a new paradigm for addressing these limitations through unified, language-native mechanisms that can generalize across tasks, domains, and modalities. In this paper, we present a comprehensive technical survey of how LLMs can be leveraged to tackle key challenges in modern recommender systems. We examine the use of LLMs for prompt-driven candidate retrieval, language-native ranking, retrieval-augmented generation (RAG), and conversational recommendation, illustrating how these approaches enhance personalization, semantic alignment, and interpretability without requiring extensive task-specific supervision. LLMs further enable zero- and few-shot reasoning, allowing systems to operate effectively in cold-start and long-tail scenarios by leveraging external knowledge and contextual cues. We categorize these emerging LLM-driven architectures and analyze their effectiveness in mitigating core bottlenecks of conventional pipelines. In doing so, we provide a structured framework for understanding the design space of LLM-enhanced recommenders, and outline the trade-offs between accuracy, scalability, and real-time performance. Our goal is to demonstrate that LLMs are not merely auxiliary components but foundational enablers for building more adaptive, semantically rich, and user-centric recommender systems

Figures (9)

• Figure 1: Taxonomy of recommender system challenges and corresponding LLM-based solutions. The framework categorizes challenges into four major areas—data-centric, modeling and algorithmic, evaluation, and privacy/security—each with sub-problems and LLM-driven strategies for addressing them.'=
• Figure 2: Illustration of representation bootstrapping solution using LLM to generate dense user and item embeddings from textual or metadata descriptions. Newly onboarded users and newly listed items are mapped to the embedding space via LLM-based encoders, enabling effective integration into the recommender system without historical interaction data.
• Figure 3: Overview of the Proposed Multi-Modal Sequence Learning Framework.(a) The model processes timestamp, ID, text, and image sequences using BERT, ViT, and ID embeddings, followed by Interactive and Temporal MoE modules for expert routing. Outputs are fed into Transformers for prediction. (b) Multi-task learning includes category prediction, contrastive learning on ID embeddings, and placeholder-based contrastive learning across modalities zhang2025hierarchicaltimeawaremixtureexperts.
• Figure 4: The image compares MAML-based few-shot adaptation (left) to standard fine-tuning without MAML (right) for a sine wave regression task. MAML generalizes well, even in regions without data, by capturing the underlying periodic structure. In contrast, the pretrained model struggles to adapt from few examples, failing to extrapolate due to conflicting gradients from pretraining finn2017modelagnosticmetalearningfastadaptation
• Figure 5: Overview of the LLM4CDSR framework for Cross-Domain Sequential Recommendation. The architecture first constructs hierarchical user profiles by partitioning user history and generating summaries using LLMs. A tri-thread framework integrates user behavior across source and target domains through self-attention layers. An adapter connects local embeddings with a global LLM-based representation to enable unified personalization across domains. liu2025bridgedomainslargelanguage