Scholarlink

<p>The emergence of large language models (LLMs) has fundamentally redefined the landscape of artificial intelligence, offering remarkable capabilities in natural language understanding, generation, and reasoning. In the domain of recommender systems (RS), these advancements present new opportunities for addressing enduring challenges such as the cold-start problem, data sparsity, domain adaptation, explainability, and user engagement. Yet the substantial cost, complexity, and resource demands of fine-tuning these models limit their widespread adoption. This review investigates an alternative paradigm-leveraging <i>untuned</i> LLMs for RS—where general-purpose language models operate without task-specific parameter updates, relying on prompt engineering, zero-shot learning, and in-context learning. By analyzing 41 state-of-the-art articles, we find that untuned LLMs can adeptly generate context-aware, personalized, and human-aligned recommendations at reduced computational overhead, exhibiting notable performance in synthetic user–item interactions, semantic preference inference, zero-shot retrieval, and cross-modality reasoning for text, images, videos, and graph-based data. Our synthesis highlights successful integrations with graph neural networks (GNNs), vision language models (VLMs), and retrieval-augmented frameworks that improve recommendation accuracy, novelty, and user satisfaction, while uncovering key limitations such as hallucination, evaluation inconsistencies, fairness gaps, and inefficiencies in large-scale deployment. To address these concerns, we propose a structured set of future research directions, including scalable model compression (e.g., quantization, LoRA), fairness-aware and human-aligned evaluation metrics, synthetic data quality assurance, and domain adaptation strategies applicable to e-commerce, education, health care, and entertainment. We position untuned LLMs as a transformative force for next-generation RS, offering a scalable, explainable, and adaptive foundation for developing intelligent, user-centric systems in diverse real-world scenarios.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Unleashing the power of untuned large language models in recommender systems: a thorough investigation of current approaches, challenges, and future research directions

  • 期刊论文

  • 发表日期：

    2025年7月16日

查看全文
  • Muniraja Pasupuleti,
  • Shashank Mouli Satapathy

摘要

The emergence of large language models (LLMs) has fundamentally redefined the landscape of artificial intelligence, offering remarkable capabilities in natural language understanding, generation, and reasoning. In the domain of recommender systems (RS), these advancements present new opportunities for addressing enduring challenges such as the cold-start problem, data sparsity, domain adaptation, explainability, and user engagement. Yet the substantial cost, complexity, and resource demands of fine-tuning these models limit their widespread adoption. This review investigates an alternative paradigm-leveraging untuned LLMs for RS—where general-purpose language models operate without task-specific parameter updates, relying on prompt engineering, zero-shot learning, and in-context learning. By analyzing 41 state-of-the-art articles, we find that untuned LLMs can adeptly generate context-aware, personalized, and human-aligned recommendations at reduced computational overhead, exhibiting notable performance in synthetic user–item interactions, semantic preference inference, zero-shot retrieval, and cross-modality reasoning for text, images, videos, and graph-based data. Our synthesis highlights successful integrations with graph neural networks (GNNs), vision language models (VLMs), and retrieval-augmented frameworks that improve recommendation accuracy, novelty, and user satisfaction, while uncovering key limitations such as hallucination, evaluation inconsistencies, fairness gaps, and inefficiencies in large-scale deployment. To address these concerns, we propose a structured set of future research directions, including scalable model compression (e.g., quantization, LoRA), fairness-aware and human-aligned evaluation metrics, synthetic data quality assurance, and domain adaptation strategies applicable to e-commerce, education, health care, and entertainment. We position untuned LLMs as a transformative force for next-generation RS, offering a scalable, explainable, and adaptive foundation for developing intelligent, user-centric systems in diverse real-world scenarios.