LLM-Rec: Personalized Recommendation via Prompting Large Language Models

The evolution of recommendation systems has shifted from collaborative filtering towards content-based and knowledge-enhanced methods. Deep learning models like DeepCoNN and BERT4Rec have improved user preference modeling, but they still depend heavily on rich, structured data. In many real-world scenarios, item descriptions are brief or noisy, limiting the effectiveness of traditional approaches. The advent of large language models (LLMs) such as GPT series and LLAMA has demonstrated their capacity for extensive knowledge storage, complex reasoning, and natural language understanding. Prior work has explored knowledge graph integration and feature engineering, yet these methods often require domain expertise and manual effort. Recent developments suggest that prompting LLMs can serve as a flexible, scalable alternative for text augmentation, but systematic strategies for optimizing prompts remain underexplored.

The core challenge lies in the incomplete or superficial descriptions of items, which hinder accurate personalization. Existing models struggle to extract meaningful features from sparse text, especially in user-generated content domains like recipes or social media posts. External knowledge bases can help but are costly and domain-specific. Moreover, manual feature engineering is labor-intensive and lacks scalability. How to leverage the intrinsic reasoning and knowledge capabilities of LLMs to automatically generate richer, more informative descriptions remains an open question. Addressing this gap could significantly enhance recommendation accuracy, interpretability, and user satisfaction.

This paper introduces a multi-strategy prompting framework that systematically guides LLMs to produce enriched item descriptions. The four prompts—basic paraphrasing, tagging, inference, and combined recommendation+engagement—are designed to capture different aspects of item features. This approach is needed to address the limitations of original descriptions and to exploit LLMs’ reasoning abilities fully. Unlike prior methods that rely on external data or complex feature extraction, this framework directly manipulates natural language inputs, making it adaptable and easy to implement. The integration of user behavior into prompts further personalizes descriptions, bridging the gap between generic content and individual preferences. This innovation creates a new paradigm for text-based recommendation enhancement, emphasizing prompt design as a key lever.

• �� Input original item description into LLM with specific prompts tailored for different strategies.
• �� Basic prompts (para, tag, infer) instruct the model to paraphrase, summarize, or categorize the description.
• �� Recommendation-driven prompts (prec) ask the model what additional info or tags would improve recommendation.
• �� Engagement-guided prompts (peng) incorporate user interaction data, combining descriptions of related items to generate personalized summaries.
• �� The generated enriched texts are concatenated and used as input to an MLP model for predicting user-item interactions.
• �� The training objective maximizes the likelihood of observed interactions, using cross-entropy loss.
• �� Evaluation involves metrics like Precision@10, Recall@10, and NDCG@10, comparing the impact of different prompts.
• �� Experiments test various prompt combinations and models (GPT-3, LLAMA-2) across datasets, validating the framework’s robustness and effectiveness.

The experimental setup involves two benchmark datasets: Movielens-1M for movies and Recipe for culinary content. Baselines include traditional content-based models, feature engineering approaches, and recent text augmentation methods. The models employ GPT-3 and LLAMA-2 for prompt generation, with hyperparameters tuned via validation sets. The negative sampling ratio is set at 1000:1, ensuring evaluation fairness. The recommendation module remains a simple MLP, emphasizing the contribution of augmented text. Multiple prompt strategies are tested, with performance measured through Precision@10, Recall@10, and NDCG@10 across five data splits. Ablation studies analyze the effect of each prompt type, confirming the superiority of combined prompts.

Results show that LLM-Rec consistently outperforms baselines, with NDCG@10 gains of up to 12.57%. GPT-3-based prompts yield the highest improvements, demonstrating the importance of prompt design. The enriched descriptions contain more detailed and expressive features, which help the MLP model better understand item characteristics. Ablation results indicate that combining recommendation and engagement prompts (prec+peng) produces the best performance, highlighting the synergy of semantic richness and personalization. The models maintain robustness across different datasets and prompt variations, confirming the framework’s generalizability.

This approach is highly applicable in e-commerce, streaming platforms, and social media where item descriptions are often sparse or user-generated. It enables rapid deployment of personalized recommendation engines with minimal domain-specific tuning. The method also facilitates explainability, as enriched texts can be directly inspected to understand recommendation rationale. Future integration with multi-modal data (images, audio) could further enhance personalization, making the system adaptable to emerging content types and user preferences.

The reliance on large pre-trained models incurs high computational costs, limiting real-time deployment in resource-constrained environments. Prompt sensitivity may cause variability in outputs, requiring careful tuning. The approach’s effectiveness depends on the quality of initial descriptions and may degrade with extremely noisy or irrelevant inputs. Future work should focus on automating prompt optimization, reducing model size, and exploring multi-modal extensions to overcome these limitations.