LLM-Generated Description and Reasoning: Use-case for Library Recommendations

Recent advances in recommender systems have begun to explore how large language models can enhance user and item representations through natural language. However, the effectiveness of using LLM-generated descriptions and internal reasoning directly within collaborative filtering models remains underexplored. In this work, we evaluate various content-based embeddings, either by directly encoding metadata or by constructing richer user and item profiles using LLMs with internal reasoning. These embeddings are integrated into standard collaborative filtering models without modifying their architectures. Using a real-world library dataset enriched through external book databases, we show that on average, content-based embeddings improve recommendation performance by 40% in recall and NDCG over baseline models, and LLM-generated profiles further improve performance and provide compact, interpretable representations. Our findings suggest that LLMgenerated profiles and reasoning not only enhance explainability, but also improve recommendation quality in sparse, real-world environments.