Application of Large Language Models in Recommendation Systems: Overview and Future Directions

Introduction: This article gives an overview of the application of large language models (LLMs) in recommendation systems.

It first introduces two application paradigms: LLM+RS, where LLM is integrated as a part of the recommendation pipeline, and LLM AS RS, where the LLM serves as an end‑to‑end recommender.

LLM+RS impacts various stages such as user/item representation, recall matching (i2i, u2i), ranking (point‑wise, pair‑wise, list‑wise), generation, and addresses fairness, bias, privacy, and explainability.

Prompt‑based approaches influence LLM input/output, affecting user/item representation, recall, ranking/generation, and help solve issues like fairness, cold‑start, and explainability.

LLM AS RS treats the LLM as a complete recommender, employing pre‑training & fine‑tuning for top‑K recommendation, rating prediction, conversational recommendation, and explainability, as well as prompting for lightweight recommendation.

Future directions include automatic prompt design using user context, multimodal recommendation capabilities, solving hallucination and bias, improving system performance to meet C‑end requirements, and exploring agents that leverage short‑ and long‑term user context.

The article concludes with acknowledgments of the speaker and editor.