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

The article introduces the application of large language models (LLMs) in recommendation systems, outlining two primary paradigms: LLM+RS (LLM as a component of a recommendation system) and LLM AS RS (LLM as an end‑to‑end recommendation system).

LLM+RS (LLM as a part of the recommendation system)

LLM can be incorporated during pre‑training and fine‑tuning stages to enhance user and item representations, improve recall matching (i2i and u2i), and influence ranking through point‑wise, pair‑wise, or list‑wise methods. Additionally, LLMs address fairness, bias, privacy, and explainability concerns.

Prompt‑based techniques further extend LLM usage by leveraging in‑context learning to affect user/item encoding, recall logic, and ranking/generation without additional model training, thereby reducing resource consumption.

LLM AS RS (LLM as a complete recommendation system)

In this paradigm, a large model directly handles the entire recommendation pipeline, including top‑K recommendation, rating prediction, conversational recommendation, and generative recommendation, using both pre‑training/fine‑tuning and prompting strategies. Techniques such as role injection, task description, and multimodal context are employed to improve performance and interpretability.

Future Outlook

The article highlights several open challenges: automatic prompt design using user and item context, enhancing multimodal recommendation capabilities, mitigating hallucination and bias, meeting high‑performance requirements on the client side, and exploring agent‑based recommendation that can invoke diverse tools and leverage short‑term and long‑term user contexts.

Overall, the survey offers readers a thorough understanding of the current state and future prospects of LLMs in recommendation systems.