Large Language Model (LLM) Revolution in Recommendation Systems: Overview, Techniques, and Future Directions

The rapid development of large language models (LLMs) such as ChatGPT is driving a revolutionary change in recommendation systems, which have traditionally relied on historical user‑item interaction data for prediction.

Conventional recommender models face several difficulties: massive user and item scales, unobservable external factors, and an over‑reliance on ID‑based features that lead to poor generalization and modeling challenges.

LLMs bring strong generalization, rich text understanding, and the ability to replace ID representations with textual ones. Prompt‑learning enables the formulation of recommendation tasks as natural‑language instructions, allowing unified task handling and leveraging powerful pre‑training paradigms.

Representative works include: (1) ID‑free item representation by converting item attributes into long sentences and using BERT‑style encoders; (2) Prompt learning for tasks such as CTR prediction, where prompts describe user history and desired recommendations; (3) M6‑Rec, which combines ID‑free textual representations with prompt‑based task formulation; and (4) two‑stage frameworks that first use LLMs for language‑space understanding, then a recommendation‑space module, and finally an item‑space scoring stage.

Key challenges remain: model interpretability, privacy protection, bias mitigation, high inference costs, and difficulty of fine‑tuning massive models for recommendation-specific objectives.

Future research directions highlighted are personalized prompt optimization, robust prompt design under distribution shift, new recommendation paradigms that integrate generative capabilities for cold‑start items, and methods to reduce LLM‑induced societal bias.

Practical recommendations suggest using the largest feasible foundation models (e.g., GPT‑4), preserving generative ability during fine‑tuning, and fusing statistical signals with LLM outputs to achieve better performance.

The article concludes with a brief introduction to the Data Space Institute of the University of Science and Technology of China, emphasizing its focus on big data, AI, and cyber‑security research.