LLM Knowledge Distillation Enhances Sequential Recommenders Efficiently
Researchers propose a novel knowledge distillation method that uses pre-trained Large Language Models (LLMs) to generate textual user profiles, which are then integrated into sequential recommender systems. This approach improves user understanding without requiring LLM inference at serving time, maintaining the efficiency of traditional models. The method avoids architectural modifications or LLM fine-tuning, addressing prohibitive inference costs of existing LLM integration. The paper is submitted to arXiv under Computer Science > Information Retrieval.
Key facts
- Sequential recommender systems model temporal user behavior but lack rich user semantics.
- LLMs offer reasoning capabilities to enhance user understanding.
- Existing LLM integration approaches create prohibitive inference costs in real time.
- The proposed method uses textual user profiles generated by pre-trained LLMs.
- No LLM inference is required at serving time.
- The approach maintains inference efficiency of traditional sequential models.
- No architectural modifications or LLM fine-tuning are needed.
- The paper is available on arXiv (2604.21536).
Entities
Institutions
- arXiv