Reproduce “Rethinking Large Language Model Architectures for Sequential Recommendations” on OpenResearch

About this paper

Hanbing Wang, Xiaorui Liu, Wenqi Fan, Xiangyu Zhao, Venkataramana Kini, Devendra Yadav, Fei Wang, Zhen Wen and 2 more

Lite-LLM4Rec introduces a novel architecture for Large Language Model (LLM)-based sequential recommendation that significantly enhances efficiency while maintaining or improving accuracy. The model achieves an approximate 99% reduction in inference time by employing a hierarchical LLM structure and replacing beam search decoding with a direct item scoring mechanism.

Key insights

• Replacing text generation and beam search with a direct item scoring approach (Item Projection Head) fundamentally addresses the primary efficiency bottleneck in LLM-based recommendation.
• A hierarchical LLM structure, separating item metadata processing (Item LLM) from sequential user behavior modeling (Recommendation LLM), allows for significant reduction in input length and enables pre-computation of item embeddings.
• Careful architecture design, rather than merely adapting text generation for recommendation, is crucial for making LLM-based recommender systems practical and efficient for real-world applications.

Summary by alphaXiv.