REPREC: Lightweight LLM-Based Sequential Recommendation via User Embedding Alignment

A team of researchers has introduced REPREC, an efficient framework designed for sequential recommendations utilizing large language models. This method transforms a fixed-size user embedding from a static sequential encoder into soft tokens through a simple MLP injector, allowing a frozen LLM to operate without the need for fine-tuning or extra components. This innovation lowers both training complexity and deployment expenses while preserving personalization. Notably, it circumvents item-level conditioning across extensive histories and keeps the pretrained backbones intact. Experimental results validate its effectiveness.