L2Rec: Dual-View Personalized Recommendation via LLM Parameter-Level Alignment
L2Rec, an innovative technique, merges behavioral and semantic insights for tailored recommendations by synchronizing these perspectives at the parameter level of extensive language models. Current methods typically combine signals at the input or output stages, leading to distribution discrepancies or insufficient end-to-end guidance. Utilizing a Dual-view Personalized Mixture-of-Experts (DPMoE) framework, L2Rec implements view-specific low-rank adjustments, allowing one LLM backbone to generate both behavioral and semantic representations that complement each other. This research has been made available on arXiv under the identifier 2605.26717.
Key facts
- L2Rec unifies behavioral and semantic understanding at the parameter level of LLMs.
- Existing approaches integrate signals at input or output levels.
- Existing methods suffer from distribution gaps or lack of end-to-end task supervision.
- L2Rec uses a Dual-view Personalized Mixture-of-Experts (DPMoE) mechanism.
- DPMoE applies view-specific, personalized low-rank perturbations.
- A single LLM backbone produces complementary behavioral and semantic representations.
- The paper is on arXiv with ID 2605.26717.
- The method is designed for personalized recommendation.
Entities
Institutions
- arXiv