L2Rec Fuses Behavioral and Semantic Signals in LLMs for Personalization

Adapting large language models (LLMs) for personalized recommendation has been challenged by the need to align general-purpose capabilities with user-specific preferences from behavioral and semantic signals. Existing approaches either inject behavioral embeddings at the input level or use contrastive alignment at the output level, suffering from distribution gaps or lack of end-to-end supervision. To solve this, researchers from an academic-industrial team propose L2Rec (accepted at SIGIR 2026), which unifies both views at the parameter level of LLMs. The key innovation is a Dual-view Personalized Mixture-of-Experts (DPMoE) mechanism that applies view-specific, personalized low-rank perturbations to the same Transformer backbone, enabling a single LLM to produce complementary behavioral and semantic adaptations with minimal misalignment. An adaptive cross-view fusion module then integrates these outputs into a unified user preference.

L2Rec was evaluated on four benchmark datasets, consistently outperforming state-of-the-art baselines across all metrics. Crucially, the method was validated via online A/B testing on a large-scale industrial recommendation platform, where it drove significant improvements in key engagement metrics such as click-through rate and session length. This demonstrates that parameter-level fusion of behavioral and semantic information is not only technically effective but also practically deployable. The work opens a new direction for LLM-based recommendations by eliminating the need for separate encoders or token-space injections, offering a more elegant and performant solution for real-world personalization systems.