L2Rec: Towards Dual-View Understanding of LLMs for Personalized Recommendation

L2Rec introduces a novel framework that adapts large language models for personalized recommendations by unifying behavioral and semantic signals at the parameter level using a Dual-view Personalized Mixture-of-Experts mechanism. The approach demonstrates superior performance across multiple datasets and validates real-world applicability through industrial A/B testing.

L2Rec addresses a fundamental challenge in applying large language models to recommendation systems: the need to balance general-purpose language understanding with user-specific behavioral patterns. Traditional approaches have struggled with either input-level integration (injecting behavioral data into token spaces) or output-level fusion (separate encoders with post-hoc alignment), both creating representation misalignment and limiting end-to-end optimization.

The research emerges from the broader trend of specialized fine-tuning for LLMs, where practitioners recognize that pre-trained models require domain-specific adaptation to maintain quality. Personalized recommendation systems represent a high-stakes application where users expect relevance beyond generic content ranking. The dual-view architecture allows a single LLM backbone to maintain both semantic understanding and behavioral sensitivity through low-rank perturbations, a parameter-efficient approach that resembles techniques like LoRA.

For practitioners building recommendation platforms, L2Rec offers measurable advantages: consistent improvements over baselines across four datasets and validated gains in engagement metrics from industrial deployment. This validates that parameter-level unification outperforms separated signal processing. The adaptive cross-view fusion module enables dynamic weighting of behavioral versus semantic preferences, potentially improving performance across diverse user segments.

Future development hinges on scaling validation across additional platforms and testing robustness to distribution shifts in user behavior. The framework's computational efficiency compared to alternatives remains unclear, which matters for cost-sensitive production deployments. Broader questions about whether this approach generalizes to other dual-signal recommendation tasks could determine its long-term impact on LLM applications beyond personalization.

• →L2Rec unifies behavioral and semantic signals at the LLM parameter level using Dual-view Personalized Mixture-of-Experts, eliminating representation gaps inherent in input/output-level integration approaches.
• →Industrial A/B testing validates significant improvements in engagement metrics, demonstrating real-world applicability beyond academic benchmarks.
• →The parameter-efficient design enables a single LLM backbone to produce complementary user-specific adaptations with minimal computational overhead.
• →Cross-view fusion module dynamically integrates dual-view outputs, allowing adaptive weighting between behavioral and semantic preferences.
• →Consistent improvements across four datasets suggest the approach generalizes across recommendation domains without domain-specific retraining.