Abstract
Sequential recommenders have recently achieved strong gains by modeling the temporal structure of user behavior, yet most systems still rely on shallow item representations that capture co-occurrence statistics rather than the underlying semantics of content. This paper proposes the Semantic Resonance System (SRR), a unified architecture that integrates deep semantic signals from a large language model (LLM) with standard behavioral embeddings for sequential recommendation. SRR uses a domain-adapted Llama 2 13B model, fine-tuned via ParameterEfficient Fine-Tuning (LoRA), to produce high-fidelity semantic item embeddings that are fused with behavioral embeddings through a Dynamic Cross-Attention module. A central component of SRR is the Semantic Resonance Optimization (SRO) loss, which jointly optimizes (i) a Dense All-Action (DAA) loss for high-recall retrieval, (ii) a Binary Cross-Entropy (BCE) loss for calibrated scores, and (iii) a Contrastive Alignment loss $\mathcal{L}_{C A}$ that explicitly aligns semantic and behavioral latent spaces. On the large-scale MovieLens $\mathbf{2 5 M}$ dataset, SRR achieves a $\mathbf{1 6. 4 \%}$ improvement in NDCG@10 over a strong transfer-learning baseline, while also improving diversity, catalog coverage, and cold-start performance. A latency-effectiveness study further shows that SRR can be configured along a smooth Pareto frontier between NDCG@10 and inference cost, supporting deployment under different resource budgets.