Large language model with retrieval semantics for cold-start recommendations.

Abstract

The item cold-start problem remains a major challenge in recommender systems. Existing approaches struggle with three limitations: semantic bias caused by new items that fall outside the pre-training corpus, semantic sparsity caused by limited item attributes, and the absence of interactions, which creates a behavioral gap. To address these challenges, we propose a method called Large Language Model with Retrieval Semantics (LaReS). First, to reduce semantic bias, we introduce self-retrieval contextualization, which retrieves semantically similar warm items as auxiliary knowledge sources. This mechanism grounds cold items in reliable contexts and helps the model infer more reliable item characteristics from warm-item neighbors. Second, to enrich sparse semantics, we propose a large language model-based semantic encoding strategy that transforms augmented textual information into dense, task-specific embeddings. A frozen large language model (LLM) and trainable semantic layers capture deep semantic features beyond simple metadata. Third, to bridge the interaction gap, we implement a joint preference learning framework through multi-objective optimization. This approach distills preference signals from warm items and aligns latent distributions, granting cold items visibility comparable to established items. Experiments and analyses show that LaReS achieves competitive performance in both cold-start and overall recommendation scenarios. For example, LaReS improves cold-start P@5 on MovieLens by 7.98%. The code is available at https://github.com/wuzelong/LaReS.