CITE: Contextual Item Transition Embeddings for Generative Recommendation with Parallel Semantic Identifiers

Recommendation systems are widely used to help users discover relevant products, media, and services from large catalogs. Generative recommendation has emerged as a promising direction because it formulates next-item prediction as a generation task and supports structured representations for large catalogs. In particular, parallel semantic identifiers predict all identifier positions simultaneously, making it possible to use long semantic identifiers. However, scoring candidates solely through parallel semantic code probabilities models item transitions indirectly and lacks an explicit term for the directed relation between the latest item and each candidate. To address this limitation, we propose Contextual Item Transition Embeddings (CITE). CITE introduces an item-level transition score conditioned on the encoded history and the latest item, with separate source and destination embeddings to model the two roles of a directed transition. It keeps the parallel semantic identifier generator fixed and adds the transition score to the original semantic score using a coefficient selected on validation data. Experiments on four Amazon categories using five random seeds show that CITE improves all reported ranking metrics over the parallel semantic identifier baseline, with a mean relative improvement of 23.6% in NDCG@10.