Semantic IDs in Generative Recommendation: Understanding and Improving Item Selection

Semantic IDs (SIDs) have become a cornerstone of modern generative recommendation systems, serving multiple roles: organizing related items via shared prefixes, uniquely identifying individual items, and narrowing down potential recommendations during token generation. This study delves into the intricacies of SIDs, from their construction and item encoding to their impact on autoregressive generation and final recommendations. It uncovers that while SIDs effectively provide broad organizational structure, they often fail to preserve the finer local relationships between items that are captured by the encoder. The research demonstrates that SID neighborhoods recover only a fraction of the encoder's nearest neighbors, indicating a loss of fine-grained item structure. Furthermore, even minor changes in item descriptions can alter exact SIDs, suggesting that SID tokens are not solely determined by item meaning. This loss of fine structure becomes particularly problematic during the generation phase, where SID filtering can prematurely discard a significant portion of plausible recommendations. To mitigate these issues, the paper proposes Item-Supported Decoding (ISD), a lightweight inference-time method. ISD allows a user-specific item ranking to guide the selection of SID prefixes before beam search, and then uses the same ranking to order the final generated items. Crucially, ISD requires no additional parameters or retraining. Empirical evaluations across various settings show that ISD consistently improves NDCG@10 scores, with relative gains up to 31.2%, by preventing the premature pruning of relevant items and better leveraging the underlying item representations.