Understanding Semantic IDs: From Item Representation to Item Selection in Generative Recommendation

Generative recommendation systems often use Semantic IDs (SIDs)—discrete token sequences derived from item content—to represent items. While these IDs are intended to both organize related items and uniquely identify them, they also serve as the basis for autoregressive generation. The authors investigate whether the dual role of SIDs as both semantic organizers and selection filters creates a bottleneck, where the model prematurely discards relevant items during the beam search process.

To evaluate SIDs, the authors perform a controlled behavioral analysis. They create alternative descriptions for items by reordering their fields (e.g., title vs. description) to test if the SID constructor is truly semantic or if it relies on arbitrary presentation. They then trace the generation process of held-out target items to see at which point they are pruned from the beam search. Finally, they propose Item-Supported Decoding (ISD), an inference-time intervention that uses an external item ranking (e.g., from SASRec or UniSRec) to boost the probability of SID prefixes associated with highly ranked items, preventing them from being discarded during beam reduction.

• Loss of Local Structure: While SIDs maintain broad semantic organization, they fail to capture the fine-grained local structure of the continuous embedding space, recovering only 32.2% of the encoder's top-10 neighbors on average.
• Premature Pruning: The autoregressive nature of SID generation means that early token decisions act as irreversible filters. The authors found that after the final semantic token, models like TIGER retain only 29.9% of targets that were plausible recommendations before filtering.
• Performance Gains: ISD consistently improves NDCG@10 across all evaluated domains and models, with relative gains of up to 31.2%. Because it operates at inference time, it requires no retraining of the underlying SID constructor or the decoder.