TopoGR: Revealing and Preserving Latent Structure of Semantic ID in Generative Recommendation

Generative recommendation (GR) models typically represent items as sequences of discrete semantic IDs (SIDs). A major limitation of current methods is the structural mismatch between the tokenization stage—which learns a structured code space—and the generation stage, which treats these SIDs as independent categorical symbols. This causes the model to ignore the latent semantic proximity between items that do not share identical IDs. To solve this, the authors propose TopoGR, a framework that utilizes Bit-decomposable Semantic IDs (Binary SIDs).

TopoGR introduces a Bit-Decomposable Quantizer (BDQ) that represents each item as a sequence of binary codes. Unlike standard integer IDs, these binary codes are learned through a bit-compositional reconstruction mechanism, which ensures that the Hamming distance between codes reflects the semantic similarity of the items. The model incorporates this topology into the recommendation process through three mechanisms: using binary features as input, applying Hamming-aware supervision during training, and employing Hamming-consistent reranking during inference.

TopoGR demonstrates that the latent topology of the SID space can be effectively preserved and exploited for better recommendation. By feeding binary SID features directly into the sequence model, the generator gains explicit access to the structural relationships between items. The Hamming-aware supervision further improves the model by penalizing incorrect predictions based on their distance from the target in the binary space, rather than treating all errors as equally wrong. Experimental results across four benchmark datasets show that TopoGR consistently outperforms existing state-of-the-art generative and discriminative recommendation baselines.

This work addresses a fundamental bottleneck in generative recommendation: the loss of semantic information during the discretization process. By moving from opaque categorical tokens to structured binary representations, TopoGR allows generative models to generalize better to semantically related items, even when those items have different IDs. This approach bridges the gap between the structured latent spaces learned by tokenizers and the sequential generation process, providing a more robust framework for next-item prediction.