TopoTok preserves item relationships, boosts RecSys recall by 9.42%

Generative recommendation models often tokenize items to enable autoregressive generation, but this quantization process can distort item relationships in the embedding space. The authors observe that intrinsic adjacency relationships from pretrained semantic embeddings are significantly disrupted after tokenization, misleading the model's perception of item similarity and bottlenecking accuracy. They call this 'topology distortion' and propose TopoTok to solve it.

TopoTok introduces a three-level distillation scheme: Inter-Group Distillation captures global cluster-wise relations, Intra-Group Distillation refines local structures within clusters, and Inter-Item Distinction enforces fine-grained alignment at the item level. On three benchmark datasets, TopoTok consistently outperforms existing tokenizers, achieving up to a 9.42% gain in Recall@5. This work, accepted to RecSys 2026, demonstrates a practical path to more accurate and context-aware sequential recommendations.