SPARC: Sequence-aware Progressive Attribute Routing and Compression Framework for Generative Recommendation

Generative recommendation models typically represent items as discrete Semantic IDs (SIDs). While these SIDs capture stable item identity, they often fail to incorporate dynamic, interaction-specific context—such as price, behavior type, or timestamp—without significantly increasing the input sequence length. The authors investigate how to enrich historical item representations with these heterogeneous attributes without incurring the prohibitive computational costs associated with longer input sequences in Transformer-based generative backbones.

To address this, the authors propose SPARC (Sequence-aware Progressive Attribute Routing and Compression). The framework operates in three stages:

1. Field-wise Context Modeling (FCM): It models the sequential evolution of each field type (e.g., price trends over time) to create context-aware representations.
2. Context-aware Attribute Routing (CAR): It separates SID identity tokens from side attributes. It then uses a routing mechanism to dynamically assign side information into a fixed number of 'slots' based on the current sequence context, ensuring that only the most relevant information is preserved.
3. Sequence-level Token Consolidation (STC): It performs lightweight cross-interaction modeling to integrate these intermediate tokens into a single, compact representation for each historical interaction.

By contextualizing and compressing information before it reaches the generative backbone, SPARC allows models to leverage rich, multi-field industrial data without the memory and computational overhead of longer input sequences. This enables better performance in industrial recommendation scenarios where dynamic states (like price or intent) are just as critical as the item's static identity. Experiments on Taobao and Amazon datasets show that SPARC consistently outperforms both conventional and generative baselines, confirming that the performance gains stem from context-conditioned information retention rather than simply increasing model capacity.