CogRec: Structure-Cognitive Fast-and-Slow Reasoning for Generative Recommendation

Generative recommendation models often treat Semantic IDs (SIDs) as target sequences to be memorized, failing to utilize the underlying hierarchy and semantic relations as an explicit reasoning space. The authors investigate whether grounding intermediate reasoning within the SID topology—rather than relying on natural-language rationales—can improve recommendation accuracy and provide a more structured, interpretable generation process.

CogRec introduces a "structure-cognitive" framework that transforms the SID space into a navigable graph. The authors augment the standard vertical SID hierarchy with two additional components: intra-layer semantic graphs (connecting semantically related codes at the same level) and item-level neighborhoods.

This topology supports "SID Routing," a reasoning process defined by three operations:

• Match: Fast semantic localization for cases where the anchor and target share the same SID prefix.
• LateralJump: Navigation between semantically related codes at the same layer.
• Explore: Handling transitions where no direct graph edge exists, inferred from user interest patterns.

The model is trained using a multi-stage pipeline that aligns SID tokens and establishes a shared checkpoint for both direct generation and SID-routing branches, ensuring that the reasoning format is the primary variable under investigation.

Empirical results on three sequential-recommendation benchmarks demonstrate that SID Routing consistently outperforms direct generation in specific scenarios. The study reveals that structure-grounded reasoning is particularly effective when the user's intent is complex enough that simple prefix matching fails, yet the target item remains reachable through learnable transitions in the SID space. Conversely, the authors identify that excessive reliance on long or weakly supported routes can lead to increased decoding latency and higher error rates, suggesting that the model's "reasoning" effort should be balanced against the complexity of the recommendation task.