CogRec paper grounds recommendation reasoning in Semantic ID space

anshulkundaje · x · 2026-07-28

CogRec grounds generative recommendation reasoning in Semantic ID space

The post links to a paper, CogRec: Structure-Cognitive Fast-and-Slow Reasoning for Generative Recommendation. The method treats recommendation as constrained sequence generation over hierarchical Semantic IDs, but argues that existing approaches mainly memorize IDs and underuse the hierarchy, intra-layer relations, and item neighborhoods as an explicit reasoning space.

The proposed framework, CogRec, adds:

It introduces SID Routing to express recommendation reasoning with three operations: Match, LateralJump, and Explore. The idea is to combine fast semantic localization with slower structural navigation, then train natural-language reasoning and SID-routing branches from a shared checkpoint in a multi-stage supervised pipeline.

Related event: CogRec Introduces Fast-and-Slow Reasoning for Generative Recommendation(2 posts)→

Original post →

More from Research

Research channel →