CogRec brings fast-and-slow reasoning to Semantic-ID recommendation

_reachsumit · x · 2026-07-28

CogRec proposes fast-and-slow reasoning for generative recommendation

The paper CogRec: Structure-Cognitive Fast-and-Slow Reasoning for Generative Recommendation argues that Semantic-ID-based recommendation methods underuse the structure of the discrete ID space.

Main idea:

CogRec introduces SID Routing, which models recommendation reasoning with layer-wise operations such as Match, LateralJump, and Explore. The paper frames Match as fast semantic localization and the other operations as slower structural navigation, then trains natural-language and SID-routing reasoning branches from a shared checkpoint.

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

Original post →

More from Research

Research channel →