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

Researchers introduced CogRec, a structure-aware reasoning framework for generative recommendation.

CogRec grounds intermediate reasoning in the same Semantic ID (SID) space used for item generation, bringing hierarchy, graph, and neighborhood structure into recommendation.

It jointly trains natural-language reasoning and SID-routing branches under constrained decoding, reducing reliance on memorized item codes.

Experiments on three sequential recommendation benchmarks showed that routing can improve generation, especially when prefix matching fails but valid SID-space transitions remain.