CogRec: Structure-Cognitive Fast-and-Slow Reasoning for Generative Recommendation
Xiang Liu, Jingsong Su, Shuqi Zhao, Pengbo Mo, Yiming Qiu, Huimu Wang, Mingming Li, Jiao Dai, Jizhong Han, Songlin Hu
cs.IR
2026-07-27
CogRec moves recommendation reasoning out of natural-language rationales and into the same Semantic-ID topology used to generate targets, via layer-wise Match/LateralJump/Explore operations (fast/slow). On three benchmarks, SID Routing beat direct generation on Hit@10 in two and hurt on Toys; gains concentrate where prefix matching is insufficient.
Generative recommendation represents each item as a hierarchical discrete Semantic-ID (SID) sequence and turns next-item prediction into constrained sequence generation. Most existing work treats the SID only as a target sequence to memorize, leaving its hierarchy, intra-layer relations, and item neighborhoods unused as an explicit reasoning space. Another line adds reasoning by generating a natural-language rationale before the item, but that rationale couples weakly with the discrete SID space where the final prediction happens. CogRec grounds the reasoning process in the same SID topology used to generate the target.
CogRec augments the SID topology with three layers of structure: a vertical hierarchy (4 layers, 256 codes each), intra-layer semantic graphs (built with residual K-means, top-16 neighbors per centroid, similarity threshold 0.15), and item-level neighborhoods (an HNSW index).
On this topology it defines three layer-wise operations that make up SID Routing:
Match is fast semantic localization; LateralJump and Explore are slower structural navigation. Training is multi-stage: align the SID vocabulary, train direct SID generation, then from one checkpoint train both the natural-language reasoning branch and the SID Routing branch under a shared trie-constrained output space. The backbone is Qwen3-1.7B.
Three public sequential-recommendation benchmarks (Beauty, Sports, Toys), compared against direct generation.
| Dataset | Direct Hit@10 | Routing Hit@10 | Delta |
| Beauty | 0.0842 | 0.0854 | +0.0012 |
| Sports | 0.0427 | 0.0473 | +0.0046 |
| Toys | 0.0901 | 0.0877 | -0.0024 |
Routing improves Hit@10 in two datasets, most clearly on Sports, and hurts on Toys. Against the OneRec-Think baseline (Direct 0.0838/0.0416/0.0843), CogRec's own direct generation is only marginally ahead, strongest on Toys. The conclusion is restrained: structure-grounded reasoning helps only when prefix matching is insufficient but learnable SID-space transitions remain; long or weakly supported routes just add decoding cost and accumulated error.
The value is the angle. Reasoning is not the same as generating longer natural-language explanations; grounding it in the same structured space as the target is more effective. The idea of turning the SID topology (hierarchy, intra-layer graph, neighborhoods) into an explicit reasoning space is portable to other generative recommenders. Be honest about it: the numbers are incremental, the routing gain is small and conditional, and this is not a SOTA to sell.
The authors state the limits plainly. Routing generates longer, more structured traces and needs 5 samples at temperature 1.5 with beam size 10, up to 50 hypotheses per instance, so decoding cost is real. Every extra autoregressive token is another chance for the trace to drift. Gains concentrate in a specific difficulty band; on hard instances long-range SID navigation stays hard and absolute numbers are low across the board. One more thing on reading it: the Routing-versus-Direct gap is negative on Toys, so this structured reasoning is not monotonically beneficial. Where it is worth the decoding cost is a conditional answer, not a universal guarantee.