TGR: unified generation, ranking, and offline reasoning injection¶
Core mechanism on public data. Combines hierarchical semantic-ID paths, list objectives, and reason-token injection.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv 2609.00986 |
| Organization | Tencent |
| First public release | 2026-09-01 (arXiv v1) |
| Authors' code | No public author code was found when checked on 2026-09-05. |
| Adapter / method | tgr |
| Local reproduction code | src/auto_research/reproductions/tgr/ |
Original paper¶
Motivation and method¶
TGR combines three industrial recommendation approaches. CCFormer supplies scalable ranking; BARGE and HiGR generate single items or whole-page lists with hierarchical semantic IDs; TGR-Reason injects reason tokens generated offline into online decoding, avoiding costly request-time rollouts.
flowchart LR
F[Unified feature tokens] --> C[CCFormer ranking]
S[Hierarchical semantic IDs] --> B[BARGE and HiGR generation]
O[Offline reason tokens] --> B
C --> Y[Online multi-objective results]
B --> YKey figure from the paper¶
Figure 1: Core architecture, modules, and their connections. From the original paper; copyright remains with the authors.
Core equation¶
The local path accumulates hierarchical-path confidence and combines it with the transition/content list objective: \(s_i=\sum_l p(z_i^l\mid H)\prod_{j\le l}(0.6+p(z_i^j\mid H))+\lambda s_i^{list}\).
Paper results¶
CCFormer reports CTR +3.57% and ad revenue +1.71%. BARGE, HiGR, and TGR-Reason are also validated through online experiments or full releases.
Local reproduction¶
The fixed three-seed MovieLens 100K protocol increases NDCG@10 from 0.05401 to 0.05953. See the complete artifact.
Local comparison: Baseline NDCG@10=0.05401, treatment NDCG@10=0.05953, or +10.24% relative.
Reproduction boundaries¶
The private semantic-ID catalog, production CCFormer checkpoint, and online generation service were not reproduced. This covers only the auditable core mechanism on public data.