简体中文 · English
TSGR: Generative retrieval for Taobao Search¶
Fidelity: core-mechanism reproduction. Runs parallel value Semantic IDs and a jointly trained VRM. Private Taobao logs and production serving are not reproduced.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv 2607.18796 |
| Institution | Taobao & Tmall Group of Alibaba / Zhejiang University |
| First public release | 2026-07-21 (arXiv v1) |
| Author code | No official or author-released code found in the paper (checked 2026-08-09). |
| Local adapter or method key | tsgr |
| Local reproduction code | src/auto_research/reproductions/tsgr/ |
Original paper summary¶
Background and main changes¶
Taobao Search must retrieve relevant products while incorporating commercial value into generation. TSGR uses residual quantization for a semantic prefix and assigns global and query-conditioned value codes in parallel to products sharing that prefix. A Value-aware Ranking Module (VRM), trained jointly with the generative backbone, does fine ranking without appending an independent pre-ranker to the generator.
flowchart LR
A["Product content and behavior"] --> B["Residual quantization"]
B --> C["Semantic prefix"]
C --> D["Global value code"]
C --> E["Query-conditioned value code"]
D --> F["Generative retrieval"]
E --> F
F --> G["Joint VRM ranking"]Key figure from the paper¶
Figure 1: The core architecture, its main modules, and their connections. From the paper; copyright belongs to the authors. Click for the source.
Core equations¶
The item identifier combines semantic and parallel value codes:
VRM jointly scores the generative state, product-side information, and value signals:
Offline and online results¶
Paper offline HR@1000 improves 9.16% relative. In a 38-day A/B test on 1% of Taobao Search traffic, IPV rose +0.43%, completed transactions +1.12%, and GMV +1.64%. The paper reports full deployment.
Local reproduction¶
On MovieLens 100K, the local run builds a two-level residual semantic prefix and global/conditional value codes. It fits the joint VRM using 23,027 weighted positives and 115,135 negatives, choosing the blend weight on validation only.
Local comparison: The baseline uses semantic retrieval and a proxy for traditional pre-ranking. The experimental arm uses QP-SID and joint VRM. At seed 42, NDCG@10 rises from 0.02629 to 0.05672, a relative gain of +115.73%.
Stable metrics: metrics/movielens-100k-seed42.json. This verifies that the mechanism runs on public substitute data; it is not Taobao's private queries, commercial labels, or H20 beam-search serving stack.