CQ-SID¶

Core mechanism: category-constrained semantic IDs, residual quantization, multi-signal rewards and EG-GRPO updates.

Paper information¶

Field	Details
Paper	arXiv 2605.14434
Institution	Alibaba Taobao & Tmall Group.
First public release	2026-05-14 (arXiv v1)
Author code	No official author code linked by the paper as checked on 2026-08-09.
Local adapter or method key	`cq-sid`
Local reproduction code	`src/auto_research/reproductions/cq_sid/`

Paper method and evidence¶

Constrain the first semantic-ID level by product category and encode within-category variation with residual codes. Expert-guided RL combines clicks, purchases, semantics and constraints for generative retrieval.

flowchart LR
  X[Items and categories] --> Q[Category-constrained residual quantization]
  Q --> S[CQ-SID]
  H[Query and history] --> G[Generative policy]
  S --> G
  E[Expert rewards] --> R[EG-GRPO]
  G --> R

Original paper figure¶

Figure 1 shows category-constrained IDs and expert-guided training. From the original paper; copyright remains with the authors.

Core equation¶

\[ s_i=(c_i,q_i^{(1)},q_i^{(2)}),\qquad A_j=\frac{r_j-\operatorname{mean}(r_{1:G})}{\operatorname{std}(r_{1:G})+\epsilon}. \]

Reported results¶

Two-week TmallAPP Search A/B: GMV +1.15%, UCTCVR +0.40%. The paper also reports semantic click hit rate +26.76%.

Local reproduction¶

Against collaborative item-ID retrieval, CQ-SID with 240 group-relative updates gives +1.66% local NDCG@10. Structured metrics.

auto-research reproduce --paper cq-sid

EG-GRPO updates execute, but the 64-GPU curriculum, private Tmall query logs, and production beam service were not reproduced.