Alibaba's new generative retrieval boosts Tmall search GMV 1.15%
New system accounts for 72% of purchases with semantic IDs and expert-guided RL.
A new retrieval framework called CQ-SID and EG-GRPO positions generative retrieval as a recall-stage supplement rather than an end-to-end replacement. Using category-and-query constrained semantic IDs via Residual Quantized VAEs and expert-guided reinforcement learning (EG-GRPO), it achieves up to 26.76% relative gains in semantic and personalized click hitrate over RQ-VAE baselines while halving beam search size. In production, it accounts for over 72.63% of purchases, with GMV up 1.15% and UCTCVR up 0.40%. The work demonstrates a viable path for deploying generative retrieval in real-world e-commerce.
- CQ-SID reduces beam search complexity by 50% using hierarchical semantic cluster IDs.
- EG-GRPO aligns generative recall with downstream ranking, achieving GMV +1.15% and UCTCVR +0.40% in A/B tests.
- Generative recall now drives 72.63% of Tmall purchases, proving scalable for production.
Why It Matters
As e-commerce scales, this shows AI can unify search and ranking, driving measurable revenue gains with lower latency.