Generative Retrieval for E-commerce with Semantic IDs and RL
A research paper proposes CQ-SID, a generative retrieval framework for e-commerce search that uses category-aware and query-item contrastive learning with Residual Quantized VAEs to encode products into hierarchical semantic cluster IDs, reducing beam search complexity. The method also introduces EG-GRPO, a reinforcement learning approach using expert-guided group relative policy optimization to align retrieval with downstream ranking goals. The framework is designed as a recall-stage supplement rather than a full end-to-end replacement, addressing challenges of massive dynamic catalogs and strict latency in industrial settings. The paper is available on arXiv under ID 2605.14434.
Key facts
- Paper ID: arXiv:2605.14434
- CQ-SID uses category-aware and query-item contrastive learning
- Residual Quantized VAEs encode items into hierarchical semantic cluster IDs
- EG-GRPO is an expert-guided reinforcement learning method
- Framework is a recall-stage supplement, not end-to-end replacement
- Addresses massive dynamic product catalogs and latency requirements
- Published on arXiv
- Announce type: cross
Entities
Institutions
- arXiv