CaLIR: Faster, smarter generative retrieval for e-commerce via latent intent reasoning

Generative retrieval for e-commerce maps user queries directly to product Semantic Identifiers (SIDs), but short, noisy, attribute-heavy queries create a gap between natural language intent and artificial SIDs. Explicit Chain-of-Thought (CoT) reasoning can bridge this gap but adds unacceptable latency for online systems. To solve this, researchers propose CaLIR (Category-guided Latent Intent Reasoning), which learns continuous latent intent states before SID decoding. Instead of generating text rationales, CaLIR uses product category hierarchies as a natural scaffold for coarse-to-fine intent reasoning, aligning latent states with category-level shopping intent. It also models diverse intent paths for multi-positive queries and incorporates a query-specific dynamic prefix trie assembled from pre-indexed category-level tries.

Experiments on multilingual e-commerce search datasets demonstrate that CaLIR achieves a better balance between retrieval effectiveness and inference efficiency than existing methods. The framework shows strong transferability and robustness across induced hierarchies and different generative backbones. By avoiding explicit CoT generation, CaLIR maintains low latency while still reasoning about user intent implicitly. This makes it particularly suited for real-time e-commerce environments where both speed and accuracy are critical. The paper is available on arXiv (2606.07075) and represents a practical step toward more intelligent and efficient product search.