Cascaded Generative Framework for E-Commerce Recommendations
A new cascaded merchandising framework decomposes storefront construction into placement-level theme generation and constrained keyword generation. This approach addresses rigidity in traditional personalized storefronts, which rely on static themes, retrieval systems, and pointwise rankers. Teacher-student fine-tuning improves scalability under production constraints. Fine-tuned model ablations approach closed-weight LLM performance.
Key facts
- Framework decomposes storefront construction into two generative tasks: placement-level theme generation and constrained keyword generation.
- Traditional storefronts use static themes, retrieval systems, and pointwise rankers.
- Teacher-student fine-tuning improves scalability under latency and cost constraints.
- Fine-tuned model ablations approach closed-weight LLM performance.
- The framework is designed for large e-commerce marketplaces.
- It aims to improve personalization and semantic cohesion across the page.
- It supports dynamic objectives and merchandising requirements over time.
- The approach is introduced in arXiv:2605.11118.
Entities
Institutions
- arXiv