Reproduce “LEAPS: An LLM-Empowered Adaptive Plugin in Taobao AI Search” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Lei Wang, Jinhang Wu, Zhibin Wang, Biye Li
Researchers from Alibaba's Taobao & Tmall Group developed LEAPS, an LLM-powered adaptive plugin for e-commerce search, designed to handle complex natural-language queries using a "Broaden-and-Refine" paradigm. The system improved F1@20 scores to 0.81 on an internal Taobao dataset and increased online Click-Through Rate by 27.1% and Clicks per Session by 43.9% in A/B tests on Taobao AI Search.
Key insights
- A multi-stage, system-aware training approach for query expansion, which includes using search system yield as a supervision signal and a diversity-aware RL reward, is crucial for maximizing item coverage and minimizing redundancy.
- Integrating heterogeneous data sources and Chain-of-Thought reasoning within the Relevance Verifier enables accurate and explainable relevance judgments for complex, multi-attribute queries, addressing nuances beyond surface-level metadata.
- A modular, plug-in based "Broaden-and-Refine" architecture effectively balances recall (via query expansion) and precision (via relevance verification) to resolve the zero-result/decision-overload dilemma without requiring a full overhaul of existing search infrastructure.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.