Reproduce “QP-OneModel: A Unified Generative LLM for Multi-Task Query Understanding in Xiaohongshu 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
Jianzhao Huang, Xiaorui Huang, Fei Zhao, Yunpeng Liu, Hui Zhang, Fangcheng Shi, Congfeng Li, Zechen Sun and 4 more
QP-OneModel, a unified generative Large Language Model, addresses multi-task query understanding in Xiaohongshu Search by reformulating diverse query processing tasks into a single sequence-to-sequence problem. It achieves a 7.35% overall score gain over traditional pipelines in offline evaluations and delivers improved online user engagement and retrieval relevance on a live social media platform.
Key insights
- Unifying diverse query processing tasks into a single generative framework effectively captures semantic synergy and mitigates error propagation, outperforming task-isolated models.
- Incorporating configurable business rules and dynamic contexts directly into the LLM's prompt significantly enhances operational efficiency and adaptability to evolving platform policies without requiring model retraining.
- A multi-reward Reinforcement Learning stage allows the LLM to internalize complex business logic, leading to a deeper understanding of rules beyond rote memorization from supervised fine-tuning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.