Recently, the International Conference on Machine Learning (ICML) was held in Vancouver, Canada. A paper presenting the research results (Performance-Oriented Query Decomposer, POQD) by the teaching team of the "Big Data and Machine Intelligence" course at the Engineering Practice Education Center of Tsinghua University (hereinafter referred to as the "Center") has been accepted as a poster paper for presentation at ICML (https://icml.cc/).
The teaching team of the course consists of faculty members from the Center, the National Research Center for Information Science and Technology at Tsinghua University, and the School of Computer Science at Peking University. This research was jointly conducted by Chen Zhen, Senior Engineer and leader of the course team, in collaboration with the School of Computer Science at Peking University.
Building on Retrieval-Augmented Generation (RAG) technology—which enhances domain knowledge in current Large Language Models (LLMs)—this research proposes the Performance-Oriented Query Decomposer (POQD) , a performance-optimized query decomposition method designed for the Multi-Vector Retrieval (MVR) framework. Experimental results demonstrate that POQD outperforms existing query retrieval strategies in RAG-based question-answering systems.
LLM-Based Subquery Generation Framework (Comprising Two LLMs: a Prompt Optimizer and a Query Decomposer)
ICML Conference Poster Presentation
Since the course was launched in 2016, its teaching team has maintained a strong focus on cutting-edge developments in artificial intelligence, continuously refining the course content and guiding students in exploring frontier issues in the AI field. Since 2024, the "Big Data and Machine Intelligence" course has been offered as a general elective at Tsinghua University and has served as a pilot course for AI-empowered teaching reform. In line with the university's requirement that general electives be "free of disciplinary barriers while maintaining academic depth," the teaching team has been progressively restructuring the course content. The course is also designated as a Global Hybrid Course by the university's Academic Affairs Office, featuring bilingual instruction in Chinese and English. It is open to undergraduate students from various schools and departments at Tsinghua, as well as international undergraduate students enrolled in the Global Hybrid Course program.
ICML is a top-tier international conference in the field of artificial intelligence and machine learning, holding significant global influence. It is classified as a Grade A conference—the highest category—recommended by both the China Computer Federation (CCF) and Tsinghua University. ICML 2025 was held in Vancouver, Canada, from July 13 to July 19.
This research received strong support from the Undergraduate Teaching Reform Project of the Academic Affairs Office at Tsinghua University, the Experimental Innovation Fund Project of Tsinghua University, and the Center’s innovation and entrepreneurship programs such as iTeach. The Industrial Cloud Platform of the Center—a project under the National Demonstration Base for Mass Innovation and Entrepreneurship initiated by the National Development and Reform Commission—provided model computing power support, while the Artificial Intelligence Laboratory of the Center offered operational support for the computing power platform.
Paper Links:
https://icml.cc/virtual/2025/poster/44047
https://arxiv.org/abs/2505.19189
https://pku-sds-lab.github.io/POQD/
