DocDB: A Database for Unstructured Document Analysis
PVLDB 2025Demo
Query unstructured documents with SQL-like operations. Two-level retrieval and document-specific execution plans reduce the amount of text and the number of calls sent to LLMs.
Incoming Ph.D. student · Tsinghua University
I am an incoming direct-entry Ph.D. student in the Department of Computer Science and Technology at Tsinghua University (expected start: 2027). My advisor is Prof. Jiannan Wang. My undergraduate major is Computer Science and Technology (Elite Class) at Beijing Institute of Technology.
My research interests lie at the intersection of data systems and large language models, with related work in natural language processing and computer vision. I enjoy building efficient, practical AI systems.
我是李泽群,已获清华大学计算机科学与技术系直博录取,预计于 2027 年入学,导师为王健楠老师。本科就读于北京理工大学计算机科学与技术拔尖班。我的研究兴趣主要包括数据系统与大语言模型。目前,我正在探索异构架构上的高效多模态数据处理,重点关注 Apple Silicon 的统一内存架构、CPU–GPU 协同、资源竞争与自适应执行策略。
I am currently exploring efficient execution of multimodal data processing pipelines on heterogeneous architectures, with a particular focus on Apple Silicon's unified memory architecture. My research investigates CPU-GPU collaboration, resource contention, and adaptive execution strategies to improve end-to-end throughput.
I have been admitted to the direct Ph.D. program at Tsinghua University (expected start: 2027). My advisor is Prof. Jiannan Wang.
Our CCL 2025 paper was selected for an oral presentation. Our system placed 1st nationwide in the Chinese Factivity Inference evaluation.
DocDB was published in PVLDB 2025 and invited for a special demonstration at the 42nd National Database Conference.
My undergraduate project on efficient instruction-tuning data extraction received support from the Beijing Natural Science Foundation.
PVLDB 2025Demo
Query unstructured documents with SQL-like operations. Two-level retrieval and document-specific execution plans reduce the amount of text and the number of calls sent to LLMs.
CCL 2025Oral
Inferring whether an event is factual from Chinese text. Macro-pattern prompting and efficient Qwen3-32B fine-tuning reach 94.01% and 92.61% accuracy in their respective tracks.
Extract high-value QA pairs for LLM instruction tuning without scanning every document. Contrastive representations, adaptive sampling, and optimal-transport feedback guide where to extract next.
Beijing Natural Science Foundation · CNY 50,000 research grant
SIGMOD 2025Demo
An interactive system that explains and summarizes outliers through compact, human-readable rules.
Contributed visual-analysis modules for font and font-size recognition, icon extraction, and container detection, reconstructing static diagrams as editable DrawIO elements.
Ph.D. in Computer Science and Technology (direct-entry)
Department of Computer Science and Technology
Advisor: Prof. Jiannan Wang
Expected start
Bachelor's degree in Computer Science and Technology (Elite Class)
–Expected graduation