Optimizing the Knowledge Lifecycle for LLM Reasoning: From Reliable Construction to Cost-effective Reasoning
The Hong Kong University of Science and Technology (Guangzhou)
数据科学与分析学域
PhD Thesis Examination
By Ms. Xinyi ZHU
摘要
Large Language Models (LLMs) increasingly support various applications, including question answering, evidence synthesis, and decision support. However, their parametric knowledge is bounded by pretraining and often lacks the domain-specific, up-to-date, or private information in domains such as finance, law, and medicine. Knowledge-Driven LLM Reasoning addresses this limitation by incorporating external knowledge through three stages: knowledge construction, knowledge retrieval, and knowledge-grounded reasoning. Each stage poses a distinct challenge. Construction must turn noisy, weakly labeled data into reliable knowledge without costly manual annotation. Retrieval must locate the correct entities and evidence in large collections despite incomplete and noisy inputs. Grounded reasoning must cover facts scattered across long documents without repeatedly processing large contexts. Together, addressing these challenges contribute to the reliability, accuracy, and scalability of downstream LLM applications.
This thesis develops three complementary frameworks along this lifecycle. Triple-D transforms noisy distant-supervision data into reliable training data through context-preserving preparation and scalable denoising, supporting relation extraction and Knowledge Base (KB) population. OpenMEL then links multimodal mentions to entities in the KB without extensive labeled data by adaptive noise filtering, incomplete multimodal evidence enrichment and entity global coherence construction. RAG-Sketch finally gathers dispersed evidence from long document collections and coordinates the reuse of shared evidence across queries, enabling more accurate reasoning with minimum context processing. Together, the three frameworks form a continuous path from raw data to reliable knowledge, from knowledge to relevant evidence, and from evidence to grounded answers.
Extensive experiments on real-world datasets and public benchmarks demonstrate the effectiveness and efficiency of the proposed approaches in their corresponding stages. Overall, this thesis presents a coherent set of techniques for improving the quality, accessibility, and cost-effective use of external knowledge throughout the LLM reasoning lifecycle.
TEC
Chairperson: Prof Xun WU
Prime Supervisor: Prof Lei CHEN
Co-Supervisor: Prof Ke YI
Examiners:
Prof Jing TANG
Prof Yongqi ZHANG
Prof Guang ZHANG
Prof Haibo HU
Note
Audience may attend only on-site.
日期
24 November 2026
时间
10:00:00 - 12:00:00
地点
E3-401, HKUST(GZ)
主办方
Data Sciencen and Analytics Thrust
联系邮箱
dsarpg@hkust-gz.edu.cn