Enhancing the verifiability of large language model based medical question answering systems

Wang, Xiao

Permalink

https://hdl.handle.net/2142/129236 Copy

Description

Title

Enhancing the verifiability of large language model based medical question answering systems

Author(s)

Wang, Xiao

Issue Date

2025-05-02

Director of Research (if dissertation) or Advisor (if thesis)

Zhang, Minjia

Department of Study

Electrical & Computer Eng

Discipline

Electrical & Computer Engr

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Large Language Models
• Retrieval-Augmented Generation
• Trustworthy AI
• Verifiability
• Citation Generation
• Biomedical Question Answering

Abstract

Large language models (LLMs) have demonstrated impressive capabilities across a wide range of natural language processing tasks. However, their tendency to generate fluent yet unverifiable statements poses a fundamental challenge for deployment in high-stakes domains such as medicine, law, and education. This thesis addresses the central question of \textit{verifiability}: how can LLMs produce outputs that are not only accurate but also supported by transparent, checkable evidence? Focusing on the concrete case of medical question answering (QA), this work investigates citation generation as a mechanism for enhancing verifiability. Rather than adhering to a fixed pipeline, the thesis follows an iterative, design-driven approach to evaluate how system-level decisions—including the use of parametric versus non-parametric knowledge, retrieval-augmented generation (RAG), and fine-grained attribution strategies—affect the alignment between model-generated content and external sources. Based on these insights, a two-pass citation framework is proposed. The approach first encourages in-context citation generation during answer formulation, followed by a post hoc retrieval and reranking stage that refines attribution at the statement level. This pipeline improves citation recall and precision while maintaining fluency and factual correctness. Additionally, a human annotation study reveals that recent general-purpose LLMs can serve as effective automatic judges of citation quality, often outperforming domain-specific NLI models in aligning with expert judgments. In summary, this thesis contributes both practical methods and conceptual frameworks for improving LLM verifiability in biomedical QA, with broader implications for developing trustworthy, evidence-supported AI in other knowledge-intensive fields.

Graduation Semester

2025-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/129236

Copyright and License Information

Copyright 2025 Xiao Wang

Owning Collections