My research focuses on making large language model systems more trustworthy, with an emphasis on retrieval-augmented generation (RAG). I study how to test these systems for hallucination, how to evaluate the benchmarks used to measure them, and how to apply this work in high-stakes settings such as healthcare.
Research Interests
Trustworthy Retrieval-Augmented Generation
Methods for detecting and measuring hallucination in RAG systems, and for judging whether a system’s answers are faithful to the evidence it retrieves.
Benchmark Evaluation and Auditing
Examining the quality, coverage and realism of the benchmarks used to evaluate language models, so that reported results reflect how systems behave in practice.
Related: A Graph-Based Audit Framework for Healthcare Prompt-Response Benchmarks (HealthCom 2026)
Language Models for Healthcare and Public Health
Applying language models, prompt engineering and structured representations to health-related text, with attention to safety and reliability.
Related: Structured Insight Generation from Mental Health Counseling Reviews (AAAI’25 Workshop)
Knowledge Graphs and Structured Representations
Using graph-based structure to organize, compare and audit prompts, questions and system behavior.
Earlier work: Design and Implementation of a Strong Representation System for Network Policies (ICCCN’22)
Current Directions
I am currently working on evaluation methods that make RAG benchmarks more representative of real user questions, particularly in medical settings, and on ways to keep such benchmarks efficient to run without losing the signal they provide.
Collaboration
I am happy to discuss collaborations or questions about this work. See my publications for papers and links.