Dissertation Documents

My PhD research on making medical vision-language models more robust and safe

← Main Index

Core Dissertation Documents

• Dissertation Proposal — Formal research proposal with questions and methods
• Introduction — Background, motivation, and coupled failure modes
• Research Thrusts — Four thrusts from measurement to deployment
• Timeline & Resources — When things will happen and what I need
• Publication Plan — Papers I plan to submit and where
• High-Level Overview — Accessible summary of the dissertation for broader audiences

🎯 What I’m Working On

My research focuses on a big problem with medical AI models: they change their answers when you ask the same question in different ways. This is dangerous in hospitals because doctors need reliable AI assistance.

Main Problems I’m Solving:

• Phrasing-Sensitive Failure - models give different answers to the same question asked differently
• Misleading Explanation Effect - explanation metrics look better for wrong answers than correct ones
• Safety Issues - current evaluation doesn’t catch these problems

📊 What I Will Deliver

1. Measurement tools to find these problems systematically
2. Better models that are more stable and reliable
3. Safety framework for using AI in hospitals safely
4. Open-source toolkit so others can use and improve my work
5. Research papers to share findings with the community

🗓️ When Things Happen

• Now - March 2026: Measure the problem and find causes
• March - August 2026: Build better models
• May - October 2026: Test safety in realistic scenarios
• Throughout: Write papers and share results

🎯 Target Models

• LLaVA-Rad - my main focus for testing improvements
• MedGemma - comparison model to see if findings generalize
• Both models work with chest X-rays and medical questions

Related Resources

• How I Measure Robustness — Technical details
• LLaVA-RAD Model — Primary model under study
• MedGemma Model — Comparison model
• Safety System — How to deploy safely
• Additional Notes & Archived Plans — Earlier terminology (FSF/EFG) and planning docs