Research
I am conducting this dissertation research in the Secure and Assured Intelligent Learning Lab (SAIL Lab) at the University of New Haven under the supervision of Dr. Vahid Behzadan.
My dissertation investigates robustness, safety, and interpretability failures in medical Vision-Language Models (VLMs). The central finding is that models that appear reliable by standard metrics are often exploiting text shortcuts rather than analyzing the medical image. A model can achieve near-perfect consistency while ignoring the chest X-ray entirely.
Interactive failure gallery
I have also published a poster-friendly interactive gallery with representative failure cases from PSF-Med, my paraphrase sensitivity benchmark. It shows how semantically equivalent clinical questions can trigger contradictory answers on the same chest X-ray while suppressing raw image filenames and internal example IDs.
Open the PhD dissertation companion Open the failure gallery
Key contributions
Every number below is traceable to a chapter, a sample size, and a source file in the dissertation companion.
| Contribution | What I found |
|---|---|
| PSF-Med | A benchmark of 92,856 audited question-paraphrase pairs, built from 26,850 chest X-ray questions across three countries. On the binary yes/no subset, six medical VLMs flip on 6.4% to 54.7% of pairs. |
| Consistency is not safety | Averaged across ten model-dataset settings, 81% of each model’s consistent predictions are image-invariant: the answer does not change when I remove the image. A model can look reliable and still be reading only the question. |
| Mechanistic diagnosis | Sparse Autoencoders point to Feature 3818 at layer 17 as a clinical-query operator gate, and the answer commits at layer 16. This is a candidate account, not a proven circuit: the feature is the largest layer-17 delta in 37 of 76 operator-preserving flips, and ablating it alone restores the original answer in only 6 of them. |
| Targeted repair | A Low-Rank Adaptation (LoRA) on layers 15 to 19, touching 0.1% of parameters, cuts the pairwise flip rate by about 59% (8.5% to 3.5% over five seeds) on a patient-disjoint test, with no observed accuracy reduction. It buys that consistency by leaning harder on the question text, so it does not preserve visual grounding. |
| Deployment audits | Offline audit rules admit 33% of PadChest cases at 96.8% accuracy, but they admit the cases a text prior already answers. On the slice where the image is needed, the same model scores 2.9%. No single internal monitor transfers across model families. |
Publications
PSF-Med: Measuring and Explaining Paraphrase Sensitivity in Medical Vision-Language Models
B. Sadanandan, V. Behzadan. MMFM-BIOMED Workshop, CVPR, 2026.
arXiv:2602.21428Mechanistically Guided LoRA Improves Paraphrase Consistency in Medical Vision-Language Models
B. Sadanandan, V. Behzadan. CHIL, 2026.
arXiv:2603.00148Consistent but Dangerous: Per-Sample Safety Classification Reveals False Reliability in Medical VLMs
B. Sadanandan, V. Behzadan. CVPR MedReasoner Workshop, 2026.
arXiv:2603.20985VSF-Med: A Vulnerability Scoring Framework for Medical Vision-Language Models
B. Sadanandan, V. Behzadan. IEEE ISBI, 2026. Poster.
arXiv:2507.00052When Chain-of-Thought Backfires: Evaluating Prompt Sensitivity in Medical Language Models
B. Sadanandan, V. Behzadan. 2nd International Conference on Applied Artificial Intelligence (2AI), 2026.
arXiv:2603.25960Predictive Entropy Links Calibration and Paraphrase Sensitivity in Medical VLMs
B. Sadanandan, V. Behzadan. Under review, 2026.
arXiv:2604.08941Attention Without Grounding: Causal Evaluation of Visual Explanations in Medical VLMs
B. Sadanandan, V. Behzadan. iMIMIC Workshop, MICCAI, 2026.Consistency Is Not Safety: Family-Specific Deployment Audits for Medical VLMs
B. Sadanandan, V. Behzadan. Under review, 2026.
Datasets & Code
- PSF-Med Benchmark: 92,856 audited question-paraphrase pairs, six VLMs, and three chest X-ray datasets: MIMIC-CXR, PadChest, and VinDr-CXR. The release carries questions, paraphrases, and audit verdicts, not images.
- Models: Base, targeted LoRA, and full LoRA checkpoints on Hugging Face.
News
- 2026: Attention Without Grounding accepted at the iMIMIC Workshop, MICCAI 2026.
- 2026: PSF-Med accepted at the MMFM-BIOMED Workshop, CVPR 2026.
- Apr 2026: PSF-Med poster at SMLM, Yale.
- Apr 2026: Mechanistically Guided LoRA Improves Paraphrase Consistency accepted at CHIL 2026.
- Mar 2026: Chain-of-Thought paper accepted at 2AI 2026.
- Mar 2026: Consistent but Dangerous, CVPR 2026 MedReasoner paper accepted.
- Feb 2026: VSF-Med poster accepted at IEEE ISBI 2026.