Everything this work has produced

PSF-Med measures whether medical vision-language models give the same clinical answer when a question is reworded without changing its meaning. The benchmark pairs original chest-radiograph questions with meaning-preserving paraphrases that pass a rubric-based equivalence audit, and a clinician-adjudicated sample validates the automated equivalence judge. Reworded but clinically identical questions flip a model's yes/no answer at rates spanning 6.4% to 54.7% across six models and three datasets, showing that paraphrase sensitivity is a measurable and model-dependent reliability gap.

Accepted at the MMFM-BIOMED workshop at CVPR 2026.

This scoping review maps how the literature evaluates the trustworthiness of medical vision-language models across four axes: robustness, visual grounding, hallucination, and uncertainty. Following PRISMA-ScR, it charts which evaluation methods and datasets are used and where coverage is thin, to frame the paraphrase-sensitivity gap this dissertation addresses.

Submitted to JMIR AI on 25 May 2026, in open peer review from 29 May to 24 July 2026. JMIR posts the preprint publicly under the authors’ names. It is not peer reviewed or edited, and it should not guide clinical practice.

VSF-Med is an early pilot that scores the vulnerability of medical vision-language models, establishing the initial paraphrase-vulnerability and attention-stability setup. It is the precursor that matured into the PSF-Med benchmark at the center of this dissertation.

Admitted as an early pilot abstract, IEEE ISBI 2026; not a peer-reviewed full paper. The precursor that matured into PSF-Med.

This companion study examines prompt sensitivity in text-only medical language models. It shows that chain-of-thought prompting can backfire, degrading answers on some clinical prompts where it was expected to help, and connects prompt-induced instability in language-only models to the paraphrase-sensitivity theme of the main work.

Accepted at the 2AI Conference 2026. Companion study on text-only language models; it sits outside the dissertation's five thrusts.

This work asks whether the layers implicated in paraphrase-driven answer changes are the right place to intervene. A low-rank adapter trained on layers 15-19 (0.10% of parameters) cuts the pairwise flip rate by about 59% on a clean patient-disjoint evaluation with no observed accuracy reduction. The adapter achieves this consistency by relying more heavily on question text than on image evidence, so lower flip rates do not imply better visual grounding.

Accepted at CHIL 2026; camera-ready complete.

This work tests whether attention maps from medical vision-language models actually mark the image regions the model uses to answer. Using bounding-box coverage, patch-rank agreement with occlusion, and region-based causal scoring, it finds that attention only marginally beats a shifted box and does not agree with causal patch importance, so attention behaves as a coarse localizer, short of a faithful explanation, and should not be treated as safety evidence.

Accepted at the iMIMIC 2026 workshop (Interpretability of Machine Intelligence in Medical Image Computing) at MICCAI 2026.

This paper argues that a model giving the same answer to reworded questions is not the same as a model that reads the image. A per-sample four-quadrant audit crosses paraphrase consistency against image dependence and shows that a large share of each model's consistent predictions are unchanged when the image is swapped or removed. Because correctness is a separate axis, a text-reliant but consistent prediction can even be more accurate than a grounded one, so a behavioral consistency screen is not a per-prediction safety verdict.

Accepted at the CVPR 2026 MedReasoner workshop.

This work studies whether a single predictive-entropy signal can flag both likely errors and likely paraphrase flips in medical vision-language models, so that a deployment audit can screen the same uncertainty for two failure modes at once. It reports how well entropy ranks flip-prone and error-prone cases across models and datasets, and finds that softmax entropy, temperature-scaled entropy, and margin are rank-equivalent while a deep ensemble is distinct but fails out of distribution.

Accepted as a poster at UNSURE 2026 (Uncertainty for Safe Utilization of Machine Learning in Medical Imaging), 8th edition, a satellite workshop of MICCAI 2026; archival Springer LNCS.

Selective prediction is the standard safety layer for medical vision-language models: abstain on uncertain cases and admit only confident ones. This work shows that such gates can certify a text-driven subset as safe. Across three model families and two chest X-ray datasets, gates that raise admitted-case accuracy select the cases where the image barely contributes, while accuracy on the slice where the image must change the answer collapses. The pattern survives frozen thresholds, prevalence matching, and a standardized swap protocol, and it motivates grounding-constrained gates that admit a prediction only when it is both reliable and demonstrably image-dependent.

Manuscript in revision.