Binesh Sadanandan PhD Dissertation Companion
Read thesis

Papers and resources

Where this work has been published

5 of these 9 are public. Work still under review is named here with its status, but the manuscript is not posted: several are anonymous submissions, and posting them would break review. The findings themselves are all in the dissertation.

AcceptedPublishedPreprintUnder review

Thrust 1: Measurement

How often do equivalent questions change model answers? Open the thrust

Accepted MMFM-BIOMED Workshop, CVPR 2026 (non-archival), 2026

PSF-Med: A Clinician-Audited Benchmark for Paraphrase Sensitivity in Medical Vision-Language Models

Binesh Sadanandan, Vahid Behzadan, Lekshmy Jayan, Arun Gopinatha Kurup

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 (non-archival).

Under review Under review, 2026

Trustworthiness Evaluation of Medical Vision-Language Models: A Scoping Review of Robustness, Grounding, Hallucination, and Uncertainty

Binesh Sadanandan, Amirhossein Karimi, Bibek Upadhayay, Vahid Behzadan

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.

Manuscript under review.

Manuscript not postedThis work is under review, so the manuscript is not distributed here. The findings it reports are in Chapter 2 of the dissertation.

Thrust 3: Mechanisms and Mitigation

Can paraphrase flips be reduced without degrading observed accuracy? Open the thrust

Accepted Conference on Health, Inference, and Learning (CHIL) 2026, 2026

Mechanistically Guided LoRA Improves Paraphrase Consistency in Medical Vision-Language Models

Binesh Sadanandan, Vahid Behzadan

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. A key caveat is that the adapter buys this consistency by leaning harder on question text rather than image evidence, so lower flip rates do not imply better visual grounding.

Accepted at CHIL 2026; camera-ready complete.

Thrust 4: Safety Evaluation

Do consistency improvements produce safer, more image-dependent behavior? Open the thrust

Accepted MedReasoner Workshop, CVPR 2026, 2026

Consistent but Dangerous: Per-Sample Safety Classification Reveals False Reliability in Medical Vision-Language Models

Binesh Sadanandan, Vahid Behzadan

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.

Under review Under review, 2026

Attention Without Grounding: Causal Evaluation of Visual Explanations in Medical VLMs

Binesh Sadanandan, Vahid Behzadan

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 rather than a faithful explanation and should not be treated as safety evidence.

Manuscript under review.

Manuscript not postedThis work is under review, so the manuscript is not distributed here. The findings it reports are in Chapter 6 of the dissertation.

Thrust 5: Deployment

Can uncertainty and audit signals identify predictions that should be escalated? Open the thrust

Under review Under review, 2026

Predictive Entropy as a Joint Screen for Error and Paraphrase Instability in Medical Vision-Language Models

Binesh Sadanandan, Vahid Behzadan

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.

Manuscript under review.

Manuscript not postedThis work is under review, so the manuscript is not distributed here. The findings it reports are in Chapter 7 of the dissertation.
Under review Under review, 2026

Consistency Is Not Safety: Deployment Audits Reveal Text-Driven Selection in Medical Vision-Language Models

Binesh Sadanandan, Vahid Behzadan

This work frames an offline readiness audit that admits a prediction only when a model agrees across paraphrases and depends on the image. Applied across models and datasets, the audit shows that consistency-based gates preferentially admit text-answerable cases rather than image-grounded ones, so apparent reliability from low flip rates can reflect text-driven selection rather than sound visual reasoning.

Manuscript under review.

Manuscript not postedThis work is under review, so the manuscript is not distributed here. The findings it reports are in Chapter 7 of the dissertation.

Companion and precursor work

Related work that sits outside the five thrusts: the pilot that became PSF-Med, and a companion study on text-only medical language models.

Accepted 2AI Conference 2026, 2026

When Chain-of-Thought Backfires: Evaluating Prompt Sensitivity in Medical Language Models

Binesh Sadanandan, Vahid Behzadan

This companion study, which sits outside the dissertation's five thrusts, examines prompt sensitivity in text-only medical language models. It shows that chain-of-thought prompting can backfire, degrading rather than improving answers on some clinical prompts, 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.

Accepted IEEE ISBI 2026 (pilot abstract), 2025

VSF-Med: A Vulnerability Scoring Framework for Medical Vision-Language Models

Binesh Sadanandan, Vahid Behzadan

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.

Early pilot abstract, IEEE ISBI 2026; the precursor that matured into PSF-Med.

Code, data, and models

The benchmark and the adapter weights are public. The image corpora are not mine to redistribute, so they stay with their original providers.

Public

PSF-Med benchmark code

Benchmark construction and evaluation code for PSF-Med: paraphrase generation, equivalence auditing, and flip-rate evaluation (Thrust 1).

Open

Public

PSF-Med dataset (Hugging Face)

The released PSF-Med paraphrase benchmark data on Hugging Face; image files are not redistributed and must be obtained from the original credentialed sources.

Open

Public

Mechanistically-guided LoRA weights (Hugging Face collection)

Targeted and full LoRA adapter checkpoints for MedGemma-4B from the CHIL 2026 paper, trained at n=12K.

Open

Public

Paraphrase-consistency mitigation code

Training and evaluation code for the targeted LoRA mitigation and its safety re-audit (Thrust 3).

Open

Public

VSF-Med project site

Project page for VSF-Med, the vulnerability scoring framework companion work on medical VLM safety.

Open

Access on request

Weights & Biases experiment tracking

Experiment tracking workspace for the uncertainty-quantification and robustness runs.

Open

Private

medical-vlm-robustness research repository

The full research monorepo behind the dissertation (pipelines, analyses, dissertation sources); private during the pre-defense period.

github.com/UNHSAILLab/medical-vlm-robustness

Dataset access, in short

DatasetHow to get itLicence
MIMIC-CXRCredentialed PhysioNet access (CITI training plus a signed data use agreement); redistribution of images or reports is prohibited, so the benchmark releases only question/paraphrase text and identifiers.PhysioNet Credentialed Health Data License 1.5.0
PadChestRegistration with the Medical Imaging Databank of the Valencia Region (BIMCV) is required before download. The PadChest Dataset Research Use Agreement grants research use only and prohibits reproducing or publishing any portion of the dataset without prior written permission, so the benchmark releases only question and paraphrase text.PadChest Dataset Research Use Agreement (BIMCV; research use only, redistribution prohibited without written permission)
VinDr-CXRCredentialed PhysioNet access (CITI training plus a signed data use agreement); redistribution prohibited.PhysioNet Credentialed Health Data License 1.5.0

PSF-Med distributes questions, paraphrases, judge labels, and audit verdicts. It does not distribute images.

Cite

Permanent link

Type to search. Press Escape to close.