Binesh K Sadanandan

Binesh K Sadanandan

Senior Principal R&D Engineer (Technical Fellow) at Medtronic | Ph.D. dissertation successfully defended at the University of New Haven

Research

I conducted 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. I successfully defended the dissertation on August 26, 2026.

My dissertation, Paraphrase Sensitivity in Medical Vision-Language Models: Measurement, Mechanisms, Mitigation, and Deployment Safety, investigates robustness, safety, and interpretability failures in medical Vision-Language Models (VLMs). Its central finding is that a low paraphrase-flip rate is not sufficient evidence of reliable visual reasoning: models that appear reliable by standard metrics may be exploiting text shortcuts rather than analyzing the medical image.

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 PSF-Med failure gallery Read the final dissertation

Key contributions

Every number below is traceable to a chapter, a sample size, and a source file in the dissertation companion.

ContributionWhat I found
PSF-MedA benchmark of 92,856 final evaluation 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 safetyAveraged 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 diagnosisSparse 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 and causal probeA 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. In a controlled replica, broader training-phrasing coverage reduces held-out wording sensitivity, identifying training coverage as a causal lever.
Deployment auditsFor Targeted LoRA on the PadChest flip bank, single-pass predictive entropy ranks paraphrase flips at AUROC 0.823 and errors at 0.862. The gate remains an offline readiness audit: it admits text-answerable cases over grounded ones, and no single internal monitor transfers across model families.

Publications

Browse the full publication list for papers, preprints, and patents. Selected dissertation publications follow.

  • Trustworthiness Evaluation of Medical Vision-Language Models: A Scoping Review of Robustness, Grounding, Hallucination, and Uncertainty
    B. Sadanandan, A. Karimi, B. Upadhayay, V. Behzadan. JMIR AI preprint; manuscript under review, 2026.
    JMIR preprint

  • PSF-Med: A Clinician-Audited Benchmark for Paraphrase Sensitivity in Medical Vision-Language Models
    B. Sadanandan, V. Behzadan, L. Jayan, A. G. Kurup. MMFM-BIOMED at CVPR, 2026.
    arXiv:2602.21428

  • Mechanistically Guided LoRA Improves Paraphrase Consistency in Medical Vision-Language Models
    B. Sadanandan, V. Behzadan. CHIL, 2026.
    arXiv:2603.00148

  • Consistent but Dangerous: Per-Sample Safety Classification Reveals False Reliability in Medical Vision-Language Models
    B. Sadanandan, V. Behzadan. MedReasoner at CVPR, 2026.
    arXiv:2603.20985

  • Predictive Entropy as a Joint Screen for Error and Paraphrase Instability in Medical Vision-Language Models
    B. Sadanandan, V. Behzadan. UNSURE at MICCAI, 2026. Poster.
    Paper

  • Attention Without Grounding: Causal Evaluation of Visual Explanations in Medical VLMs
    B. Sadanandan, V. Behzadan. iMIMIC at MICCAI, 2026.
    arXiv:2607.18577

Datasets & Code

See the project portfolio for applied work and the dissertation companion for the thesis and reproducibility materials.

  • PSF-Med Benchmark: 92,856 final evaluation 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.
  • PSF-Med Dataset: public questions, paraphrases, and audit verdicts.
  • Targeted LoRA code and models: training, evaluation, and adapter resources for the mitigation study.

News

  • September 2026: Predictive Entropy accepted as a poster at UNSURE at MICCAI 2026.
  • August 2026: Successfully defended the Ph.D. dissertation on August 26; final dissertation published on August 28.
  • August 2026: Attention Without Grounding accepted at iMIMIC at MICCAI 2026.
  • June 2026: PSF-Med accepted at MMFM-BIOMED at 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 accepted at MedReasoner at CVPR 2026.
  • Feb 2026: VSF-Med poster accepted at IEEE ISBI 2026.