Hi, I’m Binesh Sadanandan. This is my blog where I write about AI, machine learning, and my PhD research on making medical AI safer and more reliable.
I’m a PhD student at the University of New Haven, working with Dr. Vahid Behzadan. My research focuses on the robustness of Vision-Language Models (VLMs) used in clinical diagnostics, specifically understanding why these models give contradictory answers when clinicians rephrase the same question.
You can find more about me at bineshkumar.me.
My Research
Medical AI tools are already deployed in clinical settings, reading chest X-rays and flagging abnormalities. But we’ve found a critical problem: these models are sensitive to how you phrase your question. Ask about lung “volumes” vs lung “capacity”, same clinical meaning, and the model can give completely different diagnoses on the same image.
What makes this worse is that the model’s attention maps stay almost identical (correlation > 0.99) across phrasings. It looks at the right place but still produces contradictory answers. And standard explanation metrics can’t tell the difference between correct and incorrect predictions.
We call these two failure modes:
- Phrasing-Sensitive Failure (PSF): Equivalent questions yield contradictory diagnoses
- Misleading Explanation Effect (MEE): Explanations appear faithful even when the answer is wrong
What We’re Doing About It
Our research has four parts:
Measuring the problem: We’re building the VSF-Med benchmark (16,000+ question-image pairs) with metrics like Paraphrase Flip Rate and Attention Stability Index to systematically quantify these failures.
Finding the root cause: Using mechanistic interpretability (activation patching, attention decomposition) to trace failures to specific model layers and attention heads. Early evidence points to the cross-modal fusion layers.
Fixing it: Developing parameter-efficient fine-tuning (LoRA on <1% of parameters) with consistency and representation losses that enforce stable answers across paraphrases.
Safe deployment: Building a clinical safety layer that combines confidence scores, paraphrase agreement, and faithfulness checks to know when the model should defer to a human radiologist.
Early Results
Our pilot (1,600 samples) on MedGemma-4B and LLaVA-Rad shows flip rates up to 17% across paraphrase pairs, with 100% of flips occurring while attention remains stable, confirming the problem lives in language processing, not vision.