Context-Aligned Reasoning for Medical Vision-Language Models [arXiv]
Apr 2026KnowledgeMR @ CVPR 2026 (poster) · Second author
- Co-authored a framework in which a frozen vision–language model must reconcile radiomic statistics, Grad-CAM activations, and vocabulary-grounded cues before returning a structured chest X-ray answer with evidence, uncertainty, and safety notes—reducing fluent but weakly grounded conclusions.
- Conducted cross-dataset evaluation on CheXpert; radiomic features achieved AUC 0.71 while text embeddings reached 0.50.
- Found that image-plus-text conditioning reduced average hallucinated keywords from 1.14 to 0.25; tool-augmented reasoning shortened evidence explanations from 19.4 to 15.3 words without increasing model confidence.
- Designed the methodology figure, quantitative tables, and conference poster.