Spectral diagnostics for modality imbalance in medical VLMs

Arxiv CS.CV (computer vision)

Paper introduces a spectral diagnostic tool to detect modality imbalance in medical vision-language models. Unlike symmetric alignment metrics, it pinpoints which modality (image or text) is underperforming. Applied to clinical benchmarks, it reveals common over-reliance on text.

vision-language-modelscomputer-vision

Arxiv CS.CV (computer vision)·Alessandro GambettiQiwei HanCl\'audia SoaresHong Shen

7 days ago