Jun Li

Apparent SOTA gains in semi-supervised 3D medical image segmentation may be illusory, driven by confirmation bias and test-set overfitting, not genuine progress.

Frozen vision-language models can dramatically improve abnormality grounding in rare disease imaging by iteratively refining decisions through optimized instructions and visual perturbations.

Text-to-image diffusion models leak backdoor secrets under cross-attention scaling, revealing a surprising divergence between benign and poisoned inputs.