Yijun Yang

Refusal policies in language models can inadvertently compromise their ability to handle benign requests, revealing a complex interplay that could redefine safety evaluation metrics.

Achieve state-of-the-art ultrasound video segmentation with only a single point click and anatomical category name, surpassing even finetuned specialists, thanks to a novel training-free framework.

Bidirectional interaction between enhanced understanding, controllable spatial editing, and novel-view-assisted reasoning enables a unified multimodal model to achieve spatial intelligence beyond general visual competence.

Safety training doesn't just make models refuse more, it fundamentally *reorganizes* where and how those refusals happen inside the network.