Junjie Yang

Junjie Yang

Multimodal medical AI researcher

I am Junjie Yang, currently pursuing a master’s degree in Applied Machine Learning at the University of Maryland, College Park. I earned my bachelor’s degree in Computer Science and Engineering from South China University of Technology (SCUT), where I studied in the English-Taught Innovation Class. My academic work has included collaborations with Weiwei Lin, Changmiao Wang, and Yao Wan.

My research interests include computer vision, medical AI, 3D vision, and agentic reinforcement learning.

Selected Projects

TEF
TEF overview comparing direct prediction, voting, and evidence fusion

TEF: Count Evidence, Not Sentences

Tempered Evidence Fusion (TEF) is a training-free method for measuring value orientations in long text. It gives more weight to decisive sentence-level LLM judgments and discounts uncertain ones. The work also introduces MIND, a benchmark of 8,358 Chinese and English posts across six value dimensions.

Paper · Code

VERPO
VERPO method overview

VERPO: Verified Evidence Regularized Policy Optimization

VERPO turns evidence-conditioned guidance into reward-aligned token-level credit assignment for reinforcement learning. Its Fisher-aware controller and Fisher Evidence Contrast help stabilize optimization while retaining outcome-level rewards.

Paper · Code

MedGEN-Bench
MedGEN-Bench overview figure

MedGEN-Bench: A Contextually Entangled Benchmark

MedGEN-Bench is a benchmark for open-ended multimodal medical generation. The reported evaluation snapshot contains 6,422 expert-reviewed image–text pairs spanning six canonical imaging modalities, 15 clinical tasks, and 27 named subtasks, including visual question answering, image editing, and contextual multimodal generation.

Open project page · Paper · Dataset

Publications

  • Count Evidence, Not Sentences: Tempered Evidence Fusion of LLM Judgments for Long-Text Value Measurement, Yuhe Wu, Rui Qian, Guangyu Wang, Yuran Chen, Yuanchao Zhu, Junjie Yang, et al. arXiv · Code

  • VERPO: Verified Evidence Regularized Policy Optimization, Haijiang Li, Chengyu Lv, Yi Zhang, Rui Qian, Zhibing Zhang, Xiangqing Shen, Junjie Yang, et al. arXiv · Code

  • MedGEN-Bench: Contextually entangled benchmark for open-ended multimodal medical generation, Junjie Yang, Yuhao Yan, Rui Qian, et al. arXiv · Project page

News