James Song

I am an undergraduate student at the University of Michigan majoring in Computer Science and Mathematics. I currently work as a research assistant within the ECE Department, where I am fortunate to be advised by Prof. Liyue Shen. Concurrently, I’ve also worked at the InfiniAI Lab at Carnegie Mellon University with Prof. Beidi Chen and many wonderful collaborators. Before that, I also had a wonderful time working with Prof. Jingwen Hu at the UMTRI.

My research interests mainly lie in generative models and agents:

• Multimodal Generative Models: designing architectures that jointly process and generate across modalities such as images, tabular data, video, and text. For example, I designed a multimodal latent diffusion model that fuses medical imaging with clinical text embeddings, and structured the latent space via representation learning to encode clinically meaningful attributes.
• Spatiotemporal Modeling and World Models: learning predictive models of how environments evolve over time, and using them to inform decision-making. For instance, I have worked on longitudinal progression in lung LDCT scans for clinical prediction.
• LLM Agents and Security: evaluating and improving the robustness of LLM-based code agents in sandboxed environments.

Currently, I am exploring action-conditioned world models that simulate individualized disease progression, with the goal of learning optimal treatment policies.