Wenqian Ye 叶文谦 (he/him)

I am a PhD candidate in Computer Science at the University of Virginia, advised by Prof. Aidong Zhang, and I also work closely with Prof. Henry Kautz. Previously, I received my M.S. in Computer Science at NYU Courant and my B.S. in Mathematics from UIUC, with minors in Computer Science and Electrical Engineering. My research has been recognized by the KDD 2025 Best Paper Award and the John A. Stankovic Graduate Research Award.

Research Interests

My north star is trustworthy AI that can safely learn recursively and drive new scientific discovery, from AGI to ASI. This begins with ML Foundation that keeps the trajectory both safe and interpretable, building models that generalize under distribution shift, represent their internal beliefs faithfully, and know how far their own predictions can be trusted. As models gain long-horizon reasoning, tool use, and autonomy, I work on Safety Alignment and scalable oversight so that their objectives stay grounded in human intent. My ultimate goal is AI for Scientific Discovery, where agents achieve recursive self-improvement through closed-loop discovery: they generate hypotheses, design and run experiments, learn from evidence, and refine their own reasoning and methods across iterations. My work also has real-world impact on Applications including healthcare, autonomous driving, and embodied AI.

• ML Foundation: out-of-distribution generalization, spurious correlations, Bayesian machine learning, and uncertainty estimation. Representative papers: NeuronTune, Evidential Alignment, SAGE.
• Safety Alignment: reward hacking, emergent misalignment in large language models, and agentic misalignment in autonomous agents. Representative papers: PRISM, AEA, MM-SpuBench.
• AI for Scientific Discovery: self-evolving and recursively self-improving agents, task emergence in reinforcement learning, and accelerating scientific discovery. Representative papers: DFE, EHRMamba, MpoxVLM.
• Applications: embodied AI and autonomous driving, and AI for health (medical vision-language models, foundation models for electronic health records). Representative papers: LaMPilot, MAPLM, MaRs-VQA.

For a full list of papers, see the Publications page.

News