Some Random Stuff 👋

I am Mingyuan Wu, a part-time researcher at Meta. I am currently a final-year PhD candidate in Computer Science at University of Illinois, Urbana Champaign. I am fortunate to be advised by Prof. Klara Nahrstedt, and and to work with Prof. Minjia Zhang and Prof. Chengxiang Zhai. Before starting my PhD, I spent wonderful time at UIUC and Shanghai Jiao Tong University.

I work on vision language model agents and multimodal reasoning. In 2025, I have been cooking VLMs and LLMs with recipes of multi-turn reinforcement learning fine-tuning and inference-time scaling, to boost self-improvement, reasoning, and memory use. I also investigate VLM interpretability through circuit tracing.

My research goal is to build human-level agents that partner with people: humans set the intent with prompts; agents reason, execute and suggest new possibilities (via recommendations).

When I’m not teaching large models, I prototype augmented-reality systems for fun, hoping that, one day, agentic models will power these interactive, human-centered interfaces.