I lead a machine learning group, where we want to develop simple and robust machine learning methods that can improve generalization. My current research interests are:
How do we build robust machine learning models from imperfect supervision? What new possibilities do foundation models unlock? [C16, C18, C21, C24]
How do we adapt foundation models to downstream tasks while avoiding catastrophic forgetting? Can we improve multi-task performance by merging single-task models? [C28, C29]
How can we develop reasoning LLMs and agents without relying on large-scale annotations? Can reinforcement learning enables models to solve complex tasks through self-improvement?
I am also interested in diffusion models and unified multimodal understanding and generation models.
