I’m interested in studying and improving the fundamentals of modern machine learning through data (often known as data-centric AI). On the model training side, I work on data mixing, synthetic data, data representations, and data labeling. On the inference side, I work on test-time algorithms to produce higher-quality model generations, such as ensembling and routing. Currently, I am thinking about how to develop and operationalize a more principled understanding of how models learn from data (what skills does data teach the model? Does it matter if the data is synthetic or real?)