Yekun Chai

My research centers on pre-training for foundation models — what to train on, at what scale, and how those early choices set the ceiling on reasoning, code, and agentic behavior later. I am particularly interested in the pre-training ↔ post-training interface, and in understanding it in a principled way rather than through trial and error.

I have contributed to ERNIE, ERNIE-Code, and StarCoder 2.