About Me
Hi, welcome to my homepage. I am Chen Liang (梁辰), a Member of Technical Staff at Microsoft AI. Before Microsoft, I completed my Ph.D. at Georgia Tech, advised by Prof. Tuo Zhao.
My current work focuses on post-training SWE agent models for GitHub Copilot in VS Code and CLI, across Microsoft and OpenAI model families. My work spans long-horizon RL, on-policy distillation, and format fine-tuning, with an interest in research-production co-design: turning production model failures into stable recipes for data efficiency, training stability, length control, and production-aligned behavior. Previously, I worked on mid-training for VS Code code completion models and pre-training for Phi models.
More broadly, my research interests lie in efficient and generalizable LLM training, guided by a “less is more” principle: efficiency can improve model quality by exposing essential learning signals. I study this through data-efficient training, parameter-efficient adaptation, stable optimization, and shorter yet stronger inference-time trajectories.