Xuesong Ye worked on the inclusionAI/AReaL repository, focusing on optimizing the training loop for deep learning workflows. By refining the onload and offload sequences of model parameters between GPU and CPU, Xuesong reduced unnecessary memory transitions that previously caused latency and instability during large-scale training. The approach maintained parameter residency on the GPU through critical stages such as compute_values, ppo_update, and checkpointing, minimizing data transfers and improving throughput. This optimization was validated on a 4×H100 setup using Python, PyTorch, and advanced performance tuning techniques, demonstrating a strong understanding of deep learning system bottlenecks and scalable engineering practices.