Research
Today's foundation models are remarkably capable, but the cost and complexity of training, adapting, and serving them continue to grow. I am interested in closing this gap through algorithm–system co-design: designing model architectures, post-training objectives, compression methods, and systems that make frontier models more capable, efficient, and practical at scale.
Across the four connected areas below, I work from model design and learning objectives down to kernels, runtimes, and serving infrastructure, then apply these ideas to coding, tool-using, and scientific agents. I value open, reproducible research that improves both model quality and real-world usability. For paper details and authors, see Selected Publications or my Google Scholar profile.
I am interested in architecture changes that make generation cheaper without compromising model capability. The central question is how attention, residual pathways, routing, and speculators can expose more parallelism, reduce communication and KV-cache overhead, and make reasoning faster.
I study how to improve models after pretraining while controlling memory and compute. This includes stable reinforcement-learning objectives, loss and scaling-law analysis, and low-bit weight, activation, and KV-cache quantization that preserves accuracy under real serving constraints.
I build end-to-end training and serving stacks that turn efficient algorithms into practical systems. I am especially interested in asynchronous RL, training–serving loops, KV infrastructure, scheduling, and multi-tenant isolation across kernels, runtimes, and clusters.
I explore how efficient ML infrastructure can support capable coding, tool-using, and scientific agents. The goal is to improve trajectory quality, adaptation, coordination, and reproducibility while lowering the cost of training and real-world deployment.
Feel free to reach out — zhongzhu.zhou@sydney.edu.au — if you have aligned interests in efficient ML systems, LLM training/serving infrastructure, quantization, or coding-agent research. For a complete role-by-role breakdown of contributions (motivation + specific contributions), see the Experience page.