Abstract

The rapid evolution of foundation models is increasingly bottlenecked by a widening gap between algorithmic demands and system efficiency. As model scale and context lengths explode, infrastructure efficiency plateaus. Addressing these challenges requires a full-stack rethinking of machine learning systems. In this talk, I will present a research framework centered on algorithm–system co-design to push the efficiency frontier across the ML lifecycle. I first demonstrate how system-level support for algorithm advancement can drastically reduce the cost of large-scale hyperparameter exploration. Next, I tackle the long-tailed execution bottlenecks in post-training reinforcement learning, introducing a co-designed system approach that delivers substantial efficiency gains while preserving on-policy RL training. I also introduce system designs that enable vision–language models to scale to million-token contexts by resolving fundamental memory and communication constraints. I conclude by discussing the future of system support for agentic models at scale.

About the speaker

Qinghao Hu is a Postdoctoral Associate at the Massachusetts Institute of Technology, advised by Professor Song Han. His research focuses on efficient machine learning systems, spanning datacenter scheduling, distributed training, reinforcement learning, and model serving. His work has been recognized with the ASPLOS Distinguished Paper Award and the WAIC Best Paper Award. He is a recipient of the Google Ph.D. Fellowship, the ML and Systems Rising Stars Award, and the Best Ph.D. Thesis Award. His research has been featured in MIT News, NTU News, and the USENIX ;login: newsletter, and his open-source projects have attracted more than 6,000 GitHub stars. He received his Ph.D. from Nanyang Technological University and was previously a visiting scholar at ETH Zürich.