| Back to Index | Full Publications | Research Statement |
Our lab at the University of Science and Technology of China develops AI-centric networked systems. We study how model behavior, application workflows, and runtime conditions can guide the movement and placement of computation, memory, and communication.
Our current focus is on model- and workflow-aware systems for efficient, adaptive, and accessible LLM execution. This agenda grows from our earlier work on datacenter networking: moving from sensing packets, queues, and paths to sensing model and workflow semantics.
Adaptive Distributed Learning
We investigate efficient LLM training and post-training across slow, heterogeneous, and geographically distributed resources. Topics include low-communication training, asynchronous reinforcement learning, on-policy learning, and optimizer-runtime co-design.
Stateful and Agentic Inference
We build serving systems that manage evolving model state and complete agent workflows rather than isolated requests. Our interests include KV-cache migration and offloading, workflow-aware routing, prompt and state reuse, and model-guided runtime decisions.
Heterogeneous Runtime Foundations
We explore the compiler, memory, and communication foundations needed for efficient AI execution on both datacenter and commodity hardware. This includes LLM compilers, kernel-aware memory management, and high-speed interconnects for consumer systems.
Temporal sparsity for KV-cache migration, offloading, and memory-tier placement
Cross-region and low-communication LLM training
Efficient systems for reinforcement learning and on-policy distillation
Workflow-aware routing and scheduling for LLM agents
Prompt and KV-cache reuse in tool-calling workflows
Model-internal signals for resource management and predictive scheduling
LLM compilers, memory allocators, and inference on consumer hardware
We are a growing group of doctoral, master's, and undergraduate researchers. Because many projects are in an early exploratory stage, we currently present our shared research themes rather than individual student assignments.
We welcome motivated Ph.D., M.Sc., and undergraduate students interested in computer systems, networking, and machine learning. If you would like to work with us, please email snowzjx _at_ ustc _dot_ edu _dot_ cn with a brief introduction, your CV, and the research questions that interest you.
