Ziwei Song
About Me
I am a 4th-year Ph.D. student in the Department of Computer Science at City University of Hong Kong, advised by Prof. Nan Guan, and a member of CityU MLab. My research sits at the intersection of real-time systems and AI: making large-model inference — especially mixture-of-experts — predictable under tight memory and latency budgets, and building autonomous driving systems that can be reconfigured on the fly. More recently, I have been exploring real-time scheduling of LLM agents and LLM agents for OS-kernel formal verification. I have also worked on real-time task execution in ROS 2 robotic systems.
Prior to my Ph.D., I received my B.S. and M.S. degrees in Software Engineering from the University of Electronic Science and Technology of China in 2019 and 2022.
// the real-time atelier
Deadline Express
A brass-and-steam locomotive runs an MoE engine on a fixed memory budget. Keep the express on time.
Hover or tap the underlined labels — including the platform automaton — to see the paper behind each part.
Try Brake for signal with 2 experts at 30 ms — or switch to LRU and watch the tail.
What am I looking at?
Each control tick sends a batch of tokens from the locomotive to a mixture-of-experts engine. Only a few experts fit in the engine; loading an absent expert adds latency. If the total exceeds the control deadline, the express hesitates.
LRU evicts the least recently used expert. RT-MoE protects hot experts and uses free slots for prefetch.
The line outside the cab is also the research route: UESTC (the DNN years), CityU Shenzhen × Hon Hai (the Autoware / ATER autonomous truck years), and CityU MLab (the real-time MoE years). The locomotive stops briefly at each station and checkpoint.
Inspired by Autoware.Flex (RTCSA 2025) and RT-MoE (RTSS 2026). This is an illustrative simulation, not a benchmark.
Education
-
City University of Hong KongDepartment of Computer Science
-
University of Electronic Science and Technology of ChinaSchool of Information and Software Engineering
-
University of Electronic Science and Technology of ChinaSchool of Information and Software Engineering
Experience
-
City University of Hong Kong Shenzhen Research Institute · Hon Hai Research InstituteAutonomous Driving Light Truck Development · with Prof. Nan Guan and Prof. Jianping Wang
-
Chengdu Taiji Bear Technology Co., Ltd.FPGA Group
-
International Technological UniversityEmbedded System Group
Academic Service
-
IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (IEEE TCAD)
-
ACM Computing Surveys (CSUR)
Honors & Awards
-
Outstanding Graduate, Sichuan Province2022
-
Outstanding Graduate, University of Electronic Science and Technology of China2022
-
National Scholarship, Ministry of Education2021
News
Selected Publications view all
RT-MoE: Real-Time Expert Management for Memory-Constrained MoE Decoding
IEEE Real-Time Systems Symposium (RTSS) 2026 Accepted
Real-time expert management for mixture-of-experts decoding under tight memory budgets.
RT-MoE: Real-Time Expert Management for Memory-Constrained MoE Decoding
IEEE Real-Time Systems Symposium (RTSS) 2026 Accepted
Real-time expert management for mixture-of-experts decoding under tight memory budgets.
RTCSA
Autoware.Flex: Human-Instructed Dynamically Reconfigurable Autonomous Driving Systems
IEEE 31st International Conference on Embedded and Real-Time Computing Systems and Applications (RTCSA) 2025 Conference
A human-instructed, dynamically reconfigurable autonomous driving system built on the Autoware stack.
Autoware.Flex: Human-Instructed Dynamically Reconfigurable Autonomous Driving Systems
IEEE 31st International Conference on Embedded and Real-Time Computing Systems and Applications (RTCSA) 2025 Conference
A human-instructed, dynamically reconfigurable autonomous driving system built on the Autoware stack.
RTCSA
ATER: Adaptive Task Execution Rate Regulation for Enhanced Real-Time Performance in ROS 2
IEEE 31st International Conference on Embedded and Real-Time Computing Systems and Applications (RTCSA) 2025 Conference
Regulates ROS 2 task execution rates to improve real-time performance in robot software stacks.
ATER: Adaptive Task Execution Rate Regulation for Enhanced Real-Time Performance in ROS 2
IEEE 31st International Conference on Embedded and Real-Time Computing Systems and Applications (RTCSA) 2025 Conference
Regulates ROS 2 task execution rates to improve real-time performance in robot software stacks.
TNNLS
Critical Path-Based Backdoor Detection for Deep Neural Networks
IEEE Transactions on Neural Networks and Learning Systems (TNNLS) 2024 Journal
A critical-path-based detector that uses DNN interpretability to identify backdoor attacks across models.
Critical Path-Based Backdoor Detection for Deep Neural Networks
IEEE Transactions on Neural Networks and Learning Systems (TNNLS) 2024 Journal
A critical-path-based detector that uses DNN interpretability to identify backdoor attacks across models.
Beyond Research
the paper that started it all
Semi-Federated Scheduling of Parallel Real-Time Tasks on Multiprocessors
IEEE Real-Time Systems Symposium (RTSS) 2017 · DOI 10.1109/RTSS.2017.00015
In my first year as a master's student, I came across this paper by Nan Guan and his co-authors. I still remember how elegant the idea felt: a heavy DAG task with processing capacity x + δ is granted only x dedicated processors, while the remaining δ part is scheduled together with light tasks on shared processors. It is still one of my favourite papers in real-time scheduling. Years later I never became a pure real-time scheduling theorist, but I am still working every day toward becoming a real-time systems researcher — and this paper is still the standard I measure my own work against.
Read the paper at IEEE