Safety, Evidence and Inference Progress
In depth: CLEW, KernelArmor and Stint, with GPU systems, agent evaluation and community engineering evidence.
Read noteI’m Guangba Yu. My research focuses on the reliability and optimization of cloud and AI systems.

🏆 Our paper PreServe received the ACM SIGSOFT Distinguished Paper Award at ICSE 2026. CUHK announcement ↗
🏆 Our paper iKnow received the ACM SIGSOFT Distinguished Paper Award at ASE 2025. CUHK announcement ↗
🎉 Our Alert Management and Cloud Ops RAG Paper has been accepted by ASE'25.
🎉 Our Network Telemetry RCA Paper has been accepted by ToN. Congras to Hongyang!
🎉 Our Cloud Incident Management Paper has been accepted by TSE. Congras to Zilong!
🎉 Our Survey on Failure Analysis and Fault Injection in AI Systems has been accepted by TOSEM.
🎉 Log-based LLM training failures diagnosis framework L4 has been accepted by FSE'25. Congras to Zhihan!
🎉 Code knowledge enhanced root cause analysis framework CoCA has been accepted by ICSE'25. Congras to Yichen!
🎉 Microservice Trace Framework Mint has been accepted by ASPLOS'25. Congras to Haiyu!
🎉 FaaS configuration optimization Paper FaaSConf has been accepted by ASE'24 (118/587). Congras to Yilun!
🏆 Trace sampling paper TraStrainer has won Best Paper Award at FSE'24.
🎉 Request-level fault injection paper MicroFI has been accepted by TDSC. Congras to Hongyang!
🎉 Root cause change analysis paper ChangeRCA and trace sampling paper TraStrainer have been accepted by FSE'24 (56/483).
🎉 Multi-modal observability data RCA paper Nezha and database configuration optimization paper Diagconfig has been accepted by FSE'23.
In depth: CLEW, KernelArmor and Stint, with GPU systems, agent evaluation and community engineering evidence.
Read noteCoding-agent waste often spans sessions: rediscovering a test command confirmed yesterday, repeatedly loading unused tool descriptions, or spending more turns on the same failure. TokenEconomizer uses repository history to adjust initial knowledge, tool interfaces and failure intervention without modifying the underlying agent or model. It targets the cost of sustained repository work.
Sparse attention reads only a subset of keys and values, but selection can still scan the full history. NVIDIA separates execution into an indexer, Top-k and core attention: the last operates on selected k entries while the first may still scale with context N. Selection rules, tensor shapes and kernels must be designed together, or sparsity simply relocates the bottleneck.
SRE = Software Reliability Engineering
Ph.D. in Computer Science and Technology
Advisor: Prof. Pengfei Chen
M.Eng. in Computer Technology
Advisor: Prof. Pengfei Chen
B.Eng. in Medical Information Engineering
Advisor: Prof. Xiufeng Liu