Ph.D. student · HKUST
Building efficient systems for learning.
I am a Ph.D. student in Computer Science at the Hong Kong University of Science and Technology. My research focuses on machine learning systems, especially efficient LLM reinforcement learning and training infrastructure.
I work across algorithms and systems—from reusing shared computation and scheduling agent rollouts to making distributed training faster and more reliable. I am currently an RL infrastructure intern at Tencent Hunyuan.
01 / RESEARCH
From an idea to a working system.
Training-Time KV Caching for Schedule-Level Prefix Reuse in LLM RL Training
Reuse a prompt’s forward and backward computation across RL trajectories through training-time KV caching, gradient accumulation, and a three-phase schedule.
1.90× end-to-end RL speedup on long-prefix RAG
Paper ↗Understand when throughput gains outweigh sample-efficiency costs. Combine larger batches with learning-rate retuning to reduce the time needed to reach a target reward.
29% maximum time-to-target reduction in evaluated GRPO settings
Paper ↗Unlocking Full Efficiency of Token Filtering in Large Language Model Training
Turn token filtering into practical speedups: filter attention-backward activations and transform irregular sparse GEMM into smaller dense GEMM.
Huifang Huang, Zhuo Jin, Pengbo Li, Fuke Wu, Hailiang Yang
Insurance: Mathematics and EconomicsQijing Li, Xinyang Huang, Bowen Liu, Pengbo Li, Junxue Zhang, Kai Chen
APNET 2025Huifang Huang, Ting Gao, Pengbo Li, Jin Guo, Peng Zhang, Nan Du, Jinqiao Duan
Chaos02 / IN PRACTICE
Research, grounded in implementation.
RL Infrastructure Intern
Build checkpoint recovery for interrupted RL training, validate numerical alignment of the training engine, and improve shared-prefix reuse with KV caching and gradient accumulation.
Lab-incubated startup project
Developed a Kubernetes-based platform for training-job submission, GPU/CPU resource allocation, and lifecycle management, with REST APIs and direct SSH access to running pods.
Privacy-preserving computation · Backend development
Built a client–server backend for encrypted inference. Ported LightGBM to homomorphic computation with PALISADE and implemented privacy-preserving face recognition with TFHE-rs.
03 / BACKGROUND
Ph.D. in Computer Science
Machine Learning Systems · CGA 3.843 / 4.3Master’s · School of Mathematics and Statistics
National Scholarship · Top 10%Bachelor’s · School of Mathematics and Statistics
Outstanding GraduateLET’S CONNECT
For research collaborations and opportunities in ML systems, LLM training, and RL infrastructure, feel free to reach out.