I am currently a master’s student in Computer Science at Nanjing University, supervised by Prof. Chen Tian and Prof. Zhibin Wang, working on machine learning systems.
I’m interested in AI infrastructure, including:
🌱 RL infra, with a focus on fully async RL and improving RL efficiency through scalable rollout systems, training-serving coordination, and end-to-end system optimization.
🌲 LLM serving systems, including request scheduling, disaggregated serving, and resource-aware orchestration.
💼 Internships
2025.12 - Present, StepFun, Foundation Model Infra. Working on RL training frameworks and vLLM rollout, including post-processing, speculative decoding for post-training, RL scheduling, and end-to-end RL efficiency.
2024.06 - 2025.02, Tencent WeChat, Search Engine Backend Development. Built backend observability and log-tracing systems, developed stability analysis tools for search experiments, and improved cache components for full-page search architecture.
🚀 Projects
2025.03 - 2025.09, Decode-Phase Migration Scheduling for LLM Inference under PD Disaggregation. Collaboration with Alibaba Tongyi Lab. Developed a vLLM-based PD disaggregation scheduler and predictor that continuously estimates generation length and migrates requests across decode instances; achieved 2.24x goodput in 32K long-output settings.
2023.09 - 2024.05, Node Failure Repair Scheduling for Distributed Storage Systems. Collaboration with Sangfor. Designed scheduling algorithms for erasure-coded distributed storage under asymmetric networks, improving repair performance by 26% to 127% over SOTA in large-scale cluster evaluations.
🧩 Open Source
vLLM · Contributor
PRs: #33671, #33690, #34211. Meetup: vLLM Meetup.
📝 Publications
Scheduling Mixed RL Rollouts Beyond Prefix Locality
Zetao Hong, Song Yuan, Yuanhao Ding, Yibo Zhu, Daxin Jiang, Zhibin Wang, Chen Tian
- Presents MISA-T, a routing-layer admission policy for mixed RL rollout serving, e.g., MOPD-style workloads that mix RLVR, RLHF, and agentic sessions.
- Improves rollout throughput by up to 53.3% over a cache-aware vLLM Router, and reduces mean iteration time by 22.8% in matched end-to-end RL training.
STAR: Decode-Phase Rescheduling for LLM Inference
Zhibin Wang, Zetao Hong, Xue Li, Zibo Wang, Shipeng Li, Qingkai Meng, Qing Wang, Chengying Huan, Rong Gu, Sheng Zhong, Chen Tian
- Proposes STAR, a decode-phase rescheduling system that uses remaining generation length prediction for LLM inference.
- Reduces P99 TPOT by 75.1% and achieves 2.63x higher goodput under evolving decode workloads.
🏆 Contributed Technical Reports
🎩 Educations
- 2025 - 2028, M.Sc. in Computer Science, Nanjing University.
- 2021 - 2025, B.Eng. in Computer Science, Huazhong University of Science and Technology.