CV

Education

Ph.D. Candidate in Artificial Intelligence, The Hong Kong Polytechnic University, Jan. 2025 - Jan. 2028 (Expected)
- Supervisors: Prof. Hongxia Yang and Prof. Wenguang Chen
MEng in Artificial Intelligence, University of Science and Technology of China, Sep. 2021 - Jun. 2024
- Supervisor: Prof. Yong Wang
B.S. in Mechanical Engineering, Jiangnan University, Sep. 2017 - Jun. 2021
- Supervisor: Prof. Jun Zhang

AI infrastructure and machine learning systems for large-scale LLM training, including distributed training, memory optimization and system-algorithm co-design.
Efficient LLM training and inference, including FP8/FP4/NVFP4 training, INT4 rollout, quantization-aware training, compression and reasoning under efficiency constraints.

Ultra-Low-Bit LLM Pretraining in NVFP4 with System-Level Optimization, PolyU
- Designed end-to-end ultra-low-bit NVFP4 LLM training pipelines with framework-level optimization on GB200/B300, covering pretraining, post-training and RL.
- Developed system-algorithm co-design strategies for distributed training and investigated per-token scaling, block-level scaling and module-sensitive precision allocation.
End-to-End FP8 Training for Reasoning-Capable LLMs at Scale, PolyU / InfiX.ai
- Developed FP8 training recipes for reasoning-enhanced LLMs, covering continual pretraining, SFT, evaluation and deployment-oriented precision consistency.
- Worked on thousand-GPU training with NVIDIA H- and B-series GPUs, including numerical stability monitoring, throughput profiling and memory analysis.
INT4 RL with Real W4A16 Rollout for Train-Inference Alignment, with SGLang and LMSys
- Built an INT4 QAT-RL pipeline that combines fake-quantized training with real W4A16 serving-side rollout.
- Integrated SGLang and Marlin-based INT4 rollout kernels for MoE reasoning models.

Nov. 2024 - Present: Research Intern / Member of Technical Staff (Intern), InfiX.ai, HKSAR
- Built LLM pretraining infrastructure deployed on a 2K+ GPU H800 cluster with correctness validation and training-pipeline integration.
- Led research on ultra-low-bit LLM training and distributed optimization on GB200 GPUs.
- Developed inference acceleration and low-bit deployment pipelines for edge devices.
Jul. 2024 - Nov. 2024: Algorithm Engineer, Tencent Interactive Entertainment Group, Shenzhen
- Accelerated Stable Diffusion inference on consumer GPUs via distributed parallelism.
- Built automated web-agent pipelines for data collection.
Jul. 2023 - Apr. 2024: AI Infra Engineer Intern, Ant Group NextEVO Research Center, Hangzhou
- Contributed to ATorch, a distributed training framework built on Megatron-LM and DeepSpeed.
- Developed RLHF system components and improved reward-model training efficiency.
Oct. 2022 - May. 2023: Research Intern, AMD Xilinx DeePhi, Beijing
- Worked on LLM compression and quantization methods including SmoothQuant and AWQ.
- Deployed quantized models on AMD edge devices and contributed evaluation pipelines for Vitis-AI Model Zoo 3.0.
Mar. 2022 - Jun. 2022: Research Intern, iFlytek Technology Research Institute

Programming: Python, C++, CUDA, Golang
Systems and frameworks: Megatron-LM, DeepSpeed, vLLM, Ray, Triton
Research expertise: ultra-low-bit LLM training, FP8/FP4/NVFP4, quantization-aware training, distributed training systems, inference optimization
Additional: profiling, memory optimization, large-scale training infrastructure

Zhen Li, et al. Quantization Meets Reasoning: Exploring and Mitigating Degradation of Low-Bit LLMs in Mathematical Reasoning. Under review.
Zhen Li, Yupeng Su, Songmiao Wang, et al. InfiJanice: Joint Analysis and In-situ Correction Engine for Quantization-Induced Math Degradation in Large Language Models. Under review.
Wenjun Wang, Shuo Cai, Congkai Xie, Mingfa Feng, Yiming Zhang, Zhen Li, et al. InfiR2: A Comprehensive FP8 Training Recipe for Reasoning-Enhanced Language Models. arXiv.
Congkai Xie, Shuo Cai, Wenjun Wang, et al. InfiR: Crafting Effective Small Language Models and Multimodal Small Language Models in Reasoning. arXiv.
Qi Zhou, Yiming Zhang, Yanggan Gu, et al. Democratizing AI Through Model Fusion: A Comprehensive Review and Future Directions. Nexus, 2025.
Zeyu Liu, Yuhang Liu, Guanghao Zhu, et al. Infi-MMR: Curriculum-based Unlocking Multimodal Reasoning via Phased Reinforcement Learning in Multimodal Small Language Models. Under review.
Aofan Liu, Yuguo Yin, Hongjian Xing, Zhen Li, Yiyan Qi. MD3R: Minimizing Data Distribution Discrepancies to Tackle Inconsistencies in Multilingual Query-Code Retrieval. ACL@KnowFM, 2025.

Zhen Li, et al. InfiDeck: An AI Operating System with Integrated Hardware Extensions. China Patent.
Xiaotian Han, Zhen Li, et al. Reinforcement Learning Model Training Method and Device. U.S., China and Singapore Patent.

COMP6713 - Advanced Large Language Model and Beyond, Teaching Assistant, PolyU, 2026 Spring
COMP2021 - Object-oriented Programming, Teaching Assistant, PolyU, 2025 Fall
ML4432 - Machine Learning, Teaching Assistant, PolyU, 2025 Spring
CONT010177 - Modern Control Theory, Teaching Assistant, USTC, 2022 Fall