Ethan Zhen LI
About me
- I am a second-year Ph.D. candidate in Artificial Intelligence at The Hong Kong Polytechnic University, fortunately supervised by Prof. Hongxia Yang at PolyU and Prof. Wenguang Chen at Tsinghua University.
- Earlier, I received my Master’s degree in Artificial Intelligence from the University of Science and Technology of China (USTC), supervised by Prof. Yong Wang.
- My current research focuses on AI infrastructure and machine learning systems for efficient LLM training and inference, especially ultra-low-bit training, quantization, and reasoning-capable foundation models.
I am actively looking for academic collaboration. We are working with NVIDIA H- and B-series GPU platforms, including DGX Spark, B200/H200, GB200 and B300 systems. Feel free to contact me if you are interested in NVFP4/FP4/MXFP8 end-to-end training, low-bit RL, or efficient reasoning models.
Research Interests
- AI Infrastructure: large-scale LLM training systems, distributed training, memory optimization, and system-algorithm co-design.
- Efficient LLM Training and Inference: FP8/FP4/NVFP4 training, INT4 rollout, quantization-aware training, model compression, and inference acceleration.
- Reasoning under Efficiency Constraints: preserving and improving mathematical and multimodal reasoning capabilities in low-bit LLMs.
Research Highlights
- Ultra-Low-Bit LLM Pretraining in NVFP4 with System-Level Optimization: Designed end-to-end NVFP4 training pipelines with framework-level optimization, investigated per-token scaling, block-level scaling and module-sensitive precision allocation, and explored scalable low-bit training on large GPU clusters.
- End-to-End FP8 Training for Reasoning-Capable LLMs at Scale: Developed FP8 training recipes covering continual pretraining, supervised fine-tuning, evaluation and deployment-oriented precision consistency, with hybrid-granularity FP8 quantization and BF16-level reasoning performance.
- INT4 RL with Real W4A16 Rollout for Train-Inference Alignment: Built an INT4 QAT-RL pipeline that combines training-side fake quantization with serving-side real W4A16 rollout, improving rollout throughput on Qwen3-30B-A3B while keeping reasoning performance stable.
Publications and Manuscripts
(*Equal contribution)
InfiR2: A Comprehensive FP8 Training Recipe for Reasoning-Enhanced Language Models
2025 Under Review
Authors: Wenjun Wang, Shuo Cai, Congkai Xie, Mingfa Feng, Yiming Zhang, Zhen Li, Kejing Yang, Ming Li, Jiannong Cao, Yuan Xie, Hongxia Yang✉Democratizing AI Through Model Fusion: A Comprehensive Review and Future Directions
2025 Nexus
Authors: Qi Zhou, Yiming Zhang, Yanggan Gu, Yuanyi Wang, Zhijie Sang, Zhaoyi Yan, Zhen Li, Shengyu Zhang, Fei Wu, Hongxia Yang✉InfiJanice: Joint Analysis and In-situ Correction Engine for Quantization-Induced Math Degradation in Large Language Models
2025 Under Review
Authors: Zhen Li*, Yupeng Su*, Songmiao Wang, Runming Yang, Congkai Xie, Aofan Liu, Ming Li, Jiannong Cao, Yuan Xie, Ngai Wong✉, Hongxia Yang✉Infi-MMR: Curriculum-based Unlocking Multimodal Reasoning via Phased Reinforcement Learning in Multimodal Small Language Models
2025 Under Review
Authors: Zeyu Liu*, Yuhang Liu*, Guanghao Zhu, Congkai Xie, Zhen Li, Jianbo Yuan, Xinyao Wang, Qing Li, Shing-Chi Cheung, Shengyu Zhang, Fei Wu, Hongxia Yang✉MD3R: Minimizing Data Distribution Discrepancies to Tackle Inconsistencies in Multilingual Query-Code Retrieval
2025 ACL@KnowFM
Authors: Aofan Liu, Yuguo Yin, Hongjian Xing, Zhen Li, Yiyan Qi✉Quantization meets reasoning: Exploring llm low-bit quantization degradation for mathematical reasoning
2025 arXiv
Authors: Zhen Li*, Yupeng Su*, Runming Yang, Congkai Xie, Zheng Wang, Zhongwei Xie, Ngai Wong✉, Hongxia Yang✉InfiR: Crafting Effective Small Language Models and Multimodal Small Language Models in Reasoning.
2025 arXiv
Authors: Congkai Xie, Shuo Cai, Wenjun Wang, Pengxiang Li, Zhijie Sang, Kejing Yang, Yiming Zhang, Zhen Li, Guanghao Zhu, Zeyu Liu, Yang Yu, Yuhang Liu, Su Lu, Baoyi He, Qi Zhou, Xiaotian Han, Jianbo Yuan, Shengyu Zhang, Fei Wu, Hongxia Yang✉
Patents
- InfiDeck: An AI Operating System with Integrated Hardware Extensions · China Patent
- Reinforcement Learning Model Training Method and Device · U.S., China and Singapore Patent
Grants
- Numerical Stability and End-to-End Performance Optimization of Low-Bit Mixed-Precision Training for Large Language Models · Tsinghua-PolyU Joint Research Initiative Fund
- Scaling Low-Bit Training for Efficient Large Model Deployment · General Research Funding
- Dynamic Precision-Aware Low-Bit Training for Multimodal LLMs on Blackwell GPUs · NVIDIA Academic Grant Program Award
Industrial Experiences
- Research Intern / Member of Technical Staff (Intern) @InfiX.ai · 2024.11 - Present
- Algorithm Engineer @Tencent Interactive Entertainment Group (IEG) · 2024.07 - 2024.11
- AI Infra Engineer (Intern) @Ant Group NextEVO Research Center · 2023.06 - 2024.04
- Research Intern @AMD Xilinx DeePhi · 2022.10 - 2023.05
- Research Intern @iFlytek Technology Research Institute · 2022.03 - 2022.06
Project Experiences
Awards
- PolyU Research Postgraduate Scholarship, 2025
- Outstanding Postgraduate Student of USTC, 2024
- 1st-Class Academic Scholarship of USTC, 2021-2024
- Merit Student Award of the Province, 2021
- Annual Outstanding Student of the Province, Province Level · 2019
- Merit Student Award, 2019
- Outstanding Student Leader Award, 2021
Teaching
- Teaching Assistant, COMP6713 - Advanced Large Language Model and Beyond, PolyU · 2026 Spring
- Teaching Assistant, COMP2021 – Object-oriented Programming, PolyU · 2025 Fall
- Teaching Assistant, ML4432 – Machine Learning, PolyU · 2025 Spring
- Teaching Assistant, CONT010177 – Modern Control Theory, USTC · 2022 Fall
Skills
- Programming: Python, C++, CUDA, Golang
- Systems and Frameworks: Megatron-LM, DeepSpeed, vLLM, Ray, Triton
- Expertise: Ultra-low-bit LLM training, FP8/FP4/NVFP4, quantization-aware training, distributed training systems, inference optimization, profiling and memory optimization
Extracurricular Activities
- President of the University Student Union