Welcome to my personal website! I am Zhiyuan Liu (Chinese: 刘知远), a first-year master’s student at the School of Advanced Manufacturing and Robotics, Peking University.
My research interest centers around the fundamental question: How can we make large models more efficient? As models grow larger and more costly to train and use, I aim to explore new methods that improve their capability through innovation rather than scale.
To this end, my research spans several key areas, including efficient inference, data utilization, and model optimization.
🔥 News
- 2026.08: 🎉 We release our new preprint “MemOPD: On-Policy Distillation through Memory State Alignment for Long-Horizon Agents”.
- 2026.05: 🎉 Paper “dLLM-Cache: Accelerating Diffusion Large Language Models with Adaptive Caching” was accepted to ICML 2026.
- 2026.01: 🎉 Paper “Accelerating Diffusion Large Language Models with SlowFast Sampling: The Three Golden Principles” was accepted to ICLR 2026.
- 2026.01: 🎉 Paper “The Devil Behind the Mask: An Emergent Safety Vulnerability of Diffusion LLMs” was accepted to ICLR 2026.
- 2025.02: 🎉 Paper “Dataset Distillation with Neural Characteristic Function: A Minmax Perspective” was accepted by CVPR 2025 (Full Rating: 5/5/5). Thanks!
📝 Publications (* denotes the equal contribution.)
dLLM-Cache: Accelerating Diffusion Large Language Models with Adaptive Caching
Zhiyuan Liu *, Yicun Yang *, Yaojie Zhang, Junjie Chen, Chang Zou, Qingyan Wei, Shaobo Wang, Linfeng Zhang.
- Pioneered dLLM-Cache, a novel approach to accelerate diffusion large language models (dLLMs) by leveraging adaptive caching techniques.
- dLLM-Cache achieves up to 9.1x speedup over standard dLLM pipelines, with no performance loss on most tasks.
- Project
MemOPD: On-Policy Distillation through Memory State Alignment for Long-Horizon Agents
Zhiyuan Liu *, Tinghong Ye *, Chenghao Liu *, Yizhuo Li, Songfang Huang.
- Introduces MemOPD, which reconstructs the original memory states and token positions for valid teacher supervision during long-horizon agent training.
- MemOPD-3B improves F1 over PPO by up to 416.2%, while efficient invocation packing provides up to 1.63x actor-training speedup.
- Project
Dataset Distillation with Neural Characteristic Function: A Minmax Perspective
Shaobo Wang, Yicun Yang, Zhiyuan Liu, Chenghao Sun, Xuming Hu, Conghui He, Linfeng Zhang
- Pioneered NCFM, a novel dataset distillation approach that reframes the problem from a Characteristic Function perspective. This innovative approach casts dataset distillation within a min-max framework.
- NCFM achieves state-of-the-art performance while drastically reducing GPU memory requirements to 1/300th of prior leading methods and accelerating training by 20x.
- Project
Accelerating Diffusion Large Language Models with SlowFast Sampling: The Three Golden Principles
Qingyan Wei, Yaojie Zhang, Zhiyuan Liu, Dongrui Liu, Linfeng Zhang.
The Devil Behind the Mask: An Emergent Safety Vulnerability of Diffusion LLMs
Zichen Wen, Jiashu Qu, Dongrui Liu, Zhiyuan Liu, Chaochao Lu, Jing Shao, Conghui He, Linfeng Zhang, et al.
🎖 Honors and Awards
- 2023.12 National Scholarship (Top 1%), Ministry of Education of China
- 2023.11 Huawei Smart Base Scholarship (Top 1%), Huawei Technologies Co., Ltd.
📖 Educations
- 2026.09 -2029.06 (Expected), M.S., School of Advanced Manufacturing and Robotics, Peking University
- 2022.08 - 2026.06, Harbin Institute of Technology, Bachelor of Software Engineering
💻 Internships
-
2025.06 - 2025.09, Research Intern, Shanghai Artificial Intelligence Laboratory (Advised by Prof. Dongrui Liu).
-
2024.08 - 2026.07, Research Intern, EPIC Lab, Shanghai Jiao Tong University (Advised by Prof. Linfeng Zhang).