Yizhen Yuan
袁宜桢

Student, Researcher, Optimist


A picture taken when I live in Purdue

Welcome

Hi~ I am Yizhen Yuan. I am currently a Ph.D. candidate at Tsinghua University in Beijing, China, and fortunately advised by Professor Liu, Yunxin and Professor Li, Yuanchun. I completed my undergraduate studies at Purdue University in West Lafayette, U.S .

I am interested in LLM Security, Privacy and Reliability. This is also the section I am responsible for in the paper Personal LLM Agents.

I am deeply grateful for their help.
Without their help, I wouldn't be able to become the person I am today.

Thanks

When I come in to Purdue, Professor Turkstra have set up a good value as a computer scientist for me. When I took CS 397 (Honor Seminar), Professor Berkay chose me to do project with him amoung a lot of Computer Science Honor students. When I took CS 381 (Undergrad Crypto) and CS 580 (Grad Algorithm), Professor Blocki helped me a lot even with some material not covered in classes. During the following winter break, Professor Benotman helped me with my project. I also have received help from Professor Xiangxiong Zhang, Professor Jean Honorio, Tiantian Qin, Professor Kihong Park, Melanie and Professor Maji. Thanks them a lot. Without them, I will not be confident and enegetic in studying.

It is my greatest honor to be a student of Professor Li Yuanchun and Professor Liu Yunxin. I cannot imagine what my PhD journey would have been like without their help. No words can fully express the help they have given me and my gratitude toward them. I would gladly vouch for them: they are excellent, outstanding, and kind professors and mentors.

Education


Undergraduate - Purdue University, West Lafayette, U.S.

Major in Computer Science Honor (Machine Intelligence & Security) ,Mathematics and Statistics. Yes, I am a triple major student! (Well, that is true. However, due to the class overlapping, it is not as hard as you thought)

Graduate - Tsinghua University, Beijing, China

Major in Electronic Engineering . Belongs to Institute for AI Industry Research (AIR). Advised by Professor Liu, Yunxin and Professor Li, Yuanchun.

Publications

PatchBackdoor: Backdoor Attack against Deep Neural Networks without Model Modification

[ACM MM 2023 (CCF-A)] In Proceedings of the 31st ACM International Conference on Multimedia.

Yizhen Yuan, Rui Kong, Shenghao Xie, Yuanchun Li, Yunxin Liu
github pdf
Benchmarking LLM's Capability in Reasoning over Conflicting Web References

[ACL 2026 (CCF-A)]

Yizhen Yuan, Rui Kong, Dongze Li, Yuanchun Li, Yunxin Liu
pdf
LevelKV: Hierarchical KV Cache Pruning for Efficient and Reliable LLM Inference

[IEEE ToC (CCF-A)] Minor Revision Submitted

Yizhen Yuan, Rui Kong, Bolin Xian, Yuanchun Li, Ting Cao, Xianyuan Zhan, Ya-Qin Zhang, Yunxin Liu
ConvReLU++: Reference-based Lossless Acceleration of Conv-ReLU Operations on Mobile CPU

[MobiSys 2023 (CCF-B)] In Proceedings of the 21st ACM International Conference on Mobile Systems, Applications, and Services.

Rui Kong, Yuanchun Li , Yizhen Yuan, Linghe Kong.
pdf
An Empirical Study of LLM Reasoning Ability Under Strict Output Length Constraint

[EMNLP 2025 (CCF-B)]

Yi Sun, Han Wang, Jiaqiang Li, Jiacheng Liu, Xiangyu Li, Hao Wen, Yizhen Yuan, Huiwen Zheng, Yan Liang, Yuanchun Li, Yunxin Liu
pdf
AOHP: An Open-Source OS-Level Agent Harness for Personalized, Efficient and Secure Interaction

[arXiv Preprint]

Shanhui Zhao, Jiacheng Liu, Guohong Liu, Jichao Yan, Jialei Ye, Yuhao Yang, Hao Wen, Shizuo Tian, Yizhen Yuan, Yuxuan Chen, Yunxin Liu, Ju Ren, Ya-Qin Zhang, Chao Huang, Yao Guo, Yuanchun Li
pdf
WiP: An On-device LLM-based Approach to Query Privacy Protection

[MobiCom EdgeFM Workshop] In Proceedings of the Workshop on Edge and Mobile Foundation Models.

Yizhen Yuan, Rui Kong, Yuanchun Li, Yunxin Liu
pdf

Patents

Method and Apparatus for Key-Value Cache Data Pruning Processing
Application No.: 202611076427.9
This is a Chinese patent and has no official English title. The above name is a translation only and is not part of any official material.
Combines token attention scores with the activation importance of value vectors to perform comprehensive scoring and ranking, addressing the output distortion of traditional pruning; and implements graded caching and differentiated pruning based on the semantic attributes of prompts, improving the safety and controllability of models.

White Papers

Personal LLM Agents: Insights and Survey about the Capability, Efficiency and Security
Preprint/Survey about Personal Large Language Models.
Project Lead: Yuanchun Li
Section Lead: Hao Wen, Weijun Wang, Xiangyu Li, Yizhen Yuan, Guohong Liu
Cooperate With: Jiacheng Liu, Wenxing Xu, Xiang Wang, Yi Sun, Rui Kong, Yile Wang, Hanfei Geng, Jian Luan, Xuefeng Jin, Zilong Ye, Guanjing Xiong, Fan Zhang, Xiang Li,Mengwei Xu, Zhijun Li, Peng Li, Yang Liu, Ya-Qin Zhang, Yunxin Liu
If you have any question about "Section Security", please let me know, I am responsible for this Section.
arxiv github pdf
White Paper on the L1-L5 Safety Framework for General-Purpose AI Agents
[WAIC 2026] Co-authored with Shanghai AI Laboratory, Tsinghua AIR, Concordia AI, and Huawei.
Yizhen Yuan (sole student author from AIR)
white paper

Contact Information
yuanyz21@tsinghua.org.cn
If you want to know more about my research/Purdue/Tsinghua, I am glad to answer.