Yiwei Chen
（陈一苇）

I am a second-year CS Ph.D. student focusing on Trustworthy ML and Scalable AI at Michigan State University (MSU), where my advisor is Prof. Sijia Liu. Before that, I earned both my Master's degree and Bachelor's degree from Xi’an Jiaotong University (XJTU). Also, I was admitted to the Young Gifted Program at XJTU.

Email  /  Google Scholar  /  Github  /  X  /  LinkedIn  /  CV

profile photo
Research

My research interests lie at the intersection of robust and explainable artifical intelligence with the long-term goal of making AI systems safe and scalable. Currently, I'm working on:

  • Post-training of LLM, MLLM
  • Trustworthy algorithms on LLM, MLLM
  • Machine Unlearning, Safety Alignment
News
  • 2024-05 Joined Cisco as a Research Intern!
  • 2025-05 Unlearning Isn't Invisible focusing on the detecting unlearning traces in LLM from model outputs!
  • 2025-03 Safety Mirage focusing on the application of machine unlearning on VLM safety alignment, which is the first project during my PhD journey!
  • 2024-08 Started PhD journey at Michigan State University!
  • 2024-06 Graduated from Xi'an Jiaotong University!
Publications (* denotes equal contribution)
Unlearn Trace
Unlearning Isn't Invisible: Detecting Unlearning Traces in LLMs from Model Outputs
Yiwei Chen*, Soumyadeep Pal*, Yimeng Zhang, Qing Qu, Sijia Liu
Under Review , 2025
Short version accepted for Oral at MUGen @ ICML'25
Code / arXiv

Large language models retain persistent fingerprints in their outputs and hidden activations even after unlearning. These traces enable detection of whether a model has been “unlearned,” exposing a new vulnerability in reverse-engineering forgotten information.

VLM Unlearn
Safety Mirage: How Spurious Correlations Undermine VLM Safety Fine-Tuning and Can Be Mitigated by Machine Unlearning
Yiwei Chen*, Yuguang Yao*, Yihua Zhang, Bingquan Shen, Gaowen Liu, Sijia Liu
Under Review , 2025
Code / arXiv

Conventional safety fine-tuning of MLLMs suffers from a ``safety mirage'' caused by training bias, leading to spurious correlations and over-rejections under one-word attacks. Employing unlearning algorithms effectively remove harmful content and mitigate these issues.

Backdoor Unlearn
Forgetting to Forget: Attention Sink as A Gateway for Backdooring LLM Unlearning
Bingqi Shang*, Yiwei Chen*, Yihua Zhang, Bingquan Shen, Sijia Liu
Under Review , 2025
Code / arXiv

Demonstrated that LLM unlearning can be compromised through attention-sink-guided backdoor unlearning, where triggers placed at attention sinks enable models to recover forgotten knowledge while maintaining normal behavior without triggers.

MFTR
Tile Classification Based Viewport Prediction with Multi-modal Fusion Transformer
Zhihao Zhang*, Yiwei Chen*, Weizhan Zhang, Caixia Yan, Qinghua Zheng, Qi Wang, Wangdu Chen
ACM MM , 2023
Code / arXiv

Propose a tile classification based viewport prediction method with Multi-modal Fusion Transformer to improve the robustness of viewport prediction.