About
I am a PhD student at OPTML Group at Michigan State University, advised by Prof. Sijia Liu. I received the MS degree in Computer Science at Northwestern University (NU) in June 2025, advised by Prof. Qi Zhu and Prof. Xiao Wang. Prior to NU, I obtained my B.E. in Software Engineering at Tongji University in July 2023.
Away from research I am an avid astrophotographer. My work is collected in the photography gallery.
News
Three papers have been accepted to COLM 2026.
Started as a research scientist intern at the MIT-IBM Computing Research Lab.
New preprint "Subspace Control: Turning Constrained Model Steering into Controllable Spectral Optimization" is now on arXiv.
Our preprint "Forgetting to Forget: Attention Sink as A Gateway for Backdooring LLM Unlearning" is now on arXiv.
Started my Ph.D. in Computer Science at Michigan State University, joining the OPTML Group advised by Prof. Sijia Liu.
Our paper about private downstream task adaptation of pre-trained transformers has been accepted to CVPR 2025.
Selected Publications
View All →* denotes equal contribution.
Who Built This Model? Tracing LLM Lineage via Spectral Fingerprints in Weight Space
Yiwei Chen*, Bingqi Shang*, Sijia Liu
COLM 2026
Traces LLM lineage from weight geometry alone, with no access to input data: spectral energy separates independently trained families and subspace alignment discriminates models built on a shared base, across more than 110 open-weight model pairs.
Forgetting to Forget: Attention Sink as A Gateway for Backdooring LLM Unlearning
Bingqi Shang*, Yiwei Chen*, Yihua Zhang, Bingquan Shen, Sijia Liu
COLM 2026
Shows attention sinks act as gateways for backdooring LLM unlearning, so forgotten knowledge returns only when a hidden trigger is present.
Subspace Control: Turning Constrained Model Steering into Controllable Spectral Optimization
Yancheng Huang*, Changsheng Wang*, Chongyu Fan, Yicheng Lang, Bingqi Shang, Yang Zhang, Mingyi Hong, Qing Qu, Alvaro Velasquez, Sijia Liu
COLM 2026
Steers a model toward one objective without breaking a second: orthogonalizing the merged spectral subspace removes the interference between the two, and the intervention is applied only at the layers and steps that need it.
Split Adaptation for Pre-trained Vision Transformers
Lixu Wang*, Bingqi Shang*, Yi Li, Payal Mohapatra, Wei Dong, Xiao Wang, Qi Zhu
CVPR 2025
Splits a pre-trained vision transformer so the client runs a quantized frontend and the owner keeps a private backend, which protects the client data and the model weights at once rather than trading one for the other.