Yiwei Chen陈一苇

I am a CS Ph.D. student at Michigan State University (MSU), advised by Prof. Sijia Liu, working on AI agents and trustworthy foundation models. Before that, I earned both my Bachelor's and Master's degrees from Xi'an Jiaotong University (XJTU), where I was admitted via the Special Class for the Gifted Young.

I’m curious by nature 🔭 and enjoy learning new things just for fun ✨. One day, I hope humanity makes it beyond Earth! 🚀🌍

profile photo
I am looking for research internships for summer 2027. Feel free to reach out!
Research

My research focuses on AI agents and the foundation models behind them: how agents act, collaborate, and improve themselves, and how to make these models safe. Currently, I'm working on:

  • Agentic SystemsGUI & CLI Agents, LLM Agents, Multi-Agent Systems, Recursive Self-Improvement
  • Foundation ModelsPost-Training & RL for LLMs/MLLMs, Reasoning, Evaluation
  • Trustworthy AISafety & Alignment, Machine Unlearning, Interpretability, Model Auditing
Internships
Amazon Applied Scientist Intern
Latent Page Memory for Consistent GUI World Modeling
May 2026 – Aug. 2026
Cisco Research Intern
LLMs for CVE Exploit Generation in Cybersecurity
May 2025 – Oct. 2025
News
  • 2026-07 Two papers accepted at COLM'26, including one as first author!
  • 2026-05 Joined Amazon as an Applied Scientist Intern this summer!
  • 2026-01 Two first-author papers accepted at ICLR'26!
  • 2025-05 Joined Cisco as a Research Intern!
  • 2025-05 Unlearning Isn't Invisible on detecting unlearning traces in LLMs from model outputs!
  • 2025-03 Safety Mirage on the application of machine unlearning to VLM safety alignment, my first project as a PhD student!
  • 2024-08 Started PhD journey at Michigan State University!
  • 2024-06 Graduated from Xi'an Jiaotong University!
Selected Publications

(* denotes equal contribution)

Latent Page Memory for GUI World Modeling
Latent Page Memory for Consistent GUI World Modeling
Yiwei Chen, H. Xu, N. Vesdapunt, X. Li, H. Liu, C. P. Huynh, S. Zhu, Sijia Liu, J. Thomas, K. K. Fu
In Submission
Paper

A GUI world model that remembers pages stays consistent when agents return to them.

ReDreamer keeps each observed page in latent memory slots and recalls it by content on revisits, reaching state-of-the-art trajectory-level GUI world modeling and raising GUI-agent step accuracy.

Distributed Multi-Agent Training
Train Locally, Learn Together: Distributed Multi-Agent Training via ADMM
Yiwei Chen, Changsheng Wang, Yicheng Lang, Qing Qu, Alfred O. Hero, Sijia Liu
In Submission
CodePaper

Joint MAS training without a central trainer, comparable to centralized training.

Collaborative MAS learning cast as consensus optimization: an ADMM framework alternates role-local learning with inter-agent consensus, evaluated on distributed safety alignment and general reasoning.

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
ICLR 2026
CodePaper

Unlearning removes the harm that safety fine-tuning only masks.

Safety fine-tuning of MLLMs inherits the bias of its training data, producing spurious correlations and over-rejections under one-word attacks.

Unlearn Trace
Unlearning Isn't Invisible: Detecting Unlearning Traces in LLMs from Model Outputs
Yiwei Chen*, Soumyadeep Pal*, Yimeng Zhang, Qing Qu, Sijia Liu
ICLR 2026Oral @ MUGen ICML'25
CodePaper

Forgetting is not invisible: unlearned models still leak that they were unlearned.

Unlearned LLMs keep persistent fingerprints in their outputs and hidden activations, enough for a classifier to tell that unlearning happened at all.

LLM Lineage
Who Built This Model? Tracing LLM Lineage via Spectral Fingerprints in Weight Space
Yiwei Chen*, Bingqi Shang*, Sijia Liu
COLM 2026
ProjectCodePaper

Model provenance from weights alone, with no access to training data.

A geometric fingerprint of weight space traces which model a checkpoint descends from, using spectral energy for coarse discrimination and subspace alignment for fine-grained analysis.

Cybersecurity Exploit Benchmark
CVE-Conditioned Exploit Generation with LLMs: A Data-Centric Study of Benchmarking and Fine-Tuning
Yiwei Chen*, Lichi Li*, Kai Cheung, Vinny Parla, Ganesh Sundaram
Technical Report
Paper

With the right fine-tuning data, an 8B model matches frontier LLMs at exploit generation.

The first data-centric benchmark for CVE-conditioned exploit generation: a 6-level context hierarchy, an 8-criterion evaluation, and 17 LLMs measured against it.

Backdoor Unlearn
Forgetting to Forget: Attention Sink as a Gateway for Backdooring LLM Unlearning
Bingqi Shang*, Yiwei Chen*, Yihua Zhang, Bingquan Shen, Sijia Liu
COLM 2026
CodePaper

Attention sinks are a backdoor that quietly restores “forgotten” knowledge.

Triggers planted at attention sinks let an unlearned model recover the forgotten knowledge on cue, while behaving normally whenever the trigger is absent.

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
CodePaper

Framing viewport prediction as tile classification is what buys the robustness.

Viewport prediction recast as tile classification, using a multi-modal fusion transformer.