Research & projects

Mechanisms. Behavior. People.

I connect the study of internal model representations with questions about reliability and human understanding.

01 / MECHANISMS

What happens
inside the model?

I study attention, activation structure, and internal representations to understand how model behavior arises.

02 / RELIABILITY

Can we trust
what it predicts?

I study world-model hijacking and targeted adversarial threats to model behavior.

03 / PEOPLE

Does it understand
the person?

I build and study user simulation systems that account for evolving knowledge and multi-turn interaction.

01 / EXPERIENCE

Where I’ve worked

Mar 2026 — Present

Northwestern University

Undergraduate Research Assistant · Remote

Leading ongoing research on world-model hijacking and targeted adversarial attacks and contributing to mechanistic interpretability work on attention sinks and activation outliers.

Oct 2025 — Sep 2026

Michigan Institute for Data and AI in Society

Undergraduate Research Assistant · University of Michigan

Worked with Prof. Q. Vera Liao and Dr. Yoonjoo Lee on knowledge-state-aware user simulation and multi-turn information-seeking interactions. Built experimental pipelines and the initial human validation-study system for KnowSim.

Oct 2024 — Oct 2025

Data-informed Design and Intelligent Systems Laboratory

Undergraduate Research Assistant · Shanghai Jiao Tong University

Worked with Prof. Youyi Bi and Dr. Chang Liu on imitation learning for robotic manipulation. Reproduced MimicPlay and developed keyboard-based robotic-arm teleoperation for collecting demonstration trajectories.

02 / BUILDING

Selected projects

Explore projects ↗
Robotics · Research engineering

Learning from demonstrations

A keyboard-based teleoperation framework for robotic arms, built to collect human demonstrations for imitation learning. This work grew out of reproducing and studying long-horizon manipulation with MimicPlay.

MimicPlay reference implementation ↗
Course project · In development

DATA TAMER

A team project for EECS 448 exploring StarCraft II win-probability prediction and attribution. The planned approach combines spatial sequences and structured game-state features to study which aspects of a state influence the predicted outcome.

Current focus: data reconstruction, predictive modeling, and interpretable evaluation.

Selected work

Ongoing research and published work, side by side.

Normal and attacked pipelines: input image, imagined video, and downstream IDM action
Conceptual problem overview
Ongoing· Project lead · World-model safety

Targeted attacks on world models

I lead research on world-model hijacking: targeted adversarial attacks that aim to steer model behavior toward an attacker-chosen objective.

Methods and results will accompany a public release.

Standard AttnResidual compared with OASIS
OASIS, Figure 1 · Enlarge ↗
NeurIPS 2026· COLM 2026 Efficient Reasoning Workshop · Spotlight

Attention Sinks and Outliers in Attention Residuals

Haozheng Luo*, Haoran Dai*, Jingyuan Huang*, Shaoyang Zhang, Xi Chen, Eric Hanchen Jiang, Yijiang Li, Chenghao Qiu, Chenwei Xu, Zhenyu Pan, Haotian Zhang, Binghui Wang, and Yan Chen

Understanding how attention sinks and activation outliers interact with cross-layer routing, and using null-aware attention to improve model robustness.