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.
Research overview & my role ↗Selected systems I have built or helped develop, spanning model evaluation, user simulation, and robotics.
I lead research on world-model hijacking: targeted adversarial attacks that aim to steer model behavior toward an attacker-chosen objective.
Research overview & my role ↗I built and revised the experimental pipeline for knowledge-state-aware user simulation: information-unit extraction, knowledge updates, query generation, and evaluation. I also implemented the initial human validation-study system.
Paper & contribution ↗I contributed controlled interventions and layer-wise representation analysis to investigate attention sinks and activation outliers. This work connects internal model behavior with inference stability and robustness.
Paper & contribution ↗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 ↗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.