Publications
Preprints & Under Review
BEACON: Budget-Efficient Discovery of Policy Violations in LLMs via Cognitive-Guided Monte Carlo Tree Search
We reframe LLM safety evaluation as budget-constrained failure discovery. Our Cognitive-Guided MCTS achieves 85.5–100% ASR across 6 frontier LLMs, discovering failures 3.7× faster than the strongest baseline.
EviGuard: Evidence-Guided Connector Intervention for Cross-Modal Safety Unlearning in Multimodal LLMs
We identify that cross-modal risk concentrates in a low-rank subspace at the visual-language connector. EviGuard reduces over-refusal rate from 30.3% to 22.3%, achieving 9.8% ASR and 68.0% Safe & Effective rate on OOD benchmarks.
VidTouch: A Multimodal Benchmark Dataset for Dynamic Visuo-Tactile Fabric Recognition
The first dynamic multimodal benchmark for fabric recognition with 145 categories. Our X3D + EfficientNet + MLP fusion achieves 98.6% in-category accuracy, revealing zero-shot generalization as an open challenge.
Patents
A Lane Change Decision-Making Method for Intelligent Vehicles in Foggy Days Based on Dynamic Game Theory
A decision-making framework for intelligent vehicles in foggy weather using dynamic game theory with CarSim & Simulink simulations. Passed preliminary examination.