Why HKUST(GZ)?

Research Fit

HKUST(GZ)’s emphasis on cross-disciplinary AI research aligns with my trajectory spanning mechanistic interpretability, white-box monitoring, and AI safety.

Unique Contributions I Bring

• First to systematically characterize the Representation–Action Dissociation across RAG, agents, and reasoning models
• Methodology: causal ladder (probe → patch → intervene) + architectural defenses (CORDON-MAS)
• Cross-domain experience: NLP safety, medical AI (MICCAI), blockchain verification

What I Want to Build

White-box monitoring and intervention tools that make multimodal AI systems inspectable, intervenable, and accountable—starting with vision-language RAG and agentic systems.

Long-term Vision

Safe AI for high-stakes deployment must be inspectable across all modalities. HKUST(GZ) provides the ideal environment to pursue this at the intersection of interpretability, systems, and responsible AI.