Jiahao Zhang

Peak local similarity in untrimmed video retrieval often flags irrelevant clips on accidental matches, but calibrating those peaks against query-routed global registers reliably eliminates spurious false positives without costly feature re-fusion.

Text safety neurons in VLMs are localized and critical, while visual safety mechanisms are diffuse and complex, revealing a fundamental gap in current alignment strategies.

RAG systems are not just enhanced by external knowledge; they are also vulnerable to targeted attacks that can compromise accuracy, privacy, and fairness.

HODAgent's unified architecture enables humanoid robots to adaptively manage service tasks in real-time, achieving up to 92% success rates on physical platforms.