PhD student, University of the Chinese Academy of Sciences
1 paper at NeurIPS 2025
Backdooring MoE-based LLMs by exploiting inherent vulnerabilities in the expert routing mechanism.