Proof (LLM) robustness is to solve the 0-1 backpack problem

Talker: Huanran Chen

Bio: Huanran Chen is a PhD student from TSAIL (Fall 2025), advised by Prof. Jun Zhu, and closely collaborate with Prof. [Yinpeng Dong]. He has a keen interest in the physics of machine learning. His unattainable yet motivating dream is to elevate AI to the realm of science, making every phenomenon explainable and predictable.

Key note:

• What is worst-case robustness?
• Why worst-case robustness?
• Upper bounding worst-case robustness
• Lower bounding worst-case robustness (knapsack problem)
• Specific application to text domain
• Theoretical insights (e.g., diffusion>MaskGen>ARM)
• Applications

slides