Existing LVLM benchmarks primarily focus onsingle-choice questions orshort-form responses, which donotadequately assess the capabilities ofLVLMs inreal-world human-AI interaction applications.

Theoretically, we prove a tighter error bound for CLLR; empirically, the superiority of CLLR is demonstrated across multiple domains.

Although highly desirable in many applications and in spite of its omnipresence in machine learning, current neural network architectures do not equivary to normalization.

The study of provable adversarial robustness has mostly been limited to classification tasks and models with one-dimensional real-valued outputs.