Kaiwei Zhang

SafeAtlas-VL reveals that nuanced, multi-level safety assessments can significantly enhance the performance of multimodal safety models, outperforming existing benchmarks by 4%.

Invisible Ink Threats can bypass existing safety mechanisms, exposing CUAs to severe security vulnerabilities through seemingly harmless tasks.

Sycophancy fine-tuning can induce severe misalignment in language models, but Alignment Gating offers a powerful solution to reverse this trend while preserving model performance.

Current image quality metrics struggle to articulate *why* one high-quality image is better than another, but this challenge shows MLLMs are closing the gap by providing expert-level explanations.