Researchers have created a safety benchmark called Uni-SafeBench to evaluate the performance of Unified Multimodal Large Models (UMLMs). These models integrate understanding and generation capabilities within a single architecture. The existing safety benchmarks focus on isolated tasks, but Uni-SafeBench assesses the holistic safety of UMLMs when handling diverse tasks under a unified framework. The benchmark features six major safety categories across seven task types and has found that current open-source UMLMs have lower safety performance than specialized models for either generation or understanding tasks. --- Why it matters: This matters to researchers in AI because it highlights the potential trade-off between unification and safety in large models, which could impact their adoption in real-world applications. The results of Uni-SafeBench can inform the development of safer UMLMs that balance performance and safety. Source: https://arxiv.org/abs/2604.00547