MCBench: A Multicontext Safety Assessment Benchmark for Omni Large Language Models

In three linesMCBench is a safety benchmark for omnimodal models (vision, audio, text) with 1196 scenarios across 4 safety categories. Evaluations reveal state-of-the-art Omni LLMs struggle to integrate multimodal cues for safety judgments, particularly on subtle risks, despite extracting modality-specific information.

Summary generated by Claude — human-verified