Paper | Code
AdvBench-Omni is a dataset constructed to evaluate and reveal safety vulnerabilities in Omni-modal Large Language Models (OLLMs). It is based on a modality-semantics decoupling principle to study how OLLMs handle cross-modal safety risks and conflicts.
Omni-modal Large Language Models (OLLMs) expand multimodal capabilities but introduce new cross-modal safety risks. AdvBench-Omni reveals a significant vulnerability in these… See the full description on the dataset page:
https://huggingface.co/datasets/shipnebula/AdvBench-omni.
