Uni-SafeBench Introduces Comprehensive Safety Benchmark for Unified Multimodal AI Models
Uni-SafeBench, a new safety benchmark for Unified Multimodal Large Models (UMLMs), is introduced in a paper on arXiv (2604.00547). UMLMs merge understanding and generation within a single architecture, yet their safety implications remain underexplored. Existing benchmarks only cover isolated tasks, failing to assess holistic safety. Uni-SafeBench provides a taxonomy of six safety categories across seven task types, paired with Uni-Judger, a framework that separates contextual safety from intrinsic safety. Evaluations reveal that the original safety alignment of underlying LLMs is not consistently preserved in current unified models. The research highlights a critical gap in multimodal AI safety.
Key facts
- UMLMs integrate understanding and generation capabilities in a single architecture.
- Safety implications of unified architectures remain underexplored.
- Existing safety benchmarks focus on isolated understanding or generation tasks.
- Uni-SafeBench features a taxonomy of six major safety categories across seven task types.
- Uni-Judger decouples contextual safety from intrinsic safety.
- Evaluations show original safety alignment of underlying LLM is not consistently preserved.
- The paper is available on arXiv (2604.00547).
- The benchmark aims to evaluate holistic safety of UMLMs under a unified framework.
Entities
—