New MLLM safety guardrail reduces over-refusal by predicting unsafe outputs

Researchers have developed a new output-aware safety guardrail system for multimodal large language models (MLLMs) that aims to reduce over-refusal while maintaining safety. Unlike existing input-side guardrails that can indiscriminately block queries, this new method predicts whether a model's forthcoming generation will be unsafe by analyzing its hidden state. This allows for precise intervention only when a harmful output is imminent, preserving the model's utility and built-in safety capabilities. AI

IMPACT This new safety mechanism could lead to more useful and less frustrating interactions with MLLMs by reducing unnecessary refusals.

RANK_REASON This is a research paper detailing a new method for improving safety guardrails in MLLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• CatalyzeX Code Finder for Papers
• Connected Papers
• CORE Recommender
• DagsHub
• Gotit.pub
• Hugging Face
• IArxiv Recommender
• Influence Flower
• Litmaps
• MLLMs
• ScienceCast
• scite Smart Citations

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →