New framework ARENA automates red-teaming for audio language models

Researchers have developed ARENA, a novel closed-loop framework designed for automated red-teaming of large audio-language models (LALMs). This system addresses the unique safety challenges posed by LALMs, which can exhibit harmful behavior when combining text and audio inputs, even if the text alone is safe. ARENA utilizes a controller trained on a dataset of text-audio interactions, with rewards and feedback provided by MD-Judge and final outcomes assessed by Llama Guard 3. The framework demonstrated significant success on various LALMs, achieving high false discovery and partial success rates on objectives from AdvBench. AI

IMPACT Enhances safety testing for multimodal AI, potentially leading to more robust and secure audio-based AI applications.

RANK_REASON The cluster describes a new research paper detailing a novel framework for red-teaming audio language models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• AdvBench
• ARENA
• Audio Flamingo 3
• GPTAudio
• Hugging Face
• Large Audio-Language Models
• Llama Guard 3
• MD-Judge
• MiMo-Audio
• Qwen2-Audio

• paper
• safety

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