Zhenqi Jia

LLMs can achieve state-of-the-art audio-visual segmentation without any training by using a multi-agent system that explicitly reasons about expression difficulty and validates segmentation results.