A new arXiv preprint introduces DIVE, a benchmark designed to evaluate full-duplex interaction with visual engagement in continuous streams. The work responds to recent advances in multimodal large language models (MLLMs) that have substantially improved spoken interaction and video understanding.
According to the abstract, existing online and streaming benchmarks primarily assess causal multimodal perception. DIVE appears to shift focus toward full-duplex interaction, where a model must simultaneously perceive and respond while maintaining visual engagement over time.
The benchmark's exact methodology and results are not detailed in the abstract, but the stated aim is to address a gap in current evaluation. As MLLMs become more interactive, benchmarks like DIVE could help measure progress in real-time, visually engaged dialogue.