Google-led researchers released StreamVibe, a video agent designed to assist users by speaking up only when needed. The team compiled a benchmark of 4,100 examples from over 183 hours of everyday life and software recordings. StreamVibe-VL is a model based on Alibaba’s Qwen3.5-9B that decides every second whether to speak.
StreamVibe-VL improves response timing and quality by 10.7 percent over other models. The system uses a memory system to track pending requests and completed steps. The memory system boosts proactive interaction by 12.3 percent and reduces interruptions. Google’s proprietary Gemini 3.8 Flash scores higher than StreamVibe-VL in several tests.
Performance drops sharply when models process entire videos and build on their own earlier responses. A key content metric for responses drops by up to 61.5 percent when errors compound. Models struggle to pick the right moment in videos longer than five minutes. This research highlights the difficulty of maintaining context over long video sequences.
