The timing problem is what kills most video assistants for me. A helper that t… · @harden_369741 · TwKit

@harden_369741

The timing problem is what kills most video assistants for me. A helper that talks at the wrong moment is worse than one that stays quiet. A team spanning Google and several universities just published StreamVibe, a system built to solve exactly that. Instead of waiting for a command, StreamVibe-VL watches a live video feed and decides every second whether to stay quiet or speak. It is built on Qwen3.5-9B, and it tracks three kinds of intent: what you asked for, what you meant without saying it, and what changed mid task. On their StreamVibe-Bench, 4.1K evaluation samples against 85K training samples, the model improved response timing and answer quality by 10.7% over existing baselines. A separate memory harness that tracks pending requests and completed steps added another 12.3% on sustained interaction. I find the second number more interesting than the first. Most demos nail the first helpful line and fall apart on longer tasks, because the assistant forgets what you were doing. A system that remembers the task state is the difference between a narrator and a coworker. The bar for proactive agents should be silence by default, speech only when it moves the task forward. StreamVibe gets that bar right on paper. https://arxiv.org/abs/2610.34662 #AI #Agents

曝光 13 · 评论 0 · 点赞 0 · 书签 0