Sixteen researchers have built a video assistant that tries to work out what you need before you say it and a test set to measure it. The paper, posted to arXiv on 26 October 2026, names Chengzhi Liu first.
Most current systems wait for a direct request. The authors say people rarely spell out their goals, and those goals shift mid-task, as when someone browsing a website wonders where to click next. Their benchmark, StreamVibe-Bench, holds 4.1K test samples and 85K training samples across stated, unstated, and changing intents. Their model, StreamVibe-VL, watches a live feed and adjusts as the goal moves. A memory layer called StreamVibe-Harness keeps a running record of the latest intent and surroundings.
The authors report that the model beats other baselines by 10.7% on response timing and answer quality, and that adding the memory layer lifts proactive performance a further 12.3%. They also find that existing proactive models struggle most with deciding when to speak.
Those figures come from the authors’ own benchmark, in a preprint. Anyone building a screen-watching assistant now has a shared way to test the hardest part: knowing when to stay quiet.
Reported by arXiv on 26 October 2026.
