StreamVibe tests when a video agent should respond

Google researchers and their co-authors posted StreamVibe on October 26, a benchmark and a model built for video agents that act on user intent while a task is running.

The paper, StreamVibe: Proactive Video Agents for Evolving User Intent, is on arXiv as 2610.34662, submitted on October 26, 2026. Chengzhi Liu, affiliated with Google and the University of California, Santa Barbara, is the corresponding author. The 16-name author list also carries researchers from Emory, Stanford, Columbia, Oxford and UC Santa Cruz.

StreamVibe comes in three parts. StreamVibe-Bench is the test itself: 4.1K evaluation samples and 85K training samples that span explicit, implicit and dynamic intents. StreamVibe-VL is a streaming proactive model that reads continuous observations and infers the intent behind them. StreamVibe-Harness keeps those intents as persistent states in a hierarchical memory, so an agent tracks the latest intent and the current state of the environment as both change.

In the team's experiments, StreamVibe-VL improves response timing and answer quality by 10.7% over other baselines while preserving its general video understanding. StreamVibe-Harness adds a 12.3% gain in proactive interaction across sustained exchanges.

The paper benchmarks StreamVibe against proprietary models including Google DeepMind's Gemini-3.1-Pro and Gemini-3.8-Flash and OpenAI's GPT-Realtime-2.1, open-source offline models such as Qwen3-VL-8B, and open-source streaming models. The same work used Gemini-3.8-Flash to clean its training data and flag candidate help points.

This is research, not a shipped product. StreamVibe arrives with no pricing, no release date and no creator-facing tool. What it points at is a different kind of video agent: one that watches a session, reads a goal a user never typed, and decides on its own when to speak up. The authors also name the failure they are measuring, since existing proactive models tend to respond too early, too late or inconsistently as a task evolves.

Read the abstract and the full 43-page paper at arxiv.org/abs/2610.34662; the Hugging Face paper page is at huggingface.co/papers/2610.34662.