Google StreamVibe benchmarks video agents on evolving intent
TL;DR
- StreamVibe-Bench provides 4.1K evaluation samples and 85K training samples spanning explicit, implicit, and dynamic user intents.
- StreamVibe-VL improves response timing and answer quality by 10.7% over baselines while preserving general video understanding.
- Adding the StreamVibe-Harness hierarchical memory contributes a further 12.3% gain in proactive interaction over sustained use.
Google researchers posted on Hugging Face a benchmark called StreamVibe-Bench that scores proactive video agents on three types of user intent: explicit, implicit, and dynamic. The benchmark ships with 4.1K evaluation samples and 85K training samples. The paper describes the implicit case with an example as flat as "where should I click next?" while a user navigates a website.
The team's streaming model, StreamVibe-VL, "preserves strong general video understanding while improving response timing and answer quality by 10.7% over other baselines," per the abstract. A separate component called StreamVibe-Harness uses a hierarchical memory to maintain user intents as persistent states, and the authors report a further "12.3% gain in proactive interaction" from it.
The paper is candid about where proactive systems still break: "existing proactive models remain limited in understanding user intent and deciding when to respond, especially under implicit and dynamic intents." It lands amid a dense run of agent-stack coverage we have been tracking across the past three months.
Originally reported by huggingface.co
Read the original article →Original headline: Google's StreamVibe Benchmarks Proactive Video Agents That Track Evolving User Intent, Lifts Response Quality 10.7%