Google Team Introduces StreamVibe Video Agent; Tests Show Response Quality Drops Up to 61.5% in Continuous Long-Video Operation
A Google-led research team released StreamVibe, a video agent, and its ProactiveBench benchmark, revealing that current models struggle to time interventions accurately in long videos due to compounding errors.
A research team comprising researchers from Google, Stanford University, Oxford University, and other institutions has introduced StreamVibe, a video agent designed to understand vague user intentions and decide when to offer help. According to The Decoder, the team also released ProactiveBench, a benchmark containing approximately 4,100 examples covering clear instructions, vague goals, and requests that change over time.
The tests show that while StreamVibe-VL (reportedly trained based on Alibaba's Qwen3.5-9B) improved response timing and quality by 10.7 percent over baseline models, all tested systems performed significantly worse when processing continuous videos longer than five minutes. As video length increased, models struggled to correctly time their responses, and key content metrics dropped by up to 61.5 percent due to compounding errors.
Researchers noted that existing video assistants are typically built for clear-cut commands, making it difficult to pick up on unspoken intentions or track completed steps, which leads to repeated advice or delayed responses. StreamVibe attempts to address this by distinguishing between explicit, implicit, and changing user intentions, but its core finding reveals a fundamental limitation in current AI for continuous interaction: performance plummets as errors compound when models must build upon their own earlier responses.