StreamVibe proposes video AI that adapts to changing user goals

The StreamVibe researchers describe a video AI model designed to infer unspoken intent from continuous observations, alongside a memory system that tracks changing goals. Their benchmark includes 4,100 evaluation samples and 85,000 training samples. In their experiments, they report a 10.7% improvement in response timing and answer quality over baselines, plus a 12.3% gain in proactive interaction performance with the memory system.