Evaluating streaming systems means asking how well a generated world keeps running, not just how a video looks.
StreamAV-Bench introduces two complementary 180s evaluation tracks:
• Progressive Track: Generates 180s of audio-video from a single global prompt, evaluating instruction adherence and long-horizon stability, including temporal drift as generation unfolds.
• Interactive Track: Establishes an initial world and introduces five runtime updates at 30-second intervals, evaluating dynamic switching, response latency, and state retention and reuse.
