Gen-ViRe

Recent video generation models are emerging as potential world simulators, yet existing evaluations focus primarily on visual fidelity rather than reasoning capability. In this work, we introduce Gen-ViRe, a comprehensive benchmark designed to assess Chain-of-Frames (CoF) reasoning—the ability to solve complex tasks through continuous frame-by-frame visual simulation. By evaluating 7 state-of-the-art commercial models across six cognitive dimensions, spanning 24 tasks, we find that Sora-2 currently achieves the best performance in generative visual reasoning. However, our analysis reveals a critical nuance: despite its overall leadership, even Sora-2 exhibits notable deficits in spatial-temporal reasoning, particularly in the Spatial Obstacle task. This suggests that abstract logical reasoning and physical reality simulation seem to function as distinct capabilities for video generation models. Gen-ViRe establishes the necessary baselines to guide the transition from video generators to genuine world simulators.