这篇论文用MPMWorlds测试了AI看视频写物理代码的能力，发现代码生成稳但缺位置感知，扩散模型短时准但长期漂移，混合模型效果最好。
MPMWorlds是一个包含95,000个2D仿真视频的基准，覆盖液体、雪、沙子和弹性体。模型观看2.5秒场景后预测后续运动。代码生成方法能保持长期物理稳定性，但无法从帧中读取位置，隐藏坐标后精度骤降。扩散模型可捕获短期几何，但物体随时间消失、运动不真实。混合两种方法的简单门控机制超越单一模型。
AI can now watch a video and write the physics cod…
AI can now watch a video and write the physics code behind it.
Video models are often called world simulators.
A new paper tests whether they actually understand physics. MPMWorlds is a benchmark of 95,000 rendered 2D simulations.
It covers liquids, snow, sand, and elastic bodies. Models watch 2.5 seconds of a scene. Then they predict what happens next. The study compares two approaches: 1. A vision model writes simulation code. 2. A diffusion model predicts raw pixels. Code generation stays physically stable over long horizons.
Objects never vanish or drift mid sequence. But it cannot read positions from frames alone.
Withhold coordinates and its accuracy collapses. Diffusion captures short term geometry straight from video.
Over time, objects fade and motion turns implausible. Feeding it the exact physics config changes nothing.
It relies almost entirely on visual patterns. A simple gate combining both beat either alone.
What else could hybrid world models unlock?