The ball vanishes. Does the video understand physics?
arXiv:2609.28654

Scripted podcast dialogue; AI-generated voices.

HOST: I have seen gorgeous AI-generated videos where a ball disappears behind a box and never returns. Sometimes it comes out somewhere impossible. Is that just a visual glitch?

GUEST: It is deeper than one bad frame. When the ball is hidden, it should keep existing: object permanence. When it meets a solid wall, it should not pass through: solidity. WROP builds targeted tests of whether video models preserve those constraints when they continue a scene.

HOST: What does one of those tests look like?

GUEST: A model sees the start of a clip in which a sphere rolls towards a panel. It has to generate what happens next. The sphere may be hidden for a moment, but should re-emerge at a position consistent with its path. Other clips involve falling onto a surface, striking an obstacle or losing support. The decisive event occurs after the segment the model was shown.

HOST: How do the researchers avoid testing the exact same trick over and over?

GUEST: They build 150 three-dimensional scene generators in Blender. Lighting, camera angle and speed vary while the underlying physical question stays the same. The scenes fall into six families: three about permanence, including occlusion and containers, and three about solidity, including obstruction, support removal and collision. The corpus has 1.5 million training examples and a fixed 300-question exam.

HOST: Could a pixel metric simply grade the videos?

GUEST: Not reliably. A video can preserve colours and backgrounds and still get the crucial event wrong. For their main assessment, people make blind pairwise comparisons between videos. Automated metrics complement that judgement rather than replacing it.

HOST: What happens after targeted training?

GUEST: The authors further train a sixteen-billion-parameter continuation model called PWM-WROP. In human comparisons it ranks first among genuine video-continuation models and third among all fourteen systems compared. That is promising, but the main evaluation contains 361 pairwise judgements from twenty people. Confidence intervals overlap for some ranks, so exact positions are not always clear-cut.

HOST: Why distinguish continuation systems from the rest?

GUEST: Because the interfaces differ. Some systems continue existing footage, while others generate from a reference image or edit a scene. More freedom to regenerate may help on some tests but can make it harder to preserve the identity of an object already shown. The scenes are also synthetic and controlled, and the trained model sees tasks with the same structure as the exam.

HOST: So we should not say it has learned the physics of the whole world?

GUEST: Right. We can say targeted training helps on these defined tests. Open-ended simulations and real robots need more evidence. The paper's practical value is to separate beautiful video from event consistency: when an object is hidden or hits a barrier, watch what actually happens.