Paper Explained: Training Object Permanence in World Models — Teaching Video Models That Hidden Things Still Exist

Video generators make hidden objects vanish and let balls pass through walls. We walk through WROP, a 150-task Blender benchmark for object permanence and solidity, the 16B PWM-WROP model fine-tuned on it, and how to read the human-preference evaluation behind the results.

Training Object Permanence in World Models

What this paper asks

This article covers "Training Object Permanence in World Models" (arXiv:2609.28654, published September 23, 2026).

Here is the abstract in plain terms. Object permanence (things keep existing when you can't see them) and solidity (solid things don't pass through each other) are hallmarks of human cognition. Video generation models have started to show reasoning-like abilities, which makes them candidates for human-like physical intelligence. But do they actually have object permanence? And if not, can a dataset inspired by core cognition teach it to them? The authors build WROP (World Reasoning with Object Permanence), a data infrastructure of 150 hand-designed tasks split into six cognitive categories. Blender generators randomize speed, lighting, camera angle and other nuisance factors while keeping each task's cognitive structure intact, producing 10,000+ samples per task. They release a 1.5M-sample training corpus and a 300-question exam, and evaluate 14 video models on it. Their own 16B model, PWM-WROP, ranks first among continuation models and third overall in a blind pairwise Elo study. Data, exam, scores, weights, and a training stack for AWS Trainium2 are all released.

An analogy: video models that don't get peekaboo

Peekaboo works on a baby because they understand that your face is still there behind your hands. The paper cites developmental research showing that infants represent hidden objects by around 3.5 months and react to solidity violations within their first half-year (§1).

State-of-the-art video generators, meanwhile, produce gorgeous footage in which a ball enters a tunnel and exits from a different lane, or a ball larger than a hole sails straight through the board. The paper frames these as "structurally upstream" failures (§1): a model that permits interpenetration can't render a valid collision or fall, and any causal reasoning stacked on top inherits the error.

How WROP is built

The tasks split along two axes into six families (§3.1):

• Object permanence (OP): OP-1 a moving object passes behind an occluder and re-emerges on the far side / OP-2 a row of static objects is hidden by a moving screen and must reappear unchanged / OP-3 an object inside a container must follow the container as it moves or gets swapped
• Object solidity (OS): OS-1 a barrier whose aperture decides whether an object can pass / OS-2 an object falls when its support is removed / OS-3 objects neither vanish nor merge in a collision

There are 150 generators (90 OP, 60 OS). Each sample is a 120-frame clip at 1280×720 and 24 fps, cut at frame 60, right where the key physical event begins (§3.2). The model sees the first 60 frames and must generate the second half, which contains the event and its consequence.

The central design move is splitting parameters into two kinds (§3.1):

• Structural parameters: object count, geometry, trajectory, occlusion layout, aperture size, contact timing. These define the challenge and are varied systematically, so that memorizing a fixed outcome won't work
• Surface parameters: color, material, lighting, camera viewpoint. These scatter the appearance without changing the physics, so that visual cues can't stand in for physical reasoning

Take "Marked_Boxes_Swap": two labeled boxes cover different objects and then trade places (§3.3). The final screen position alone doesn't determine the answer; you have to track box identity. Also note that motion comes from hand-authored Blender keyframes, not a physics engine (§3.2). That lets occlusion and contact land on exact frames, and the paper states plainly that physical correctness is the scene author's responsibility.

Three interface classes, 14 models

The 14 models fall into three classes based on how they treat the input clip (§4.2):

• True continuation: treat the input as a prefix and generate what comes next. PWM-WROP, Grok Imagine (video extend), and two others
• Reference-to-video: use the input as a visual reference and regenerate the whole event on their own timeline. Wan 3.0 Prime, MiniMax H3, Seedance 2.5
• Edit/transfer: repaint the input span frame by frame. Because the output occupies the same time window as the input, these models fundamentally can't show what happens after the occlusion boundary. Seven models

Every model receives only the input video and the prompt, verbatim; the target is never shown (§4.3).