Let's start with the conclusion. PWM-WROP is the result of fine-tuning a video-generation model on 1.5 million synthetic videos modeled after cognitive-science experiments. In blind human evaluation, it placed first among models that work the same way — "continuation" models that watch the opening of a clip and generate what comes next — and third overall across all 14 models, including ones that operate on entirely different principles. And it did so from a far lower weight class in terms of resolution than its competitors. What hasn't been established yet is whether this achievement carries over to understanding the physics of real, camera-shot footage. Below is why this result is interesting, and what to read carefully.
What is object permanence? Think of peekaboo. Even when you cover your face with your hands, the baby knows Mom's face hasn't vanished. Believing that something keeps existing even when it's hidden from view — that is "object permanence."
Even a 3.5-month-old infant knows that a ball rolling behind a barrier should reappear on the other side. Developmental psychology calls this belief object permanence. On top of it sits a related intuition: that two solid things can't pass through each other, known as object solidity. The paper treats both as belonging to the very bottom layer of human cognition — what's called "core knowledge" (citing developmental-psychology work such as Baillargeon 1986).
Why does this foundation matter? Because without these intuitions, all the physical reasoning built on top of them collapses. A model that lets objects pass through one another can't properly render a collision, or a scene where something falls once its support is pulled away.
And this is exactly the point where today's video-generation models keep slipping. The picture looks plausible, but the ball passes through the wall, and an object that disappears behind a barrier pops out in the wrong place. Since these models are being eyed as candidates for "world models" (models that mimic how the world works in order to predict the next scene), the weakness stings all the more.
The trouble is that there was no proper test to interrogate this weakness. Prior work mostly stayed in 2D environments, and evaluations went no further than the "image → video" approach: give the model a single still and have it imagine what follows. They couldn't probe the "video → video" (hereafter V2V) approach — giving the model the opening footage and having it continue — that is, the context in which a physical event unfolds from before to after. Nor was there a dedicated benchmark built to measure object identity, physical constraints, and causal consequences specifically.
There's one more trap: who does the grading. Lately it's fashionable to hand grading over to a multimodal language model (a VLM, a language model that also understands images and video) — but the paper flatly argues that these very VLMs lack core knowledge (citing Li et al. 2025, among others) and so aren't qualified to judge object permanence. It's like handing grading to a proctor who doesn't even know the subject matter.
The core idea moves as a pair: WROP, a dataset modeled after cognitive-science experiments, and PWM-WROP, a model that learns physical intuition from that data.
First, WROP (World Reasoning with Object Permanence). The premise is simple: take the cognitive experiments once run on infants, recreate them in 3D, and mass-produce them. The authors hand-designed 150 "generators" in the 3D software Blender. Each generator handles one cognitive task, while randomly shaking peripheral variables — lighting, camera angle, speed, color, material — on every run. The heart of the design is splitting the variables into two kinds.
Drawing from this, they collected more than 10,000 samples per generator, releasing a training corpus totaling 1.5 million (1.5M) samples and a fixed 300-question test (two questions per generator).
The tasks come in six strands. Three on the object-permanence side (occlusion of a moving object, occlusion in a static scene, hiding inside a container) and three on the solidity side (blocking, support removal, collision). The generators break down as 90 for permanence and 60 for solidity.
For example, one task (OP-1, Baillargeonian_Occlusion) goes like this. A ball rolls behind a barrier. Only if the model "believes" the ball hasn't vanished — that it keeps rolling at the same speed even while hidden — can it render the ball reappearing on the other side right on time. The point is to see whether that belief is present. This is different from mechanically extrapolating the motion of the last visible frame.
The evaluation method is tailored to V2V too. The ground-truth clip is cut in half at a key physical event (the moment the ball touches the wall, say). The first half is given to the model as context, and the model has to generate the second half, which contains the event and its consequences. Grading is done not by a VLM but by humans, comparing against a hand-built, physically accurate reference animation.
Now the second axis, PWM-WROP. Using this WROP data, the authors post-trained (further training an already-trained model on specific data) a 16B-parameter (16-billion) continuation model.
On a task like the OP-1 above, a model lacking object permanence gives itself away fast. Behind the barrier the ball quietly disappears, or comes out a different color, or pops out in the wrong spot. A model that has permanence assumes the ball keeps rolling even through the unseen stretch, and brings it back at the physically correct position and timing. To the eye both are equally smooth videos — but they diverge in how they fill in the unseen moments. And that's precisely the difference people pick up on.
The evaluation was a blind test. Evaluators were shown two videos side by side without knowing which model made which, and picked the better one; those choices were aggregated into an Elo score for each model. Elo is a relative score that, like a chess ranking, climbs as you beat opponents. The 14 models compared each work differently: 4 continuation models (generating onward from the end of the input video), 3 reference-to-video models (taking an input video plus a prompt and regenerating the event on their own timeline), and 7 editing models (reworking an existing video). In other words, the difference between continuation and reference-to-video isn't whether the input is an image or a video — it's whether generation continues straight from the end of the input video. PWM-WROP belongs to the continuation group.
Here are the results. PWM-WROP came first among continuation models and third overall out of 14. The only models above it are two commercial reference-to-video systems — and those two are a "statistical tie," so first and second can't be separated. By the figures the introduction cites, PWM-WROP scored an Elo of 1679.5, with the two systems above it at 1723.6.
There's a condition you must read alongside the ranking. PWM-WROP outputs at 320×192, far lower than competitors that render at 720p or 1080p. In other words, it was at a heavy disadvantage in picture quality. Yet on automatic metrics that match resolution to the same conditions and compare against the ground-truth video, PWM-WROP was best on both LPIPS and MS-SSIM. Both are metrics by which a machine automatically measures how closely a generated video resembles the reference video at the pixel level. That said, keep in mind these metrics only tell you how closely the output resembles the ground-truth video — they don't directly tell you whether the model truly understood the "physics." They're also a different yardstick from the Elo ranking above, so it's more honest not to lump them together into a single achievement.
In the end, this model is a case study in how "smart data" can overcome the hardware handicap of low resolution. It gets at a human perception in which the physically-closer-to-correct option feels more "plausible" than a merely good-looking high-definition video.
There are, of course, clear limits.
First, it's third overall, full stop. Even granting the handicap of low resolution, you can't say it surpassed the frontier commercial systems. The authors don't claim that either.
Second, WROP is a synthetic 3D environment built in Blender. Randomly shaking the peripheral variables secures diversity, but whether the physical intuition learned in simulation holds up on real footage is another matter. Real video is a tangle of light reflections, shadows, and unpredictable noise. How well an ability acquired in simulation transfers to reality — this is AI research's chronic "Sim-to-Real (simulation → reality)" gap problem, and this paper's results alone can't establish that the gap has been closed.
Third, humans are at the center of the evaluation. It's a reasonable choice to sidestep the flaws of VLM grading, but relying on human hands carries costs in scalability and reproducibility.
The lesson of this paper doesn't stop at video generation. It maps almost directly onto the AI-agent problem ONTLE wrestles with every day. ONTLE is a project-manager-style service in which AI takes over people's repetitive work, so while the domain is nothing like video, the methodology is shared.
What resonates most is the "is the grader qualified" problem. The paper's central claim is: don't hand grading to a grader who lacks the ability to judge. The same applies when we auto-evaluate an agent's output with an LLM. The first thing to ask is, "Does this grading model actually possess the ability we're trying to measure right now?" Multiplying the number of judges without checking their qualifications just gives you several judges sharing the same blind spot. Having two judges is a safeguard at a different layer. It's like cross-checking a single issue against two newspaper articles of differing editorial slants. Both articles could still be wrong, but you raise your odds of filtering out a bias or factual error present in only one of them. The same principle works in AI evaluation. But this doesn't guarantee "has the ability to grade." The key is not to confuse the two safeguards.
Second, separating structural from surface parameters is a textbook of evaluation design. To tell whether an agent genuinely solved a problem or merely memorized an answer tuned to a particular phrasing of the prompt, you have to throw it variant problems that leave the essence (structure) intact while changing only the surface (phrasing, order, terminology). For instance, when you fix a search-and-answer feature and then write a regression test to confirm the old behavior didn't break, this is exactly the principle you can use.
Third, the view that "if the floor collapses, everything above it collapses." In agents too, if the lowest layer — fact lookup and state assessment — is wrong, the planning and execution above it are wholly contaminated. The message is: verify the floor before you tune the performance of the upper stages.
Yes. The authors say they're releasing the 1.5-million-sample training corpus, the 300-question test, the model answers, scores, and PWM-WROP weights, and even PWM, a pure-PyTorch training stack that runs on AWS Trainium2 chips. For the detailed links, see the arXiv and HuggingFace pages below.