PhiZero: teaching video AI to think in physics before it renders

Video generation models have a physics problem. Give one a starting frame and it can produce a perfectly plausible sequence in which a ball falls up, a cup passes through a table, or an arm bends the wrong way. A new paper from CASIA argues the flaw is structural: today's models predict pixels directly, and nothing in that pipeline ever represents the rules of the world being depicted.

The proposed fix is unusual. PhiZero does not try to make the visual predictor smarter. Instead it steps back from pixels entirely, first translating the expected evolution of a scene into what the authors call "physical language," a compact discrete representation of world-state transitions, then rendering that description into video. Reason first, paint second.

The trouble with predicting pixels

Existing physical world models, the paper notes, typically operate in pixel space: they take a frame or a prompt and predict future frames, leaving "the underlying world dynamics implicit within high-dimensional visual predictors," as the abstract puts it. That works for short, flashy clips. It does not obviously work for scenes that need consistent mechanics, because the model is never required to expose what it thinks is happening, only what it thinks it looks like.

PhiZero's premise is that humans do something different. We abstract predictive structure from visual experience and organize it in roughly language-like statements about how things change. The paper treats that observation as a design constraint rather than a philosophical one. The same wager has paid off elsewhere: on temporal reasoning tasks, neuro-symbolic models beat raw neural ones.

Physical language, learned in the wild

The "language" is not a handcrafted ontology. PhiZero learns it from in-the-wild videos through self-supervision, with no labels or annotations, as a compact discrete representation of state changes. What exactly a unit in that language encodes is not detailed in the preprint's abstract. The intent is clear enough: force the model to commit to a description of what will change before it is allowed to draw anything.

Reason, then render

The architecture follows a two-stage paradigm the authors call reason-then-render. Stage one infers the future evolution of the world as a sequence in physical language. Stage two renders those inferred transitions into actual video frames. That division of labor is what distinguishes PhiZero from approaches that bake dynamics into a high-dimensional predictor and hope they re-emerge at inference time.

What the paper claims

The authors report extensive experiments across generation and understanding benchmarks validating, in the words of the abstract, "physically coherent world evolution." They also claim potential for three things that matter beyond benchmark scores: realistic and interactive world modeling, fine-grained action-conditioned simulation, and zero-shot motion transfer, meaning a motion pattern learned on one object can be applied to another without retraining.

Those ambitions already have company. Nvidia's Cosmos-H-Dreams runs at 160 fps for interactive surgical simulation, and CHARM makes a comparable structured-context bet on zero-shot graph transfer.

The preprint was submitted to arXiv on July 30, 2026, and the project page is available at phi-zero.github.io. The paper has picked up 155 upvotes on Hugging Face at the time of writing. That is modest community interest, not a verdict on the science.

Unanswered questions

The open questions are real, and the paper invites them. What does "coherent" mean operationally in the benchmarks? How does a discrete representation scale when scenes get long and crowded? And is "language" the right metaphor, or an elaborate discrete bottleneck with a flattering name? The abstract settles none of that. The stronger claim, that explicit reasoning beats direct pixel prediction, will need third-party benchmarks and a reproduction of the reason-then-render pipeline on someone else's data before it means much. Physical-action benchmarks have humbled stronger systems before: SceneActBench found even the best VLMs fail at 3D action.

None of which makes the direction less interesting. World models are being pulled two ways at once: toward bigger context and longer prediction horizons, or toward more structured internal representations. PhiZero is a bet on the second one, that a model able to say what will happen before it shows what will happen will end up understanding the world better than one that merely renders it.