Executable world representation
A physical scene written as code a simulator can run, rather than as pixels, latent features or a sentence. Abbreviated EWR in code-as-world-report, which introduces the term. The bet is that if you can execute the representation you can also test it, edit it, and read exact quantities out of it. A pixel record and an embedding let you do none of those.
The three components
An EWR is a triple (C, E, A):
- Physical composition (C) — the objects, their geometry and mass, their relations, the coordinate system and its units. Anything that participates in support or collision lives here.
- Dynamic evolution (E) — initial states, forces, temporal change, key events, simulation duration. Executing this expands the composition into a full state trajectory with contacts, collisions and velocity changes recorded as they happen.
- Visual appearance (A) — cameras, backgrounds, materials, lighting, frame rate, resolution, rendering configuration. Explicitly not part of the physics: appearance determines how the trajectory is shown, never what it is.
That last split is the load-bearing one. Backgrounds are appearance; a wall you bounce off is composition. Getting the boundary right is what lets the same physical world be re-rendered from a new camera, or handed to a video generator for photorealism, without touching the mechanism.
Why code, per the project's four claims
Abstraction: entities, states, events and relations instead of raw observations, so task-relevant structure separates from incidental appearance. Compositionality: parts can be decomposed and recombined, so you can reason over one object while holding the rest fixed. Executability: the representation is tested against evidence by running it and comparing consequences, rather than by duplicating pixels. Controllability: every parameter is inspectable and editable, which buys interventions and counterfactual rollouts. Change the bowling ball's initial velocity, re-simulate, and the trajectory differs.
What it is not
Not a world model in the video-prediction sense. Nothing here predicts the next frame, in
latent space or in pixels. The nearest neighbours in this corpus are the world-model agent role catalogued in
awesome-agentic-time-series and the perceptual-map sense the phrase carries in
perceptual-learning (../psychology-wiki), and it is neither. An EWR is a program with named
physical parameters; its output is a trajectory, and the trajectory is the prediction.
The vocabulary overlap is worth guarding. "World representation" here means scene as MuJoCo-ready program, and nothing about representation learning, embeddings or the internal states studied at induction-head.
The standing problem
An EWR is verified by whether it renders like the evidence, which is not the same as whether it is right. code-as-world-report §4.4 concedes exactly this: outside the simulator's modelling scope the search can settle on a locally plausible world that reproduces the video without recovering the mechanism. Executable does not mean true, and the paper does not offer a way to tell the two apart.
Related
- agentic-discovery-loop — how an EWR is recovered from an observation
- code-as-world-vl — the models trained on EWR-derived labels
- training-data-quality — what an exact simulator label is worth