WorldFoundry: Towards Unified Infrastructure for World Intelligence

WorldFoundry. Towards Unified Infrastructure for World Intelligence. By Juanxi Tian and fifteen others, from OpenEnvision. Published on the seventh of November.

Visual generative models are increasingly blurring the line between synthesizing observations and simulating environments. The authors argue that looking plausible is not the same as modelling a world. A compelling output may still lack persistent state, physical dynamics, or causally consistent responses when something interacts with it. WorldFoundry is their answer on the practical side: shared infrastructure for running and evaluating such systems reproducibly, across video, 3D and 4D, embodied action, and interactive world systems.

It begins from a definition. A generative world model, in their terms, is a predictive dynamical system that keeps a persistent, history-conditioned representation of an environment, updates that representation in response to actions or interventions, and generates temporally coherent trajectories reflecting outcomes, under both observed and counterfactual interventions. From that definition comes a framework oriented toward intelligence, with five parts: perceptual grounding, persistent state modelling, controllable transformation, dynamics-aware prediction, and closed-loop interaction.

The infrastructure itself unifies the stages of the work. Discovering models and benchmarks. Preparing runtimes and assets. Orchestrating execution. Managing artifacts. Visualizing interactively. Running underneath all of it are modality-specific adapters, which preserve each system's native inputs, execution semantics, output representations and evaluation protocols, so that unification does not flatten what makes a video model different from an embodied one.

Trust rests on a separation between whether something ran and what it is claimed to show. Explicit manifests, durable artifact contracts, and scorecards that are aware of their evidence keep execution readiness apart from evaluative claims. Each claim can then be traced back to the artifacts supporting it. The catalog holds two hundred seventy model entries, seventy four benchmarks, and sixty-two registered metric entries, with readiness and execution evidence recorded separately for each.

The intended users are those comparing world systems that do not share a format: a video generator against a 3D or 4D model, an embodied policy against an interactive environment. For them, WorldFoundry offers reproducible comparison of persistent, controllable dynamics across heterogeneous systems.

The ambition stated at the end is a shift of emphasis for the field, away from perceptual synthesis and toward world intelligence that can be measured. Here the definition came first, the framework followed from it, and the infrastructure was built so that every number it reports can be traced to something that actually ran.