WorldFoundry: Unified infrastructure for generative world models

Visual generative models can synthesize realistic observations, but realistic images alone do not mean a system understands physical dynamics, maintains a persistent state, or responds reliably to interactions, according to the authors (arxiv.org). The authors show and define what separates a basic generative model from a true generative world model, and introduce open-source infrastructure to measure and compare these systems reproducibly (arxiv.org).

The problem

Recent advances in generative modeling are moving from creating individual objects or scenes toward modeling entire worlds that evolve over time (arxiv.org). However, video generators, neural scene representations, and embodied policies often use incompatible representations, runtimes, interfaces, and evaluation protocols (arxiv.org). According to the authors, this fragmentation makes capabilities difficult to compose, results difficult to reproduce, and progress difficult to measure (arxiv.org). Furthermore, visual plausibility is often mistaken for world understanding, even though compelling outputs may lack persistent state, physical dynamics, or causally consistent responses to interactions (arxiv.org). The bottleneck is not purely algorithmic, but infrastructural (arxiv.org).

The idea

The authors formulate a functional abstraction for world intelligence consisting of three tightly coupled components: state inference, forward simulation, and interactive response (arxiv.org). Building on this, the authors formally define a generative world model as a predictive dynamical system that maintains a persistent, history-conditioned environment representation, updates that state in response to actions or interventions, and generates temporally coherent trajectories reflecting observed and counterfactual outcomes (arxiv.org).

To bridge the gap between theory and execution, the authors present WorldFoundry, an open-source infrastructure designed for reproducible inference and evaluation across video, 3D and 4D, embodied action, and interactive world systems (arxiv.org). WorldFoundry unifies model and benchmark discovery, runtime and asset preparation, execution orchestration, artifact management, and interactive visualization (arxiv.org). It uses modality-specific adapters to preserve native inputs, execution semantics, output representations, and evaluation protocols (arxiv.org). Explicit manifests, durable artifact contracts, and evidence-aware scorecards separate execution readiness from evaluative claims to ensure all claims are traceable (arxiv.org).

The results

The current WorldFoundry catalog integrates 270 model entries, 74 benchmarks, and 62 registered metric entries (arxiv.org). The authors report that every model runner has been validated through a documented end-to-end reference workflow, yielding an auditable record of operational readiness under the tested configurations (arxiv.org).

Why it matters

By establishing a shared operational definition and executable infrastructure, WorldFoundry helps shift the field of artificial intelligence away from isolated perceptual synthesis and toward measurable, auditable world intelligence (arxiv.org). It allows researchers to compare persistent, controllable dynamics across heterogeneous systems on a common operational basis (arxiv.org).

Caveats

The authors note that this paper represents a preprint version (arxiv.org), meaning its findings and infrastructure claims have not yet undergone traditional peer review. The current catalog relies on specific registered models and benchmarks, and the operational readiness of each model is tied to the tested configurations within the reference workflows (arxiv.org).