WorldFoundry catalogs 270 world models and 74 benchmarks

WorldFoundry, released this week on arXiv, bundles 270 model entries, 74 benchmarks, and 62 registered metric entries for generative world models under a single open-source harness. Sixteen authors, including Ming-Hsuan Yang, Felix Heide, and lead author Juanxi Tian, argue that visual plausibility is not enough. They define a world model as 'a predictive dynamical system that maintains a persistent, history-conditioned representation of an environment, updates that representation in response to actions or interventions, and generates temporally coherent trajectories.'

The paper covers video, 3D and 4D, embodied action, and interactive systems, with modality-specific adapters that 'preserve native inputs, execution semantics, output representations, and evaluation protocols.' Five layers run through the intelligence-oriented framework: perceptual grounding, persistent state modeling, controllable transformation, dynamics-aware prediction, and closed-loop interaction.

Readiness and execution evidence are tracked separately from evaluative claims. 'Explicit manifests, durable artifact contracts, and evidence-aware scorecards separate execution readiness from evaluative claims,' the authors write, so a model being runnable inside the harness is not the same as a model having been scored. The paper's closing line doubles as its pitch: 'advancing the field from perceptual synthesis toward measurable world intelligence.'