Paper review · November 11, 2026
WorldFoundry: Unified inference and evaluation infrastructure for world intelligence
WorldFoundry: Towards Unified Infrastructure for World Intelligence
The authors define generative world models in terms of their functions and build open-source infrastructure that runs and evaluates 270 video, 3D/4D, embodied and interactive models in one place, providing a basis for reproducible comparison.
Problem
A video generation model can produce plausible frames yet lack persistent state, physical dynamics and causal responses to interventions. However, there is no shared definition of a world model, benchmarks focus mainly on perceptual quality, and workflows vary widely. This makes it hard to tell generative models apart from world models and to evaluate dynamics and controllability reproducibly.
Method
The authors define a generative world model as "a dynamical system that infers and maintains a persistent state conditioned on history and control inputs, updates that state in response to interventions, and generates consistent future trajectories." They split this into three functions: state inference, forward simulation and interactive response. Based on this definition, they built the WorldFoundry infrastructure, which combines model and benchmark discovery, runtime and asset preparation, execution orchestration, artifact management and visualization. Modality-specific adapters preserve each model's original inputs and evaluation protocol. Declarative manifests, artifact contracts and evidence-aware scorecards record "runnable status" separately from "performance claims."
Results
- The current catalog contains 270 models, 74 benchmarks and 62 registered metrics. The integration readiness and execution evidence of each entry are recorded separately.
- The authors report that they verified every model runner by running it end to end with a documented reference workflow.
- In a comparison table, the authors show that it covers video, 3D/4D, interactive and embodied models more broadly than existing world model infrastructure, toolkits and evaluation tools.
Limitations
The provided text only covers the introduction and related work, so actual performance comparisons between models cannot be checked. Verification is at the level of "runnability in the tested configurations" and does not guarantee that each model's performance is reproduced. The world model definition is a functional framework proposed by the authors, not an agreed standard.
Why it matters
It proposes that world models should be evaluated on state persistence and closed-loop interaction rather than on "good-looking video." Providing open-source shared infrastructure for running heterogeneous models under the same conditions is also useful in practice.
Key terms
- Generative world model
- A model that continuously maintains the state of an environment, updates that state in response to actions or interventions, and generates the future. The authors judge such models by whether this loop is closed (closure), not by visual plausibility.
- Closed-loop interaction
- A repeating structure in which a model's output feeds back into its next input and state. An intervention should leave a trace in the state and constrain later predictions.
- Evidence-aware scorecard
- An evaluation sheet that records separately the capabilities declared as supported, whether they actually run, and verified performance. Each claim can be traced to the artifacts that support it.
Original
Authors: Juanxi Tian, Jiadong Hou, Baiang Li, Xin Lin, Zhengbo Jiao, Weijia Zhang, Yeying Jin, Truong Nguyen, et al. · Posted 2026-11-07 · License CC BY 4.0 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.