World Models

Internal models of environments that support prediction, counterfactuals, and long-horizon decision making.

World models are compact representations that an agent can roll forward to anticipate consequences. This track covers learned simulators, latent dynamics, generative environment models, and structured world models that combine geometry, objects, and physics.

We are especially interested in models that remain stable over long horizons, that can be queried for counterfactual “what if” questions, and that expose uncertainty when the environment leaves the training distribution.

Papers should explain what is being modeled, what is left unmodeled, and how the world model is used for planning, data generation, or scientific insight.

All papers in this track must follow the PIJ Ethics Code.