Research / Engine of World

Engine of
World

Building persistent worlds in which intelligence can act, remember, and encounter consequences.

An interactive world is not a static scene. It is a causal system that continues after an action.

Physical state is only one layer. A consequential world also contains beliefs, relationships, roles, institutions, durable memory, and unrealized future branches. These hidden states determine what an action means and which futures remain reachable.

Engine of World makes long-horizon consequence observable.

We study environments that preserve latent state, generate causal event chains, and allow agents to act across extended horizons. The objective is not visual realism alone, but worlds coherent enough for counterfactual reasoning, social learning, and the evaluation of decisions whose effects emerge slowly.

Such worlds become scientific instruments: places where intelligence can form plans, revise them, influence other agents, and bear the consequences of what it sets in motion.

Selected work

MarioLMPersistent and consequential interactive worldsLong-Horizon DeceptionSafety across extended interaction