From Generation to Simulation: How Far Are World Models from Being True Simulators?

This capability audit asks whether modern generative world models are true simulators or merely convincing generators. It evaluates 200 representative works against eight simulator requirements: asset construction, physics, interaction, controllability, stability, state feedback, diversity, and evaluation. The field’s strongest areas are controllability and interaction, illustrated by systems such as DeepMind’s Genie, NVIDIA’s Cosmos platform, and interactive video models related to Sora; however, these systems generally sample plausible visual futures rather than enforce invariant physical laws. The most serious deficiency is state feedback: across 163 implementation papers, only 6 expose a runtime interface for queryable entity states or physical parameters, while 45 provide sensor-level outputs beyond RGB. This limits direct use by controllers, reinforcement-learning agents, and safety-critical evaluators. The audit also finds that controllability receives 62.5% coverage, whereas physics-engine functionality receives only 17%, revealing a major research imbalance. Joint-embedding models such as V-JEPA 2 offer an alternative by planning in latent space: they require 16 seconds per action compared with about 4 minutes for comparable pixel-generation planning, a 15 times speed advantage, but their representations are difficult to interpret. Techniques including autoregressive-diffusion distillation, latent-action learning, differentiable physics, explicit 3D memory, and causal-intervention benchmarks are narrowing the gap. The conclusion is conditional: world models can replace simulators in restricted scenarios, but formal physics guarantees, persistent long-horizon state, structured feedback, reproducibility, and downstream-utility evaluation remain unresolved. The proposed roadmap prioritizes formalized physics, unified action interfaces, first-class state feedback, stability, utility-based evaluation, and hybrid architectures.