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

TL;DR - This survey evaluates whether generative world models can replace traditional simulators by mapping 200 works across eight core simulation capabilities. Current models can support interaction and controllability in specific settings, but still lack reliable physics, structured state feedback, and reproducible long-horizon behavior.

• Compares three technical routes: latent dynamics, video generation, and joint-embedding prediction.
• Uses traditional simulators as an external benchmark across asset construction, physics, interaction, controllability, stability, state feedback, diversity, and evaluation.
• State feedback is the largest cross-route gap: only 6 of 163 implementation papers provide runtime access to entity states or physical parameters.
• Priorities include formalized physics, unified action interfaces, first-class state feedback, long-horizon stability, utility-based evaluation, and hybrid approaches.