AI models that claim to simulate the world keep getting the details right and the big picture wrong.
A new arXiv paper tests conventional world models (transformer and CNN based autoregressive models, plus diffusion models) on cellular automata like Conway's Game of Life. A CNN predicted 96.3% of individual cells correctly, but completed a fully correct rollout only 18.9% of the time. A joint diffusion model predicted most cells correctly too, yet completed zero correct rollouts. The researchers trace this to three specific failures: the models don't exactly capture spatial locality, temporal locality, or temporal stability in how information moves through a sequence of frames.
The fixes are notably small. Adding two-dimensional rotary position encoding took a transformer from 39.1% to 100% rollout accuracy on the Game of Life. Giving each token access to its cell's previous-frame neighborhood took accuracy on unseen rules from 25.8% to 99.9%. A technique called causal freezing took the diffusion model from 42.2% to 99.9%. None of these changes touch the model's architecture; they only change how information flows through it.
That distinction matters beyond cellular automata. World models are the backbone of robotics planning and some bids at general-purpose AI reasoning, and claims that such models have learned the underlying rules of the world are common in that pitch. This research suggests many such models match surface statistics, not the underlying rules, even in a toy domain with fully known, deterministic logic.
Cellular automata are about as easy as world simulation gets: fixed grid, fixed rules, full visibility. If a model can't reliably complete that, its grasp on messier real-world physics is worth doubting too.