AutomataGPT
Decoder-only transformer models for one-step forecasting and inverse ruleset inference in 2D cellular automata.
AutomataGPT trains decoder-only transformer models on simulated two-dimensional binary deterministic cellular automata on toroidal grids. The paper tests two problems: predicting the next global state from an initial condition and rules matrix, and inferring a rules matrix from an initial/next-state pair.
Paper takeaways
- The largest model setting uses around 1 million simulated trajectories spanning 100 distinct two-dimensional rules.
- For previously unseen rules from the same family, the model reaches 98.5% perfect one-step forecasts.
- Inverse experiments reconstruct governing rules with up to 96% functional application accuracy and 82% exact rule-matrix match.
Results that matter
- Training over wider regions of rule space improves both forward forecasting and inverse ruleset inference.
- Some inferred rules are degenerate alternatives: they may not exactly match ground truth while still reproducing the observed dynamics.
- The work frames transformer-based rule inference as a step toward data-efficient cellular-automata surrogates for scientific systems.