Project

LifeGPT

Topology-agnostic GPT for cellular automata; published in npj Artificial Intelligence.

Read the paper →
LifeGPT paper figure showing topology-agnostic architecture
Figure: LifeGPT architecture and training setup for topology-agnostic rollout and rule inference across multiple lattice types (from the LifeGPT paper).

LifeGPT learns CA dynamics with a topology-agnostic tokenizer and GPT-style attention, enabling one model to reason over square, hexagonal, and triangular lattices without architecture changes. The paper demonstrates strong long-horizon rollouts and rule generalization on diverse 2D automata.

Paper takeaways

  • Unified tokenization handles variable neighborhoods, letting one GPT capture CA rules across geometries.
  • Autoregressive training stabilizes long rollouts and preserves fine pattern details in chaotic regimes.
  • In-context rule inference lets the model clone unknown automata from a handful of observed steps.

Results that matter

  • Consistent accuracy on heterogeneous benchmark sets spanning Life-like, Lenia-inspired, and custom rules.
  • Topology transfer: models trained on one grid adapt to others with minimal degradation.
  • Ablations show the importance of positional encoding choices for stable pattern propagation.