Training Language Models via Neural Cellular Automata

TL;DR

This paper introduces neural cellular automata (NCA) to generate synthetic data for pre-training large language models, improving efficiency and performance while reducing reliance on natural language data.

Contribution

It demonstrates that pre-training on synthetic NCA data can outperform natural language pre-training in certain settings, offering a controllable and scalable alternative.

Findings

Pre-training on 164M NCA tokens improves downstream performance by up to 6%.

NCA pre-training accelerates convergence by up to 1.6x.

Synthetic NCA data can outperform large natural language datasets in pre-training efficiency.

Abstract

Pre-training is crucial for large language models (LLMs), as it is when most representations and capabilities are acquired. However, natural language pre-training has problems: high-quality text is finite, it contains human biases, and it entangles knowledge with reasoning. This raises a fundamental question: is natural language the only path to intelligence? We propose using neural cellular automata (NCA) to generate synthetic, non-linguistic data for pre-pre-training LLMs--training on synthetic-then-natural language. NCA data exhibits rich spatiotemporal structure and statistics resembling natural language while being controllable and cheap to generate at scale. We find that pre-pre-training on only 164M NCA tokens improves downstream language modeling by up to 6% and accelerates convergence by up to 1.6x. Surprisingly, this even outperforms pre-pre-training on 1.6B tokens of natural language from Common Crawl with more compute. These gains also transfer to reasoning benchmarks, including GSM8K, HumanEval, and BigBench-Lite. Investigating what drives transfer, we find that attention layers are the most transferable, and that optimal NCA complexity varies by domain: code benefits from simpler dynamics, while math and web text favor more complex ones. These results enable systematic tuning of the synthetic distribution to target domains. More broadly, our work opens a path toward more efficient models with fully synthetic pre-training.