Training Language Models via Neural Cellular Automata

arXiv cs.AI / 3/12/2026

📰 NewsIdeas & Deep AnalysisModels & Research

共有:

Key Points

The authors propose using neural cellular automata (NCA) to generate synthetic, non-linguistic data for pre-training LLMs, enabling a synthetic-then-natural-language pretraining approach.
NCA data exhibit rich spatiotemporal structure similar to natural language while remaining controllable and cheap to produce at scale.
Pre-training on just 164M NCA tokens improves downstream language modeling by up to 6% and accelerates convergence by up to 1.6x, even outperforming pre-training on 1.6B natural-language tokens in some settings.
The gains transfer to reasoning benchmarks (GSM8K, HumanEval, BigBench-Lite), with findings that attention layers are highly transferable and that optimal NCA complexity varies by domain, enabling targeted synthetic distributions.

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.