Search papers, labs, and topics across Lattice.
The paper explores pre-training LLMs on synthetic data generated by Neural Cellular Automata (NCA) before natural language pre-training. They 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, outperforming pre-pre-training on 1.6B tokens of natural language. Analysis reveals that attention layers are the most transferable and that optimal NCA complexity varies by domain, suggesting the possibility of tuning synthetic distributions to target specific skills.
Forget scaling laws: pre-training LLMs on just 164M tokens of synthetic, non-linguistic data can outperform pre-training on 1.6B tokens of Common Crawl, opening a new path to efficient model training.
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.
MIT CSAIL