Scientists Found the Randomness Trick That Makes Self-Growing AI More Reliable

Imagine a single pixel that knows how to grow into a picture of a lizard — and if you cut off its tail, it grows back. That's the idea behind neural cellular automata (NCAs): AI systems where each tiny cell follows simple rules and, together, they build and repair patterns. Researchers at arXiv wanted to answer a surprisingly messy question: when should you add randomness to this process? During learning, or when the system actually runs?

Their answer is practical. Adding randomness during training — so the AI doesn't see the same state twice — made learning work 10 times out of 10, compared with only 3 out of 10 when training was perfectly orderly. But here's the interesting part: once the AI had learned, it no longer needed the randomness. All ten of those models held onto their target pattern for 4,096 steps when tested with zero randomness.

They also found that how you train the system shapes how it behaves later. Of 30 models that all passed the same basic test of recreating a pattern, eight of the ten trained only to 'grow' drifted off-target after 4,096 steps. Models trained to persist or regenerate stayed on target — and the ones trained to regenerate were the only ones that could recover from damage.

The bigger lesson: 'stability' in AI isn't one single thing. It's really three separate questions — does the training work reliably, does the system need randomness at runtime, and can it handle being damaged. Researchers had been treating them as one. Untangling them could lead to AI that grows, heals, and lasts — useful for designing self-repairing materials, robots, or biological simulations. This is early lab work, though, not a product you can buy.