Stability in AI beyond reliance on pool-training

Determine effective methods to achieve stability in artificial intelligence systems without relying solely on pool-training, identifying mechanisms or training procedures that ensure robust behavior over time across distributed models such as neural cellular automata.

Background

The paper introduces EngramNCA, a neural cellular automaton with public and private cell states designed to support stable, coexisting morphologies and memory transfer. The authors argue that private, cell-internal memory can yield more stable distributed behaviors than traditional approaches relying exclusively on public states.

Within the discussion, the authors explicitly note that stability remains an open problem in AI and is typically mitigated by pool-training. They suggest exploring EngramNCA’s private memory mechanisms as a potential path forward when pool-training is not applicable, motivating the need for principled stability methods that do not depend solely on pool-based training strategies.