Cancer-inspired AI: Greulich et al. evolve neural agents via population ecology

A new paper from Philip Greulich, Michael Levin, and Rosalia Moreddu, posted on arXiv in March 2025 and revised in June 2026, proposes a cancer-inspired (oncomorphic) approach to distributed AI. The framework models populations of neural agents that can replicate, undergo architectural mutation, migrate between task environments, and compete for a finite shared resource pool—all without centralized coordination. Each agent trains only on its local task, so population ecology determines which neural network architectures persist and dominate.

In controlled synthetic nonlinear classification experiments, the researchers observed that stronger selection pressure improved the local accuracy of surviving agents. Architectural mutation played a context-dependent role: diverse initial populations performed best with low mutation rates, whereas clonal large-architecture populations benefited from mutation-generated variation. The system also demonstrated increased multi-task competence at the end of runs—surviving agents could perform well on all environments without additional training. Time-resolved entropy and dominance analyses showed concentration toward successful architectures, while finite training cycles kept agents in a non-asymptotic learning regime, consistent with efficient redistribution of limited computational resources.