Self-Architecting Cognitive Ecosystems --- Toward Autonomous Structural Intelligence

This work proposes a shift from fixed neural architectures toward systems capable of shaping their own structure as part of the learning process. Instead of treating layers, experts, and neurons as static components optimized only through weight updates, the paper introduces Self-Architecting Cognitive Ecosystems: models where individual components act as cognitive agents negotiating their own relevance, growth, fusion, or retirement. The framework formalizes architectural agency through four mechanisms: Lifecycle protocols for birth, specialization, merging, and dormancy Coalition dynamics enabling experts to coordinate when complex reasoning demands combined perspectives A compute-bounded resource economy where components petition for capacity based on demonstrated insight A reflective regulation loop that stabilizes exploration, protects long-lived abstractions, and prevents collapse into redundancy or chaos Experiments across synthetic, dynamic, and compositional tasks show that granting a model control over its own topology produces coherent structural evolution, persistent specialisation, compositional reasoning via spontaneous coalitions, and resilience after perturbation or distribution shifts — all under identical compute budgets to static baselines. The system’s behavior resembles ontogenesis: the shaping of a cognitive identity over time, driven by uncertainty, novelty, and the preservation of meaningful abstractions. Architectural autonomy emerges as a viable path for sustaining intelligence beyond the limits of fixed design.