Curiosity-driven algorithms produce 35% more complex neural architectures than reward-based training in controlled simulations

Researchers ran matched pairs of spiking neural networks on identical sensory streams. One set received only extrinsic reward prediction error; the other added a curiosity term based on prediction surprise. After 500 epochs the curiosity networks developed 2.8 times more distinct cell assemblies and retained higher entropy in connection weights, measured via graph-theoretic metrics. The result held after controlling for total energy expenditure and network size. The finding reframes classic Hebbian accounts by showing that information-seeking objectives can substitute for some forms of supervised feedback during early development. Human data remain absent; the model predicts measurable differences in resting-state connectivity between high- and low-curiosity individuals by adolescence. Follow-up work will test whether pharmacological dampening of dopaminergic novelty signals reduces the same complexity markers in rodents. If replicated, the result could shift educational interventions toward curiosity scaffolding rather than performance incentives within two years.