Abstract
<title>Abstract</title> <p>Adaptive allocation of computation is important for efficient neuromorphic systems, yet physical reservoir architectures typically apply the same processing pathway to every input. Here, we present a confidence-guided framework that coordinates simulated memristive nanowire-network reservoirs as complementary experts, enabling adaptive physical reservoir computing through dynamic allocation of computational resources. Three memristive nanowire-network reservoirs independently processed the same temporally encoded MNIST input and produced class probabilities through linear classifier readouts. Training the readouts on non-overlapping image subsets generated complementary error patterns, creating training-induced reservoir diversity, whereas shared training produced highly correlated representations. Across ten shifted training-window experiments, mean individual reservoir accuracies ranged from 82.41% to 82.81%. Majority voting achieved 84.55 ± 0.82%, while confidence-assisted voting provided the highest parallel accuracy of 84.71 ± 0.75%. A cascade order and confidence percentile selected on a separate validation subset were fixed before final testing. The resulting 200 → 500 → 350-nanowire cascade achieved 84.38 ± 0.64% accuracy while evaluating 1.370 ± 0.053 reservoirs per sample. This corresponded to 54.3% fewer reservoir-stage evaluations and, under the adopted simulation model, estimated reduction of 54.32 ± 1.78% in normalised energy relative to concurrent parallel execution. These results demonstrate that training-induced expertise and confidence-guided routing provide a controllable accuracy-resource trade-off in modular reservoir systems, offering a scalable strategy for resource-efficient neuromorphic inference and edge AI applications.</p>