Researchers used an evolutionary algorithm to design quantum circuits for image classification instead of hand-building them.
EXAQC, an evolutionary framework for automated quantum circuit discovery, has been extended to image classification, evolving parameterized quantum circuits that slot between a classical feature extractor and classical prediction layers instead of relying on fixed, hand-designed templates. Tested on MNIST, Fashion-MNIST, and CIFAR-10, the evolved circuits reached 98.42%, 90.62%, and 85.47% accuracy respectively, using gate counts comparable to other quantum architecture-search methods. A separate evolved configuration pushed CIFAR-10 accuracy to 85.68% while using more than 25 times fewer trainable parameters than a 10-layer convolutional neural network. Encoding choice mattered too: rotation-based schemes (RX, RY, U3) outperformed amplitude encoding by 22-25 points on CIFAR-10.
Manually designing quantum circuit architectures has been a major bottleneck in hybrid quantum-classical machine learning, so showing that evolutionary search can match or approach hand-tuned designs while cutting parameter counts by more than 25 times is a real result, not just a tuning trick. That matters because today's quantum processors can barely handle a few dozen noisy qubits, so any viable near-term application needs circuits that are small, not just accurate.
Worth remembering that all of this runs on classical simulators of quantum circuits, not actual quantum chips, so the parameter savings are a promising paper result well before they translate into anything you could run on real hardware.