Quantum AI Just Got 2,500x Faster to Train — Using Reinforcement Learning

Training quantum neural networks is hard: gradients vanish (the "barren plateau"), and differentiating through an n-qubit, L-layer circuit costs O(L·2^n) time and memory. The authors propose RLQ-Grad, a reinforcement-learning optimizer where a classical policy (a spectrally-normalized PPO agent) learns to propose parameter updates directly, conditioned on the QNN's current parameters, loss, accuracy, and previous update. Because the surrogate gradient comes from a classical network, its variance isn't bound by the barren plateau, and its cost scales with trainable parameters rather than Hilbert-space dimension. On a hardware-efficient ansatz across four supervised benchmarks up to n=20 qubits, it holds a near-flat gradient-variance curve while backpropagation, parameter-shift, and adjoint differentiation decay by 1 to 2 orders of magnitude — and it needs under 2 MB of memory, running 2490×, 7876×, and 673× faster per iteration than those three methods at n=20. It improves top-1 accuracy by up to +10% over gradient-based baselines on circuits up to 12 qubits, and matches dedicated barren plateau mitigation methods on CIFAR-10 at 14 to 20 qubits.