IBM-led team trains generative quantum eigensolver with hardware feedback at 70 qubits
Researchers from IBM Quantum, Mitsubishi Chemical Corporation, Keio University and NVIDIA report training a generative quantum eigensolver (GQE) with quantum hardware in the loop at a scale of 70 qubits (the article does not say whether physical or logical; the work ran on IBM hardware, so physical is likely). The transformer-based policies generate circuits for sample-based quantum diagonalization and were trained first on noiseless simulation up to 32 qubits, then for 150 iterations each at 46 and 70 qubits on the IBM Heron processor ibm_kobe. Applied without retraining to organotin hydroxide models of photoresists used in extreme-ultraviolet lithography, 27 of 28 runs at 46-70 qubits came within 15 mHa of a density-matrix renormalization group reference, with a median error of 4.5 mHa. The authors say the circuits need 34 to 72 times fewer two-qubit gates than alternative methods at 56-70 qubits. The work is an arXiv preprint and has not yet been peer reviewed.