Generative AI Cuts Quantum Circuit Tuning Time by 96%

Generative AI replaces trial-and-error in quantum optimization

A collaboration between IonQ, Oak Ridge National Laboratory, NVIDIA, and the University of Tennessee has demonstrated a generative AI method that writes quantum optimization circuits on demand, eliminating the repetitive parameter-tuning process that has constrained hybrid quantum approaches. The research was presented at IEEE Quantum Week in Toronto and won a best paper award.

Hybrid quantum optimization divides large problems into smaller subproblems, solves each with a tailored quantum circuit, then recombines the results. Traditionally, creating each circuit required hundreds of trial-and-error iterations—run the circuit, measure performance, adjust parameters, repeat. Larger subproblems yield better answers but multiply tuning costs, creating a practical ceiling on problem scale.

The team trained a transformer model—the same architecture underlying large language models—on examples of high-performing circuits generated through conventional tuning. Once trained, the model generates candidate circuits directly. In benchmark tests, researchers sampled ten candidates per subproblem, simulated all ten, and selected the best.

Benchmark results show dramatic time savings

On a 100-variable optimization benchmark, circuit-generation time using the prior state-of-the-art method climbed from approximately 34 seconds on four qubits to more than 11 minutes on 12 qubits. The generative approach held steady at roughly 28 seconds across all tested problem sizes—a 96% reduction at the 12-qubit scale.

Solution quality also improved as subproblem size increased. Model-generated answers approximately doubled in quality as researchers scaled up subproblems on the 100-variable test case, according to the study.

"Better answers in hybrid quantum optimization have traditionally come with a steep tuning tax," said Dr. Martin Roetteler, IonQ's Vice President of Quantum Applications R&D and a co-author. "Take that cost away and you can work at the size where the answer is meaningful."

Why it matters

The tuning bottleneck has forced researchers to choose between solution quality and computational cost. By replacing iterative optimization with direct generation, the approach could enable hybrid quantum methods to tackle larger, more complex problems without proportional increases in runtime. That shift matters for near-term quantum applications, where hybrid algorithms represent one of the most practical paths to useful results on current hardware.

GPU simulation provides controlled comparison

All circuits in the study were simulated rather than executed on quantum hardware. The team used NVIDIA's cuQuantum SDK and CUDA-Q platform running on a single NVIDIA H200 GPU in Oak Ridge's Defiant2 system. Both the conventional and generative methods ran on identical infrastructure, isolating the performance difference to the circuit-generation workflow itself.

"Drawing on accelerated computing and AI to make breakthroughs in quantum algorithms is one of the most promising ways to reach useful quantum applications as quickly as possible," said Sam Stanwyck, Director of Quantum Product at NVIDIA.

Oak Ridge National Laboratory led the research, with co-authors from ORNL's National Center for Computational Sciences, IonQ, NVIDIA, and the University of Tennessee. The paper is available at arXiv:2607.20225. Details were first reported by The Quantum Insider.