IonQ (NYSE: IONQ), Oak Ridge National Laboratory (ORNL), Nvidia, and the University of Tennessee, Knoxville published joint research showing that a trained generative AI model can write quantum optimization circuits directly, replacing a repetitive parameter-tuning process that had limited the scale of problems researchers could solve.

The paper was presented at IEEE Quantum Week 2026 in Toronto and received a best paper award. It is one of nine IonQ papers accepted at the conference.

In conventional hybrid quantum optimization, each subproblem requires a tailored quantum circuit built through iterative trial-and-error tuning, which becomes more time-consuming as subproblem size grows. The research team trained a transformer model — the same class of model used in large language models — on near-optimal circuits produced by the conventional method, then used it to generate candidate circuits without the tuning loop.

In benchmark tests on a problem with 100 decision variables, the conventional method's circuit-finding time rose from roughly 34 seconds on 4 qubits to more than 11 minutes on 12 qubits. The generative approach held at nearly 28 seconds across all tested sizes. Solution quality roughly doubled as subproblem size grew under the generative method.

All circuits in the study were simulated rather than run on quantum hardware, using Nvidia's cuQuantum SDK through the CUDA-Q platform on a single Nvidia H200 GPU in ORNL's Defiant2 system.

"Better answers in hybrid quantum optimization have traditionally come with a steep tuning tax. In this benchmark, generative AI replaced the iterative tuning loop, and as the quantum subproblems grew the solution quality improved," said Dr. Martin Roetteler, IonQ's Vice President of Quantum Applications R&D and a co-author of the paper.

ORNL led the study. The paper describes its results as benchmark-scale validation. The research is based on information from a press release issued by IonQ.