Insider Brief
- IonQ, ORNL, NVIDIA and the University of Tennessee, Knoxville demonstrated a generative AI method that directly generates quantum optimization circuits, replacing iterative parameter tuning.
- In benchmark experiments, the generative approach maintained circuit-generation time at about 28 seconds across tested problem sizes, while the prior method increased from about 34 seconds on four qubits to more than 11 minutes on 12 qubits.
- The study used NVIDIA H200 GPU simulation rather than quantum hardware and found that model-generated solution quality roughly doubled as subproblem size increased on a 100-variable benchmark.
PRESS RELEASE — IonQ (NYSE: IONQ), the world’s leading full-stack quantum platform and foundry, today detailed joint research with Oak Ridge National Laboratory (ORNL), NVIDIA, and the University of Tennessee, Knoxville (UT) showing that a trained generative model can write quantum optimization circuits directly, eliminating the trial-and-error parameter-tuning loop that has made the most accurate approach too costly to run. The paper is being presented this week at IEEE Quantum Week in Toronto.
Hybrid quantum optimization breaks a large problem into smaller pieces, solves each one, and recombines the results. Each piece needs a tailored quantum circuit, which traditionally required trial- and-error parameter tuning: run, measure, adjust, and repeat, often hundreds of times. Larger pieces can improve answers, but they also raise tuning costs, limiting the size of problems researchers could solve.
“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. “Take that cost away and you can work at the size where the answer is meaningful. The result provides a potential path toward scaling hybrid quantum optimization, unlocking completely new capabilities and scales that align with IonQ’s existing and future quantum computing hardware generations.”
To teach the generative model how to tune circuit instructions, the team first showed it what strong results looked like. They ran the conventional trial-and-error method across many sampled problems, kept only the near-optimal circuits, and used those examples to train a transformer, the same class of model behind large language models but trained on circuits instead of text.
The trained model generates candidate quantum circuits directly, without the repetitive parameter-tuning loop used by conventional methods. In the reported experiments, the model sampled ten candidate circuits for each subproblem; all ten were simulated and scored, and the best-scoring candidate was used to update the global solution.
On a dense, higher-order benchmark problem with 100 decision variables, model-generated answer quality roughly doubled as subproblems grew. Under the prior state-of-the-art method, circuit-finding time rose sharply from about 34 seconds on 4 qubits to more than 11 minutes on 12 qubits. The generative approach held nearly 28 seconds for every size previously tested. Both approaches are quantum methods; the study compares circuit-generation approaches, not quantum against classical solvers.
ORNL led the study. Co-authors span ORNL’s National Center for Computational Sciences and its Materials Science and Technology Division, IonQ, NVIDIA, and UT. Abhinav Rijal, a graduate researcher in the Department of Physics and Astronomy at UT is a co-author of the study.
“This work brings generative AI, quantum computing, and high-performance computing together to tackle large-scale, complex optimization problems. AI can become a new computational layer for quantum circuit synthesis, enabling the automatic design and optimization of quantum circuits for increasingly complex problems. We are now extending the framework to real-world scientific and engineering applications and scaling it across larger HPC systems to address problems of even greater scale and complexity,” said Dr. In-Saeng Suh and Dr. Seongmin Kim, National Center for Computational Sciences, ORNL.
“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, Quantum Product at NVIDIA. “By enabling developers to build quantum algorithms architected around AI from the outset, tools like CUDA-Q are laying the foundation for the next generation of advances in quantum computing and its useful application.”
The paper describes its results as benchmark-scale validation.
Every circuit in the study was simulated rather than executed on quantum hardware, using the NVIDIA cuQuantum SDK through the NVIDIA CUDA-Q open platform for quantum-GPU supercomputing on a single NVIDIA H200 GPU in the Oak Ridge Leadership Computing Facility’s Defiant2 system. That environment is what makes the comparison meaningful: both the trial-and-error method and the generative one ran on identical GPU-accelerated infrastructure, providing a controlled comparison of the two end-to-end workflows. The measured difference principally reflects the distributed quantum approximate optimization algorithm (DQAOA)-GPT’s replacement of iterative variational parameter optimization with generative circuit synthesis and a fixed number of candidate evaluations.
The paper is available at arXiv:2607.20225. It is one of nine IonQ papers accepted at IEEE Quantum Week 2026, held Sept. 13–18 at the Metro Toronto Convention Centre, and the winner of a best paper award.