IonQ (NYSE: IONQ) has announced research results showing that generative artificial intelligence (AI) can directly generate quantum optimization circuits. The findings draw attention for their potential to significantly reduce the repetitive parameter tuning that has incurred substantial time and cost in existing quantum optimization.

IonQ said that in joint research with the U.S. Oak Ridge National Laboratory (ORNL), NVIDIA, and the University of Tennessee, Knoxville (UT), it confirmed that a trained generative model can directly write quantum optimization circuits. The paper will be presented at 'IEEE Quantum Week 2026' in Toronto, Canada.

■ Why Repetitive Tuning Became a Bottleneck

Hybrid quantum optimization divides large-scale problems into several smaller subproblems, solves each one, and then recombines the results. This process requires quantum circuits suited to each subproblem.

Previously, finding suitable circuits required repeating the process of execution, measurement, and parameter adjustment—sometimes hundreds of times. Handling larger subproblems could improve solution quality, but tuning costs grew proportionally, limiting the scale of problems that could practically be solved.

In this research, generative AI replaced this repetitive tuning. Martin Roetteler, Vice President of Quantum Applications R&D at IonQ and co-author of the paper, explained, "Getting better answers in hybrid quantum optimization previously required accepting considerable tuning costs. In this benchmark, generative AI replaced the repetitive tuning process, and solution quality improved as the size of the quantum subproblems increased." He added that reducing these costs makes it possible to handle problems at a scale where meaningful solutions can be obtained, and assessed that the results suggest one possible path to scaling hybrid quantum optimization.

■ Transformer of the Same Family as LLMs, Learning Quantum Circuits Instead of Text

To teach the generative model how to adjust quantum circuit instructions, the researchers first trained it on what good results look like. They applied conventional iterative optimization methods to a variety of sample problems, then selected only the near-optimal circuits as training data for a transformer model.

The transformer belongs to the same family used in large language models (LLMs), but in this study it learned quantum circuits instead of text. After training, the model directly generated candidate quantum circuits without the conventional repeated parameter adjustment.

In experiments, the model generated ten candidate circuits per subproblem, scored them all via simulation, and used the highest-scoring circuit for the overall solution update.

■ "Solution Quality Roughly Doubled... Circuit Search Time Maintained at About 28 Seconds"

On a dense, high-dimensional benchmark problem with 100 decision variables, the quality of the model-generated solutions roughly doubled as subproblem size increased.

The time difference was also stark. The existing state-of-the-art method took about 34 seconds to find a circuit at 4 qubits, but the time surged to more than 11 minutes at 12 qubits. In contrast, the generative AI-based approach maintained roughly 28 seconds across all sizes tested.

However, the researchers emphasized that this study was not a direct comparison of quantum and classical computing performance. Both approaches are based on quantum algorithms, and the core of the study lies in comparing different 'quantum circuit generation methods.'

The research was led by ORNL, with co-authors from ORNL's National Center for Computational Sciences (NCCS) and Materials Science and Technology Division, IonQ, NVIDIA, and the University of Tennessee.

Dr. In-Saeng Suh and Dr. Seongmin Kim of ORNL said, "This research is an attempt to solve large-scale complex optimization problems by combining generative AI, quantum computing, and high-performance computing (HPC). AI can become a new computational layer for quantum circuit synthesis, automatically designing and optimizing the quantum circuits needed for even more complex problems." The team is currently extending the framework to real-world scientific and engineering problems while also working on scaling it to larger HPC systems.

Sam Stanwyck, Director of Quantum Products at NVIDIA, said, "Leveraging accelerated computing and AI to unlock breakthroughs in quantum algorithms is one of the promising ways to reach useful quantum applications quickly. With tools like CUDA-Q, developers can now design quantum algorithms with AI at the center from the ground up, laying the foundation for the advancement of next-generation quantum computing."

■ Simulations, Not Real Quantum Hardware... 'Benchmark-Scale Validation'

The results were characterized in the paper as 'benchmark-scale validation,' meaning the circuits were not run on an actual quantum computer. All circuits used in the research were executed in simulated environments rather than on real quantum hardware.

The researchers used NVIDIA's CUDA-Q open platform and cuQuantum SDK on a single NVIDIA H200 GPU installed in the 'Defiant2' system at the Oak Ridge Leadership Computing Facility. By running both the conventional iterative tuning method and the generative AI approach on the same GPU-accelerated infrastructure, they compared the entire workflows of both methods in a controlled environment.

According to the researchers, the measured performance differences mainly arose as 'DQAOA-GPT' replaced conventional iterative variational parameter optimization with generative circuit synthesis and a fixed number of candidate evaluations.

The paper is available on arXiv as '2607.20225.' It is one of nine IonQ papers accepted at IEEE Quantum Week 2026, to be held September 13–18 at the Metro Toronto Convention Centre, and it also won the Best Paper Award.

['Seohak!Star' is overseas stock content, including U.S. stocks, provided by CBC News. We will continue striving to deliver a wide range of information to our readers. Please note that all investments are made by each individual's own judgment, and the final responsibility lies with the investor.]