A research collaboration led by Oak Ridge National Laboratory (ORNL) alongside IonQ (NYSE: IONQ), NVIDIA, and the University of Tennessee, Knoxville (UT) has introduced DQAOA-GPT, a generative AI framework that synthesizes quantum optimization circuits directly to eliminate iterative parameter-tuning loops in distributed quantum algorithms. Presented at IEEE Quantum Week 2026 in Toronto, the paper received a Best Paper Award for demonstrating constant-time circuit synthesis for subproblem evaluations across scaling quantum domain widths.
The framework replaces the traditional trial-and-error variational loop of the Distributed Quantum Approximate Optimization Algorithm (DQAOA) with a transformer model trained on high-performing circuit profiles. For each subproblem, the generative transformer outputs candidate quantum circuits directly, which are evaluated in a fixed 10-candidate sampling step before updating global solution parameters. Benchmarked on a 100-variable higher-order unconstrained binary optimization (HUBO) problem using single NVIDIA H200 GPUs via NVIDIA CUDA-Q and the cuQuantum SDK, conventional variational circuit optimization times escalated from 34 seconds (4 qubits) to over 11 minutes (12 qubits), whereas the generative DQAOA-GPT approach maintained a constant synthesis runtime of approximately 28 seconds regardless of subproblem qubit count while doubling overall solution quality.
| [ DQAOA-GPT vs. Variational Quantum Circuit Synthesis Baseline ] | ||
|---|---|---|
| Optimization Metric | Traditional Variational DQAOA | Generative DQAOA-GPT Framework |
| Circuit Synthesis Runtime | • 4 Qubits: ~34 Seconds • 12 Qubits: >11 Minutes (>19.4× Growth) | • 4 to 12 Qubits: ~28 Seconds (Constant Runtime) • Zero Iteration Parameter Optimization Loop |
| Solution Quality Scaling | • Degrades/Stagnates Due to Tuning Latency Costs | • Approx. 2× Improvement in HUBO Answer Quality |
| Compute Stack Baseline | • Variational Classical Optimizer Loop (GPU) | • Transformer Circuit Synthesis + CUDA-Q Scoring |
By eliminating the classical optimization bottleneck, DQAOA-GPT enables hybrid algorithms to scale subproblem sizes without incurring prohibitive execution overheads. The framework serves as a scalable software layer across GPU-accelerated HPC systems and IonQ’s trapped-ion hardware pipeline, paving the way for larger real-world scientific and industrial optimization workloads.
Review the official press release via IonQ Newsroom here, inspect the peer-reviewed research preprint on arXiv here, and examine our prior coverage of IonQ’s IEEE Quantum Week 2026 HPC/AI Research Portfolio here.
September 16, 2026
