GPT meets Quantum Approximate Optimization
Researchers from NVIDIA (Yuri Alexeev, Marwa Farag, Elica Kyoseva), University of Delaware (Ilya Safro and Ilya Tyagin), and Virginia Tech (Kyle Sherbert and Karunya Shirali) have introduced a novel approach to quantum circuit design that leverages the performance of large language models to accelerate and scale quantum optimization workflows.
The quantum computing community has long grappled with the bottleneck of manually or iteratively constructing quantum circuits for problems like combinatorial optimization tasks. While techniques such as QAOA (Quantum Approximate Optimization Algorithm) and its adaptive versions like ADAPT-QAOA have shown promise, their classical optimization overhead and limited scalability remain major barriers. In practice, the number of distinct algorithmic classes in quantum computing is surprisingly limited. Most quantum algorithms fall into just a few well-known circuit design paradigms, such as variational, amplitude amplification, or phase estimation-based approaches. As a result, researchers are extremely interested in automating the generation of novel quantum circuits and improving the scalability and performance of existing algorithmic classes. In this collaborative effort, NVIDIA, Vtech and UD worked together to introduce the first version of QAOA-GPT, a generative transformer model that automates this process, replacing iterative ansatz construction with fast, one-shot inference.
“AI is becoming an essential tool in advancing quantum research, helping us overcome scalability and design challenges that were previously out of reach,” says Ilya Safro. “But this is just the beginning—many important developments lie ahead, and we’re actively looking for talented and motivated students to join us on this journey. Through its PhD programs in Computer Science and in Quantum Science and Engineering, the University of Delaware prepares students for exciting work in the rapidly evolving field of quantum computing.” He adds that the university has also recently introduced a new MS degree in Artificial Intelligence.
QAOA-GPT is trained to generate problem-specific quantum circuits using a custom dataset of graph-structured optimization problems paired with preoptimized high-performing circuits generated using ADAPT-QAOA. The model was augmented with machine learning representations of the optimization problem to encode its structural characteristics, which are then used to condition the circuit generation process. The model treats circuits as sequences, similar to code or natural language, allowing it to leverage the power of transformer architectures.
“Generative AI is proving to be a key enabling tool for quantum computing researchers in a number of areas, including applications development,” said Elica Kyoseva, Director of quantum algorithm engineering at NVIDIA. “The work NVIDIA has completed with Virginia Tech and the University of Delaware shows how AI can shorten the timeline to useful quantum applications, by generating compact and efficient quantum circuits for running them.”
Unlike prior approaches, QAOA-GPT sidesteps gradient-based optimization altogether at inference time. This results in orders-of-magnitude speedups while maintaining great solution quality across diverse instances. The current model is lightweight and runs efficiently on a single GPU, yet can generalize to out-of-distribution inputs, thanks to its rich training data and structural priors.
This innovation enables scalable quantum circuit generation and makes an important step toward bridging the gap between the latest developments in AI and quantum algorithm design. It illustrates how transformer-based generative models can serve as powerful assistants in quantum software engineering, just as they currently transform classical programming. Moreover, the use of NVIDIA CUDA-Q platform for high-performance training data generation exemplifies the synergy between classical GPU-accelerated AI and quantum computation.
“It will be exciting to see how combining state of the art techniques such as ADAPT-VQE and generative AI can pave the path to scalable quantum computing,” said Karunya Shirali.
As quantum hardware continues to mature, tools like QAOA-GPT will be essential for making practical use of these devices. This work sets the stage for a future in which large-scale generative AI plays a central role in automating and optimizing quantum algorithms, paving the way for impactful applications in logistics, materials, biomedical research, and beyond.