Kenya Sakka and colleagues at The University of Osaka presents an autonomous framework using large language models to iteratively design quantum circuits subject to specific constraints. The system, comprising seven integrated components for exploration, generation, and evaluation, combines online knowledge gathering with code execution and experimental results. It offers a viable pathway towards automated quantum circuit design, achieving competitive performance in both quantum machine learning and quantum chemistry applications. The framework outperforms existing quantum feature maps and rivals established ansatz constructions, highlighting the potential for AI to contribute to iterative scientific optimisation.
Learned quantum circuit design rivals established methods for molecular ground state estimation
A learned ansatz outperforms representative quantum feature maps and surpasses the classical radial basis function kernel when scaled to larger qubit counts. In molecular ground state estimation across seven molecules, the generated ansatz attains competitive accuracy with widely used chemically inspired and hardware-efficient constructions while satisfying the imposed scaling constraints. These results establish LLM driven agentic systems as a viable model for automated quantum circuit design and illustrate how AI systems can participate in iterative scientific optimisation workflows across scientific domains.
Quantum circuit structure is central to the practical performance of quantum algorithms. In quantum machine learning, the structure of a quantum feature map defines the geometry of the induced Hilbert space and directly influences generalisation performance. In quantum chemistry, ansätze determine the expressibility, trainability, and resource efficiency of variational quantum eigensolvers (VQE) used for molecular ground-state energy estimation. Despite advances in quantum hardware and algorithmic techniques, quantum circuit design largely depends on human intuition, domain-specific heuristics, and incremental modifications of established design paradigms.
This reliance on manual design presents inherent challenges, as limited theoretical understanding of the relationship between quantum circuit structure and performance makes principled circuit design difficult both within specific tasks and across different application domains. Specialisation is often unavoidable, yet it requires substantial human expertise and iterative experimentation, and the design burden increases with system size and circuit complexity. Automated quantum architecture search methods have been proposed to expand exploration beyond manual heuristics, but these approaches typically operate within predefined templates.
Large language models (LLMs) have recently demonstrated the ability to generate quantum programs and assist components of scientific workflows. However, most approaches automate isolated components, such as literature review or code generation, rather than embedding them within a closed-loop research cycle that systematically explores, evaluates, and refines circuit designs. Previous work demonstrated the feasibility of prompt-based circuit generation guided by quantitative feedback, achieving performance exceeding representative quantum baselines on benchmark datasets.
That study was restricted to a single design task and relied primarily on the internal knowledge of the LLM during initial idea generation, without systematically integrating external domain knowledge. The refinement process was driven mainly by performance-based feedback and lacked structured multi-perspective critique grounded in prior research. Consequently, candidate quality varied, and computational effort was sometimes directed toward less promising directions, limiting the framework to a specific application setting.
This work presents an extended agentic framework that advances autonomous circuit design beyond task-specific pipelines. The system embeds LLMs within a structured research loop comprising seven components: “Exploration”, “Generation”, “Discussion”, “Validation”, “Evaluation”, “Storage”, and “Review”. The “Exploration” component conducts web-based investigation before idea generation, surveying existing directions and expanding diverse seed designs. The “Discussion” component assigns domain-specific expert roles and evaluates candidate ideas alongside retrieved literature, enabling structured multi-perspective critique.
The framework generates quantum circuits whose implementation logic is independent of a fixed number of qubits, ensuring scalability across system sizes. These mechanisms enable autonomous generation of circuit candidates while incorporating task performance, scalability, and resource efficiency into the design loop. The proposed framework was evaluated on two conceptually distinct tasks: quantum feature map construction for QML and ansatz generation for molecular ground state energy estimation with the VQE. A central objective in the QML setting is to achieve performance that competes with or exceeds classical machine learning models.
Starting from diverse seed designs, the system iteratively refines candidate feature maps through quantitative feedback, yielding steady performance improvements over successive trials. The best-performing feature map outperforms representative baseline quantum feature maps as well as the classical radial basis function (RBF) kernel on MNIST, Fashion-MNIST, and CIFAR-10. Furthermore, the system produces executable feature map code whose logic remains valid across a range of qubit numbers, confirming that the learned design principles are scalable rather than tied to a fixed circuit size. For molecular ground-state estimation, the best-performing ansatz outperforms several widely used hardware-efficient ansätze, while remaining more compact than chemically motivated constructions and achieving competitive accuracy with favourable parameter scaling.
Researchers begin by introducing quantum feature maps for QML and the variational quantum eigensolver for molecular ground state estimation, together with an overview of representative circuit design principles developed in previous studies. In quantum machine learning, a quantum feature map is a data embedding circuit that maps a classical input x to a quantum state through an input dependent unitary U(x). Starting from the initial state |0⟩⊗n, the resulting quantum feature state can be written as ρ(x) = U(x)|0⟩⊗n⟨0|⊗nU(x)†. The design of U(x) determines how classical information is embedded into Hilbert space and therefore strongly affects the expressivity and generalisation behaviour of the resulting model. Researchers focus on the quantum kernel approach, where the similarity between two inputs is measured by the overlap between their corresponding quantum feature states.
A common choice is the Hilbert, Schmidt inner product k(x, x′) = Try[ρ(x)ρ(x′)]. This kernel can be used with standard kernel based learning algorithms such as support vector machines or kernel ridge regression. Because predictive performance depends directly on the structure of the feature map circuit, the design of the quantum feature map is a central problem in QML and serves as one of the main targets of the automated circuit generation framework. The VQE is a hybrid quantum, classical algorithm that estimates the ground state energy of a given Hamiltonian H. It prepares a parameterised quantum state |ψ(θ)⟩= U(θ)|0⟩⊗n using an ansatz circuit U(θ) and iteratively optimizes the parameters θ so that the expectation value ⟨ψ(θ)|H|ψ(θ)⟩becomes minimal.
Once the optimisation converges, the obtained energy corresponds to the ground-state energy, and the optimised state |ψ(θ)⟩approximates the ground state of the Hamiltonian. A key factor determining the performance of VQE is the design of the ansatz circuit. An autonomous agentic framework employing large language models (LLMs) conducts iterative quantum circuit designs under explicit constraints. This system integrates seven components, Exploration, Generation, Discussion, Validation, Storage, Evaluation, and Review, forming a closed-loop workflow that combines web-based knowledge, literature critique, code generation, and experimental feedback.
Evaluations focus on quantum feature map construction for machine learning and ansatz generation for variational quantum eigensolver applications in quantum chemistry. Generated feature maps outperform representative quantum constructions and, with increased qubit counts, surpass classical radial basis function kernels. Generated ansätze attain competitive accuracy with chemically inspired and hardware-efficient constructions while satisfying scaling constraints.
The use of AI for scientific discovery has expanded rapidly in recent years, with LLMs in particular enabling new approaches to automate and accelerating scientific research. In particular, LLMs have been used to automate and support key components of the research process, including literature analysis, hypothesis generation, experimental planning, software development, and the coordination of multistep scientific workflows. Such approaches have been actively explored across a wide range of scientific domains, ranging from the automation of computational and simulation based research workflows to the automation of experimental research involving interaction with physical systems and laboratory hardware, with the aim of accelerating discovery and reducing manual intervention. Motivated by these developments, similar ideas have begun to emerge in the context of quantum computing, where the design and optimisation of quantum circuits are vital.
Autonomous agents surpass classical methods in quantum circuit optimisation
Sakka and colleagues at The University of Osaka has achieved 96.2% accuracy with generated quantum feature maps, exceeding the performance of existing quantum designs and, in particular, surpassing the 93.8% accuracy of the classical radial basis function kernel when scaled to larger qubit counts, a threshold previously unattainable without extensive manual optimisation. This demonstrates a viable pathway towards automated quantum circuit design, enabling the creation of circuits competitive with those designed by human experts, and opens possibilities for tackling more complex quantum machine learning problems. The team’s autonomous agentic framework, utilising large language models, systematically explores, evaluates, and refines circuit designs, integrating web-based knowledge with code execution and experimental results to achieve this performance leap.
Researchers at The University of Osaka demonstrated the framework’s effectiveness across two distinct quantum applications. In quantum feature map construction, the generated designs achieved superior performance compared to existing quantum circuits, improving upon a baseline of 88.2% accuracy on image classification tasks. Furthermore, when applied to variational quantum eigensolver applications for quantum chemistry, the autonomously created ansatz matched the accuracy of established, manually designed circuits across seven different molecules. The team also imposed scaling constraints, ensuring the generated circuits remained efficient as the number of qubits increased.
Balancing chemical accuracy and circuit complexity in automated quantum design
Automated quantum circuit design promises to alleviate the bottleneck of human expertise currently limiting progress in both quantum machine learning and quantum chemistry applications. The researchers successfully demonstrated a system capable of generating circuits competitive with established methods, yet the study highlights a key tension between accuracy and resource demands. Incrementally adding modules to the generated ansatz demonstrably improved performance, particularly with the ‘Upal’ component achieving chemical accuracy for most molecules tested, but this came at the cost of significantly increased circuit complexity.
Nevertheless, increased circuit complexity accompanying improved accuracy presents a practical challenge for near-term quantum devices. Adding components within an agentic framework employing large language models iteratively designs quantum circuits under explicit constraints. This system integrates components including exploration, generation, discussion, validation, storage, evaluation, and review, forming a closed-loop workflow combining knowledge acquisition and experimental feedback.
Across seven molecules, generated ansatz attains competitive accuracy with existing constructions while satisfying scaling constraints. The team at The University of Osaka demonstrated a functioning system where artificial intelligence designs quantum circuits, moving beyond simply assisting existing workflows. This agentic framework, driven by large language models, iteratively improves designs by combining internet research with testing and evaluation, effectively automating a process previously reliant on expert intuition. Achieving competitive results in both quantum machine learning and molecular modelling signifies a potential shift in how quantum algorithms are developed, offering a pathway to explore more complex computational problems.
The research demonstrated an agentic system, driven by large language models, successfully designed quantum circuits for both quantum machine learning and quantum chemistry applications. This is significant because it offers a new, automated approach to circuit development, previously reliant on human expertise. Generated circuits achieved competitive accuracy with established methods across seven molecules and outperformed some classical approaches in image classification benchmarks. The researchers suggest this framework represents a viable paradigm for automated quantum circuit design and iterative scientific optimisation.
Stay current. See today’s quantum computing news on Quantum Zeitgeist for the latest breakthroughs in qubits, hardware, algorithms, and industry deals.
AloJapan.com