Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Log inSign up
Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Blog
Job Board
Q3AS

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

Journal of Quantum Artificial Intelligence

Warm-Started Variational Quantum Optimization for Accelerating Deep Model Training under Compute Constraints

Ashardi Abas

Published Oct 27, 2026• 0 citations
PDF (CC BY)Publisher

Abstract

Compute constraints increasingly determine whether modern deep models can be trained, tuned, and deployed in practice, yet most optimization pipelines still assume abundant iterations, repeated trials, and stable accelerator access. This paper presents Warm-Started Variational Quantum Optimization (WS-VQO), a hybrid controller that accelerates training by proposing structured updates to optimizer controls under an explicit budget of 14 GPU-hours and an upper bound of 18,000 quantum circuit evaluations. Across three architectures, ViT-Small, ResNet-24, and a 12-layer text encoder, WS-VQO improves final validation accuracy at the budget boundary from 76.9% to 78.6%, from 80.6% to 82.1%, and from 77.7% to 79.3%, respectively, while reducing time-to-target by 2.5 GPU-hours on ViT-Small (11.6 to 9.1), 2.1 GPU-hours on ResNet-24 (10.5 to 8.4), and 2.2 GPU-hours on the text encoder (11.8 to 9.6). Ablations show that removing warm start reduces ViT-Small accuracy by 1.1 points and increases time-to-target by 1.2 GPU-hours, while removing the budget penalty increases outcome variance and slows convergence despite similar peak accuracy. Under effective noise rates up to 0.04 and shot budgets of 256–1024, performance degrades smoothly, with final accuracy ranging from 78.6% to 76.4% at 1024 shots and from 77.7% to 75.2% at 256 shots, approaching classical baselines as proposal resolution diminishes. Runtime accounting indicates that quantum overhead is dominated by queue delay rather than circuit execution, and the compute cost model tracks observed consumption with a mean absolute error of 3.1%, enabling reliable budget enforcement without overruns. These results demonstrate that warm-started quantum-assisted control can deliver measurable acceleration and stable gains under strict compute ceilings.

DOI
10.63913/jqai.v1i3.20

Related papers

Reinforcement Learning Based Warm Initialization for Constrained Open-System Quantum Optimal Control: A Controlled Budget-Matched RL-GRAPE Benchmark

Daniele Gabriele, Lorenzo Ricciardi Celsi

Superconducting-qubit control is fundamentally constrained by decoherence, finite bandwidth, and hardware-limited drive amplitudes, making high-fidelity state preparation sensitive to optimizer initialization under non-convex open-system dynamics. We propose a hybrid reinforcement learning (RL)–quantum optimal control (QOC) pipeline in which a lightweight, tabular, model-free RL agent is trained offline in simulation to generate feasible, bounded seed pulses, which are subsequently refined via GRAPE under Lindblad dynamics. Hard amplitude constraints are enforced consistently across both stages, ensuring strict feasibility throughout optimization. Performance is evaluated using a budget-matched protocol based on fidelity evaluations (F-evals), enabling controlled comparison with random-start multi-start GRAPE. On a transmon-like qubit benchmark with relaxation and dephasing, RL warm-starting reduces the median online refinement effort in the adopted finite-difference GRAPE implementation from 7568 to 3543 F-evals (2.14× reduction) while achieving terminal state fidelity ≥0.995 under identical constraints and evaluation budgets. We provide a theoretical interpretation of the improvement in terms of basin-of-attraction probability shaping in constrained control landscapes and an amortized cost analysis showing that the offline RL cost is recovered after a small number of reuse cycles. The results support the view that learning-based initialization can improve warm-start quality relative to uninformed feasible multi-start in constrained open-system quantum-control benchmarks, while broader practical comparison against stronger physics-guided seeds remains for future work.

Electronics•Mar 17, 2026•0 citations

Generative flow-based warm start of the variational quantum eigensolver

Hang Zou, Martin Rahm, Anton Frisk Kockum +1 more

Abstract Hybrid quantum-classical algorithms like the variational quantum eigensolver (VQE) show promise for quantum simulations on near-term quantum devices, but are often limited by complex objective functions and expensive optimization procedures. Here, we propose Flow-VQE, a generative framework leveraging conditional normalizing flows with parameterized quantum circuits to efficiently generate high-quality variational parameters. By embedding a generative model into the VQE optimization loop through preference-based training, Flow-VQE enables quantum gradient-free optimization and offers a systematic approach for parameter transfer, accelerating convergence across related problems through warm-started optimization. We compare Flow-VQE to a number of standard benchmarks through numerical simulations on molecular systems, including hydrogen chains, water, ammonia, and benzene. We find that Flow-VQE generally outperforms baseline optimization algorithms, achieving computational accuracy with fewer circuit evaluations (improvements range from modest to more than two orders of magnitude) and, when used to warm-start the optimization of new systems, accelerates subsequent fine-tuning by up to 50-fold compared with Hartree–Fock initialization. Therefore, we believe Flow-VQE can become a pragmatic and versatile paradigm for leveraging generative modeling to reduce the costs of variational quantum algorithms.

arXiv:2507.01726•npj Quantum Information•Dec 23, 2025•6 citations

Adaptive V16: Cost-Aware Reinforcement Learning Control for Robust NISQ Variational Optimization

Kondamudi Sivaparvathi, Cheppali Prabhavathi

Variational quantum algorithms (VQAs) are promising for near-term noisy intermediate-scale quantum (NISQ) devices, but their performance is often limited by device noise, finite-shot uncertainty, and nonstationary behavior that destabilize optimization. This paper proposes a cost-aware closed-loop hybrid framework (Adaptive V16) that augments a standard VQA loop with an online decision policy to adapt measurement effort and mitigation actions during execution. At each iteration, the controller uses runtime signals such as objective progress and uncertainty to select an action comprising the shot budget, mitigation mode, and step-scaling factor, explicitly trading solution improvement against resource overhead under a fixed budget. Experimental evaluation on a representative molecular Hamiltonian optimization task shows that Adaptive V16 achieves a lower final energy than a static VQE baseline and reduces the absolute error from 0.1391 to 0.0657. The controller’s decision module also demonstrates strong predictive performance with 0.9700 accuracy, 0.9595 F1-score, and 0.9951 ROC–AUC, indicating reliable adaptive behavior across thresholds. These results suggest that policy-driven adaptation can improve robustness and budget efficiency for hybrid quantum–classical optimization in realistic NISQ settings.

International Conference on Knowledge Engineering and Communication Systems (ICKECS)•Apr 24, 2026•0 citations

Warm-Starting MaxCut Relaxation via Low-Depth Quantum Approximate Optimization Algorithm

Bao G. Bach, Ilya Safro, Filip B. Maciejewski

Quantum optimization has attracted growing interest as quantum hardware continues to improve, yet state-of-the-art classical solvers remain a formidable benchmark for practical utility. Rather than seeking a fully quantum replacement for classical optimization, we propose a hybrid strategy that uses quantum information to enhance leading classical heuristics. Specifically, we introduce a warm-start method based on local correlators obtained from the Quantum Approximate Optimization Algorithm (QAOA), and use this information to initialize the Burer-Monteiro (BM) rank-two relaxation. We demonstrate numerically that, compared to a random, multi-start initialization baseline (a standard strategy used for BM), this quantum-informed initialization offers a significant head start, i.e., high-quality solutions with very small number of iterations, for two problem classes -- random Erdős Rényi graphs with edge density of $10\%$ (ER-10) and fully-connected Sherrington Kirkpatrick (SK) spin glass models, at $n=500$ and $n=1000$ qubits. At the same time, given enough iterations, the random baseline often eventually catches up and slightly outperforms the warm-start strategy on average, an effect visibly stronger for $n=500$ than for $n=1000$. The results demonstrate an exploitation/exploration tradeoff of using WS to quickly arrive at very good solutions vs exploring slightly better solutions with a larger iterations budget via a standard strategy. Our results highlight how low-depth quantum circuits can provide useful structural information for classical optimization and suggest a promising route toward near-term quantum utility through quantum-assisted initialization.

quant-pharXiv:2608.06212•Aug 7, 2026•0 citations

A GPU-Accelerated and Geometry-Aware Quantum-Classical Framework for Optimized PES Scans with CUDA-Q

Panayiotis Christou, Shuwen Kan, Ying Mao

Hybrid quantum-classical workloads, particularly Potential Energy Surface (PES) scans in quantum chemistry using batched VQE and QAOA, involve many structurally similar Hamiltonians differing only in coefficients. Existing warm-start and meta-optimization strategies reuse parameters between instances but still consume fixed quantum-classical resources, ignoring cross-instance structural similarities. We present OptiPESS, a GPU-accelerated runtime system for adaptive resource allocation in hybrid quantum-classical execution. OptiPESS introduces the Locality-Occupancy-Variance (LOV) metric, a lightweight, geometry-aware measure of Hamiltonian similarity, coupled with efficient alignment and Pauli-classification passes. Using LOV, OptiPESS performs distance-aware scheduling of iteration and shot budgets, predicting which instances can be skipped, quick-solved, or fully optimized. Implemented atop CUDA-Q with persistent multi-GPU execution and near-linear scaling, OptiPESS reduces classical and quantum evaluations by up to 78.3\% and wall-time by up to 74.8\% while maintaining baseline accuracy. This demonstrates how lightweight static analysis and adaptive scheduling can transform quantum-classical co-execution from fixed-budget loops into a predictive, resource-aware runtime.

quant-pharXiv:2611.01389•Nov 3, 2026•0 citations

Efficient Optimization of Variational Quantum Algorithms via Gradient-Free Parameter Prediction

Satwik Kundu, Debarshi Kundu, Swaroop Ghosh

The exponential runtime of quantum simulators on classical machines, along with long queue times and high costs associated with real quantum devices, presents significant challenges for the efficient optimization of Variational Quantum Algorithms (VQAs) such as the Variational Quantum Eigensolver (VQE), Quantum Approximate Optimization Algorithm (QAOA), and Quantum Neural Networks (QNNs). To address these limitations, we propose a novel approach, DyPP (Dynamic Parameter Prediction), which accelerates the convergence of VQAs by leveraging regular trends in the parameter weights to update parameters efficiently. We introduce two techniques for optimal prediction performance: Naive Prediction (NaP) and Adaptive Prediction (AdaP). Through extensive experimentation and training of multiple QNN models on various datasets, we demonstrate that DyPP achieves a speedup of approximately 2.25×compared to standard training methods, while also providing improved accuracy (up to 2.3% higher) and lower loss (up to 6.1% reduction) with minimal storage and computational overhead. We further evaluate DyPP’s effectiveness in VQE for molecular ground-state energy estimation and in QAOA for graph MaxCut. Our results show that, on average, DyPP leads to a speedup of up to 3.1× for VQE and 2.91× for QAOA compared to traditional optimization, while requiring up to 3.3× fewer shots (i.e., repeated circuit executions). Even in the presence of hardware noise, DyPP outperforms existing optimization techniques, delivering up to 3.33× speedup and requiring 2.5× fewer shots, thereby enhancing the efficiency of VQAs.

arXiv:2307.12449•IEEE Access•Jan 1, 2026•2 citations
Compute constraints increasingly determine whether modern deep models can be trained, tuned, and deployed in practice, yet most optimization pipelines still assume abundant iterations, repeated trials, and stable accelerator access. This paper presents Warm-Started Variational Quantum Optimization (WS-VQO), a hybrid controller that accelerates training by proposing structured updates to optimizer controls under an explicit budget of 14 GPU-hours and an upper bound of 18,000 quantum circuit evaluations. Across three architectures, ViT-Small, ResNet-24, and a 12-layer text encoder, WS-VQO improves final validation accuracy at the budget boundary from 76.9% to 78.6%, from 80.6% to 82.1%, and from 77.7% to 79.3%, respectively, while reducing time-to-target by 2.5 GPU-hours on ViT-Small (11.6 to 9.1), 2.1 GPU-hours on ResNet-24 (10.5 to 8.4), and 2.2 GPU-hours on the text encoder (11.8 to 9.6). Ablations show that removing warm start reduces ViT-Small accuracy by 1.1 points and increases time-to-target by 1.2 GPU-hours, while removing the budget penalty increases outcome variance and slows convergence despite similar peak accuracy. Under effective noise rates up to 0.04 and shot budgets of 256–1024, performance degrades smoothly, with final accuracy ranging from 78.6% to 76.4% at 1024 shots and from 77.7% to 75.2% at 256 shots, approaching classical baselines as proposal resolution diminishes. Runtime accounting indicates that quantum overhead is dominated by queue delay rather than circuit execution, and the compute cost model tracks observed consumption with a mean absolute error of 3.1%, enabling reliable budget enforcement without overruns. These results demonstrate that warm-started quantum-assisted control can deliver measurable acceleration and stable gains under strict compute ceilings.
Superconducting-qubit control is fundamentally constrained by decoherence, finite bandwidth, and hardware-limited drive amplitudes, making high-fidelity state preparation sensitive to optimizer initialization under non-convex open-system dynamics. We propose a hybrid reinforcement learning (RL)–quantum optimal control (QOC) pipeline in which a lightweight, tabular, model-free RL agent is trained offline in simulation to generate feasible, bounded seed pulses, which are subsequently refined via GRAPE under Lindblad dynamics. Hard amplitude constraints are enforced consistently across both stages, ensuring strict feasibility throughout optimization. Performance is evaluated using a budget-matched protocol based on fidelity evaluations (F-evals), enabling controlled comparison with random-start multi-start GRAPE. On a transmon-like qubit benchmark with relaxation and dephasing, RL warm-starting reduces the median online refinement effort in the adopted finite-difference GRAPE implementation from 7568 to 3543 F-evals (2.14× reduction) while achieving terminal state fidelity ≥0.995 under identical constraints and evaluation budgets. We provide a theoretical interpretation of the improvement in terms of basin-of-attraction probability shaping in constrained control landscapes and an amortized cost analysis showing that the offline RL cost is recovered after a small number of reuse cycles. The results support the view that learning-based initialization can improve warm-start quality relative to uninformed feasible multi-start in constrained open-system quantum-control benchmarks, while broader practical comparison against stronger physics-guided seeds remains for future work.
Abstract Hybrid quantum-classical algorithms like the variational quantum eigensolver (VQE) show promise for quantum simulations on near-term quantum devices, but are often limited by complex objective functions and expensive optimization procedures. Here, we propose Flow-VQE, a generative framework leveraging conditional normalizing flows with parameterized quantum circuits to efficiently generate high-quality variational parameters. By embedding a generative model into the VQE optimization loop through preference-based training, Flow-VQE enables quantum gradient-free optimization and offers a systematic approach for parameter transfer, accelerating convergence across related problems through warm-started optimization. We compare Flow-VQE to a number of standard benchmarks through numerical simulations on molecular systems, including hydrogen chains, water, ammonia, and benzene. We find that Flow-VQE generally outperforms baseline optimization algorithms, achieving computational accuracy with fewer circuit evaluations (improvements range from modest to more than two orders of magnitude) and, when used to warm-start the optimization of new systems, accelerates subsequent fine-tuning by up to 50-fold compared with Hartree–Fock initialization. Therefore, we believe Flow-VQE can become a pragmatic and versatile paradigm for leveraging generative modeling to reduce the costs of variational quantum algorithms.
Variational quantum algorithms (VQAs) are promising for near-term noisy intermediate-scale quantum (NISQ) devices, but their performance is often limited by device noise, finite-shot uncertainty, and nonstationary behavior that destabilize optimization. This paper proposes a cost-aware closed-loop hybrid framework (Adaptive V16) that augments a standard VQA loop with an online decision policy to adapt measurement effort and mitigation actions during execution. At each iteration, the controller uses runtime signals such as objective progress and uncertainty to select an action comprising the shot budget, mitigation mode, and step-scaling factor, explicitly trading solution improvement against resource overhead under a fixed budget. Experimental evaluation on a representative molecular Hamiltonian optimization task shows that Adaptive V16 achieves a lower final energy than a static VQE baseline and reduces the absolute error from 0.1391 to 0.0657. The controller’s decision module also demonstrates strong predictive performance with 0.9700 accuracy, 0.9595 F1-score, and 0.9951 ROC–AUC, indicating reliable adaptive behavior across thresholds. These results suggest that policy-driven adaptation can improve robustness and budget efficiency for hybrid quantum–classical optimization in realistic NISQ settings.
Quantum optimization has attracted growing interest as quantum hardware continues to improve, yet state-of-the-art classical solvers remain a formidable benchmark for practical utility. Rather than seeking a fully quantum replacement for classical optimization, we propose a hybrid strategy that uses quantum information to enhance leading classical heuristics. Specifically, we introduce a warm-start method based on local correlators obtained from the Quantum Approximate Optimization Algorithm (QAOA), and use this information to initialize the Burer-Monteiro (BM) rank-two relaxation. We demonstrate numerically that, compared to a random, multi-start initialization baseline (a standard strategy used for BM), this quantum-informed initialization offers a significant head start, i.e., high-quality solutions with very small number of iterations, for two problem classes -- random Erdős Rényi graphs with edge density of 10%10\%10% (ER-10) and fully-connected Sherrington Kirkpatrick (SK) spin glass models, at n=500n=500n=500 and n=1000n=1000n=1000 qubits. At the same time, given enough iterations, the random baseline often eventually catches up and slightly outperforms the warm-start strategy on average, an effect visibly stronger for n=500n=500n=500 than for n=1000n=1000n=1000. The results demonstrate an exploitation/exploration tradeoff of using WS to quickly arrive at very good solutions vs exploring slightly better solutions with a larger iterations budget via a standard strategy. Our results highlight how low-depth quantum circuits can provide useful structural information for classical optimization and suggest a promising route toward near-term quantum utility through quantum-assisted initialization.
Hybrid quantum-classical workloads, particularly Potential Energy Surface (PES) scans in quantum chemistry using batched VQE and QAOA, involve many structurally similar Hamiltonians differing only in coefficients. Existing warm-start and meta-optimization strategies reuse parameters between instances but still consume fixed quantum-classical resources, ignoring cross-instance structural similarities. We present OptiPESS, a GPU-accelerated runtime system for adaptive resource allocation in hybrid quantum-classical execution. OptiPESS introduces the Locality-Occupancy-Variance (LOV) metric, a lightweight, geometry-aware measure of Hamiltonian similarity, coupled with efficient alignment and Pauli-classification passes. Using LOV, OptiPESS performs distance-aware scheduling of iteration and shot budgets, predicting which instances can be skipped, quick-solved, or fully optimized. Implemented atop CUDA-Q with persistent multi-GPU execution and near-linear scaling, OptiPESS reduces classical and quantum evaluations by up to 78.3\% and wall-time by up to 74.8\% while maintaining baseline accuracy. This demonstrates how lightweight static analysis and adaptive scheduling can transform quantum-classical co-execution from fixed-budget loops into a predictive, resource-aware runtime.
The exponential runtime of quantum simulators on classical machines, along with long queue times and high costs associated with real quantum devices, presents significant challenges for the efficient optimization of Variational Quantum Algorithms (VQAs) such as the Variational Quantum Eigensolver (VQE), Quantum Approximate Optimization Algorithm (QAOA), and Quantum Neural Networks (QNNs). To address these limitations, we propose a novel approach, DyPP (Dynamic Parameter Prediction), which accelerates the convergence of VQAs by leveraging regular trends in the parameter weights to update parameters efficiently. We introduce two techniques for optimal prediction performance: Naive Prediction (NaP) and Adaptive Prediction (AdaP). Through extensive experimentation and training of multiple QNN models on various datasets, we demonstrate that DyPP achieves a speedup of approximately 2.25×compared to standard training methods, while also providing improved accuracy (up to 2.3% higher) and lower loss (up to 6.1% reduction) with minimal storage and computational overhead. We further evaluate DyPP’s effectiveness in VQE for molecular ground-state energy estimation and in QAOA for graph MaxCut. Our results show that, on average, DyPP leads to a speedup of up to 3.1× for VQE and 2.91× for QAOA compared to traditional optimization, while requiring up to 3.3× fewer shots (i.e., repeated circuit executions). Even in the presence of hardware noise, DyPP outperforms existing optimization techniques, delivering up to 3.33× speedup and requiring 2.5× fewer shots, thereby enhancing the efficiency of VQAs.