Researchers Ernesto Acosta and Carlos Cano of the University of Granada, together with Guillermo Botella of Complutense University of Madrid, have reformulated the parameter-optimization problem of variational quantum algorithms as a Quadratic Unconstrained Binary Optimization problem, the native language of quantum annealers. Their work, published as volume 8, article 101 of the journal, demonstrates that the same machines designed to find low-energy configurations of combinatorial optimization problems can also serve as scalable optimizers for gate-based quantum models, opening a hybrid pathway that combines gate-model quantum computing, quantum annealing, and classical control in a single training loop.

The core insight behind the method is elegant. Variational quantum algorithms depend on a Hamiltonian expressed in terms of trainable rotation angles, and the exponential structure of the circuit’s unitary operator can be transformed, through a heuristic substitution that replaces imaginary phases with real-valued angles, into sums of real exponential terms. Those terms map naturally onto the quadratic binary cost functions that quantum annealers minimize. In other words, the mathematical structure of the ansatz itself provides the bridge: instead of estimating gradients with hundreds of circuit shots, the team encodes candidate parameter regions directly into QUBO form and asks the annealer to find the configuration with the lowest energy.

Because continuous rotation angles must be represented with discrete binary variables, the researchers developed a recursive refinement strategy that progressively narrows the search space. The parameter range is partitioned into segments, validation points are sampled within each partition, and the QUBO problem is solved to identify the most promising regions. The angle range is then rescaled — in their worked example, halved at each level — and centered on the best solution found so far, and the process repeats. This coarse-to-fine scheme approximates high-quality continuous solutions with a small number of discrete QUBO solves, trading resolution for tractability in a controlled way.

A key strength of the framework is its adaptability. The method exposes a rich set of configurable metaheuristic parameters, including the number of training levels, the number of partitions per angle, and the number of validation points per partition. These knobs allow practitioners to tune the balance between solution quality and computational cost according to the problem at hand and the computational resources available. The authors also integrate an adaptive metaheuristic optimization scheme rather than a fixed search rule, making the training procedure generalizable to arbitrary Hamiltonians rather than being tied to a specific circuit family or ansatz design.

To test the approach, the team benchmarked their adiabatic training scheme on publicly available classification datasets, including the Iris, Heart Disease, and Diabetes datasets from the UCI machine learning repository. The experimental evaluations show that the method achieves accuracy comparable to, and in some configurations better than, established classical and evolutionary optimizers, while significantly reducing computational overhead. The authors’ hyperparameter exploration reveals that the highest performance efficiency — accuracy gained per unit of training time — tends to occur at lower numbers of levels and partitions, meaning that modest configurations already capture most of the benefit before costs escalate.

The implications extend beyond raw benchmark numbers. Gradient-based training on near-term quantum hardware is plagued by noisy gradient estimates, since every partial derivative must be inferred from finite numbers of noisy circuit measurements. Meanwhile, barren plateaus, whose origins have been linked both to deep random circuits and to noise accumulation, can render gradients exponentially vanishing and classical optimizers effectively blind. A gradient-free, annealer-driven optimizer sidesteps both pathologies: it never estimates a derivative, and its search is driven by the energy landscape of the QUBO encoding rather than by local slope information that may not exist in any meaningful sense.

The study builds on a growing body of work connecting annealing-style hardware to machine learning. Previous research has shown that QUBO formulations can be used to train classical machine learning models, that Ising machines can serve as training engines for standard neural networks, and that universal adiabatic quantum computers can be harnessed for neural network training. An earlier preprint by the same team had already explored adiabatic training for variational quantum algorithms; the new journal publication consolidates that line of inquiry into a general, configurable framework with systematic experimental validation, code released through a public repository, and full experimental detail across multiple datasets and hyperparameter regimes.

What emerges is a compelling vision of near-term quantum computing as a genuinely hybrid enterprise. In the architecture the Granada and Madrid researchers describe, a gate-based quantum processor executes the variational circuit, a quantum annealer solves the training problem at each iteration, and a classical system orchestrates the recursive refinement and manages the workflow. Each platform contributes what it does best, and none has to shoulder the full burden alone. As annealing hardware scales and QUBO encodings grow more expressive, this division of labor could become a practical route to training quantum models that classical optimizers struggle to handle.

Challenges certainly remain. The heuristic substitution that enables real-valued QUBO encoding sacrifices strict unitarity, and the discretization of continuous parameters means solution precision depends on the depth of the refinement schedule. The authors’ own data show that pushing training levels and partitions higher does not always pay off, underscoring the need for careful configuration. Still, the demonstration that quantum annealers can train variational quantum algorithms with competitive quality and lower overhead marks a meaningful step. For a field hunting for any advantage in the noisy intermediate-scale era, recruiting one quantum machine to train another may prove to be one of the more inventive entries in the playbook.

Subject of Research: A QUBO-based method that uses quantum annealers to train variational quantum algorithms without gradients.

Article Title: QUBO-based training for VQAs on quantum annealers

Article References: Acosta, E., Botella, G., & Cano, C. (2026). QUBO-based training for VQAs on quantum annealers. Quantum Machine Intelligence, 8(2), Article 101. https://doi.org/10.1007/s42484-026-00441-1

Image Credits: AI Generated

DOI: 10.1007/s42484-026-00441-1

Keywords: quantum computing, variational quantum algorithms, quantum annealing, QUBO, optimization, barren plateaus, quantum machine learning, hybrid quantum computing, metaheuristics, NISQ era, gradient-free training, Hamiltonians

The challenge arises because quantum computers do not directly reveal a complete quantum state. Instead, researchers prepare a state, run a circuit and measure the result, repeating the process over many “shots” to estimate quantities such as the expectation value of an observable. If a machine-learning model contains many adjustable parameters and processes a large training set over numerous optimization steps, the number of required measurements can quickly become enormous. A conventional QCNN typically reduces the number of active qubits through successive pooling layers, much as a classical convolutional neural network compresses an image while retaining important features. However, the circuit may still require a measurement effort that scales linearly with the original number of qubits. For a model with n qubits, the total training cost can scale approximately as the product of the number of parameters, training examples, optimization epochs and shots per circuit. On hardware where measurements are slow or noisy, that cost can dominate the entire computation.

QCNNs are attractive partly because their hierarchical architecture is relatively shallow. Convolutional layers apply local unitary operations to extract nearby correlations, while pooling layers coarse-grain the information by reducing the number of qubits involved in later stages. Because the active system shrinks rapidly, the circuit depth can scale as order log n rather than growing directly with the number of input qubits. This structure also helps address a notorious problem in variational quantum algorithms known as the barren plateau. In a barren plateau, the optimization landscape becomes exponentially flat as the system grows, causing gradients to become so small that a classical optimizer cannot identify a useful direction for improving the circuit. Local operations, local observables and logarithmic depth allow QCNNs to avoid this failure mode under established conditions. Yet shallow circuits alone do not solve the measurement problem, and adding symmetry creates another architectural complication: pooling can destroy the very spatial relationships the model is meant to respect.

Equivariance provides a way to build those relationships into the model from the beginning. In a symmetry-aware learning problem, the desired output is unchanged when the input is transformed by an allowed operation. A molecular structure, for example, may represent the same physical object after a rotation or inversion, while a lattice system may preserve its label under a translation or reflection. Mathematically, if a density matrix describing the input is represented by ρ and a symmetry operation by a unitary (U_g), the target function obeys (f(rho)=f(U_grho U_g^dagger)) for every symmetry element g. An equivariant circuit imposes a corresponding constraint, requiring its parameterized unitary to commute with the symmetry operation: ([U(theta),U_g]=0). When the final observable is also symmetric, the circuit automatically produces the same prediction for symmetry-related inputs. This reduces the effective space of models the optimizer must explore, potentially improving trainability and generalization by preventing the network from learning irrelevant distinctions.

The difficulty is that ordinary QCNN pooling usually discards selected qubits, and the selection itself can favor one position over another. Earlier symmetry-preserving approaches addressed this problem by randomly choosing which qubits to retain for each measurement shot. For translational symmetry, one shot might retain even-numbered qubits and another odd-numbered qubits, creating a classical mixture of related circuits. The new method takes a different route. Rather than randomly selecting one branch, it splits the circuit into non-overlapping branches and executes them coherently. At each layer, the set of qubits is partitioned into disjoint subsets, with later branches allowed to split from an earlier branch but not merge with another. Each branch receives its own unitary operation, and because the operations act on separate qubits within a layer, they commute. This design makes it possible to impose more general symmetry groups through a group-theoretical construction while preserving the parallel structure needed for efficient measurements.

The measurement advantage follows from the locality of the final observable. Suppose the output is an average of single-qubit observables, (O=(O_1+O_2+cdots+O_n)/n). In a randomized QCNN, each shot effectively samples one subcircuit associated with a particular output qubit, so the expectation value is assembled by averaging results from many separate circuit executions. In the split-parallelizing version, the corresponding subcircuits coexist within the same coherent circuit. In the absence of statistical error, the two procedures produce the same expectation value because each local observable interacts only with the backward light cone—the part of the circuit that could have influenced it. The split architecture can therefore generate as many as n useful measurement outcomes per circuit execution in favorable cases, suggesting an order-n improvement in measurement efficiency. The researchers stress that this is not a universal guarantee. Quantum correlations can make outcomes statistically redundant: in a highly entangled GHZ state, for example, many measurements may carry essentially one bit of independent information rather than n.

The same architecture can accelerate the estimation of gradients, which are required to train a variational circuit. A common technique, the parameter-shift rule, estimates the derivative associated with a parameter by evaluating the circuit at shifted parameter values, typically (theta_mu+pi/4) and (theta_mu-pi/4) in the formulation used by the researchers. In a conventional randomized design, separate branch circuits may be needed for each relevant output, requiring as many as twice the number of branch-associated circuit types for one derivative. In the equivariant sp-QCNN, all terms connected to a parameter can be measured using only the two shifted circuit configurations because the relevant branches run in parallel. Moreover, derivatives associated with parameters in distinct, non-overlapping branches can be measured simultaneously: their corresponding observables act on separate qubit regions and commute. Combining these effects produces an ideal scaling advantage of order n for gradient measurements under the model’s assumptions. That could be particularly important during early training, when repeated gradient evaluations are needed and statistical noise can otherwise slow or destabilize optimization.

The researchers tested the framework on a noisy classification problem involving ground states of the Heisenberg model on a square lattice, a system whose symmetry is relevant to quantum many-body physics. Their numerical experiments found that the equivariant sp-QCNN suppressed statistical error in expectation-value estimates and accelerated training compared with a conventional equivariant QCNN when measurement resources were limited. The symmetry-aware split model also achieved high classification accuracy with fewer training examples than a non-equivariant alternative, consistent with the idea that encoding known structure can improve generalization. The study does not claim that symmetry or efficient measurement automatically delivers a quantum speedup. A circuit that avoids barren plateaus may still be simulable by a classical computer for certain locally simple datasets, and the authors emphasize that classical simulability remains a fundamental issue for many variational quantum models. Instead, the result identifies a practical route toward polynomial improvements in measurement and training efficiency, while extending split-parallelizing QCNNs beyond the translationally symmetric cases considered previously.

The proposal’s significance therefore lies less in a single benchmark than in the combination of three design principles: hierarchical quantum convolution, explicit symmetry and coherent reuse of qubits. Its measurement savings arise primarily from the splitting structure, while symmetry supplies the inductive bias expected to improve trainability and generalization. The absence of barren plateaus is established under a modest but important assumption: each branch, whose structure resembles a conventional QCNN, must itself remain free of the phenomenon, meaning that its local cost-function variance does not vanish exponentially. Likewise, the strongest measurement improvements depend on correlations in the output state; extreme entanglement can reduce the amount of independent information obtained per shot. Future applications could involve quantum materials, lattice models, molecular data and other problems with nontrivial geometric structure, but practical validation on real noisy processors will be essential. For now, the equivariant sp-QCNN offers a technically grounded way to turn symmetry and parallelism into a resource-saving strategy for quantum machine learning at a time when every reliable measurement remains costly.