Researchers map 119 studies of AI and quantum optimization

A review of 119 studies reveals a rapidly growing intersection between artificial intelligence and quantum computing aimed at solving complex optimization problems. Researchers identified a detailed analysis of 67 papers, focusing on how machine learning can both assist and be assisted by quantum workflows, an area they term “adaptive” optimization.

Learning reduces quantum evaluations, improves initialization, supports decomposition and penalty control, mitigates noise, and informs execution choices, whereas evidence that quantum computation improves learned optimizers remains largely confined to small-scale simulation. This suggests a key asymmetry in the field as researchers explore how to best allocate classical and quantum resources across the optimization process.

Adaptive Pipelines in Quantum Combinatorial Optimization

Current approaches to quantum combinatorial optimization face limitations imposed by qubit availability, circuit reliability and the costs associated with sampling and calibration, prompting a shift toward adaptable workflows. Rather than relying on static quantum algorithms, researchers are increasingly exploring machine learning to dynamically configure optimization processes, a trend reflected in the growing body of work examining decisions previously fixed in advance. These decisions now encompass not only parameter tuning but also choices regarding problem formulation, decomposition strategies and even whether to use a quantum processor at all.

A central question driving this evolution is no longer simply but rather “which formulation, subproblem, solver, and resource should be used at this point in the search?”, a shift indicating a move toward more control over the optimization process. Evidence suggests a clear asymmetry in the current research landscape. AI demonstrably improves quantum optimization components by reducing evaluations and improving initialization, but the reverse, quantum computation materially improving AI-driven optimizers, remains largely confined to small-scale simulations.

AI-for-QO: Improving Quantum Optimizer Components

Classical optimizers within hybrid quantum loops repeatedly refine angles based on samples returned from quantum processors; each iteration consumes valuable quantum processing unit time and shots, driving much of the current research focus on predicting optimal angles or minimizing evaluations. Beyond simply tuning individual components, some systems are now coordinating multiple decision dimensions or dynamically allocating tasks between classical and quantum solvers as the search progresses; these are termed AI-Quantum Co-Optimization systems.

A key element in these co-optimization systems is the use of weights, λ1 through λ5, to encode the relative importance of factors like solution quality, computation time, quantum access, sampling efficiency and constraint satisfaction. The authors note that a static pipeline represents a fixed action sequence, while most AI-for-QO methods focus on learning a single component of that action, holding others constant, whereas co-optimization learns several interacting components jointly or adapts them sequentially as the solver state evolves, including repeated allocation between classical and quantum actions. “to quantum or not to quantum,” encapsulates the core decision now being made by these adaptive systems.

Q-for-AI: Quantum Layers Within Learned Models

While most research focuses on using artificial intelligence to improve quantum optimization, a smaller but growing body of work explores embedding quantum layers directly within learned models, a strategy termed Q-for-AI. Thirteen of fifteen coded studies in this area remain at early stages of maturity, suggesting the field is still largely exploratory despite promising initial results. These investigations center on whether quantum models can enhance learned optimizers.

Existing Q-for-AI studies demonstrate that quantum policies can encode symmetry within a problem, reduce the number of parameters needed for a solution, and generalize across limited distributions of problem instances. Nine of the fifteen Q-for-AI studies include a classical baseline. However, only five studies replace a quantum layer with a classical one, and even fewer, just two, execute their models on actual quantum hardware. Of the 52 studies analyzed across all three paradigms, 22 lack any classical baseline, and 18 only compare against established heuristics or exact solvers.

AI-Quantum Co-Optimization: Allocating Classical and Quantum Work

Recent advances categorize AI’s role with quantum optimization not simply as assistance, but through increasingly integrated systems that dynamically allocate tasks between classical and quantum resources. This review identifies three distinct approaches, AI for quantum optimization, quantum for AI-driven optimization and AI-quantum co-optimization, each making unique empirical claims about where learning and quantum computation reside within the workflow. The scope of learned decisions varies, ranging from fixed circuit parameter tuning to more complex choices involving circuit structure, problem formulation and even the selection of solvers.

Coding the evidence core on fixed categories also shows thin experimental controls: of the 52 studies in the three paradigms, 22 include no classical baseline, the quantum contribution is fully isolated in 10 of the 24 where an ablation applies, and the median of the largest experiment reported per study is 16 qubits. The researchers introduce an M0-M5 evidence hierarchy that separates simulation, generalization, hardware validation, end-to-end execution and practical quantum advantage under matched resources. The researchers find no broadly convincing result at the highest level in the screened corpus.

We argue that scaling is increasingly a systems problem. The question is not only whether a problem fits on a quantum processor, but how classical and quantum resources should be allocated across the optimization process.

M0-M5 Evidence Hierarchy for Quantum Advantage

A structured evaluation of 67 core papers reveals a tiered evidence hierarchy for quantum advantage, ranging from simulation (M0) to practical advantage with matched classical baselines (M5). This framework, detailed in a review of 119 studies, assesses the strength of claims made at each stage of development, highlighting a current imbalance in the field. The researchers found no broadly convincing results at the M5 level within the analyzed corpus, indicating a gap between theoretical potential and demonstrated real-world benefit.

This allows for a focused comparison of evidence strength across these paradigms, revealing that the largest volume of hardware-grounded evidence currently resides on the AI-for-quantum optimization side. These gains are valuable even without achieving full quantum advantage, as they improve existing quantum routines. 22 of those 27 AI-for-QO studies did not include a classical solver as a comparator, focusing instead on improvements to the quantum process itself.

The review emphasizes that simply achieving hardware execution or outperforming another quantum method does not establish practical advantage; a rigorous comparison against optimized classical baselines is essential. The M0-M5 hierarchy is not a ranking of scientific quality, but a record of the strongest evidence demonstrated by each study, allowing for a detailed understanding of progress across both theoretical and engineering work. The framework acknowledges that a theoretical paper may remain at M1 while an engineering integration study reaches M3 or M4, reflecting different stages of development and validation.

Limitations of Experimental Controls in Hybrid Approaches

Experimental controls within the 67-paper evidence core reveal a frequent omission of classical baselines. This lack of comparison hinders robust assessment of whether observed improvements stem from the quantum component or the learned policies themselves. When ablation studies were performed, isolating the quantum contribution, it was fully separated from the classical system in 10 of 24 cases, limiting understanding of true interaction.

Thirteen of 15 coded studies remain at the lower end of the evidence hierarchy, classified as M1 or M2. The variational loop common to algorithms like QAOA illustrates the interplay of classical and quantum components, where quantum processors prepare and measure states.

Near-Term Quantum Methods for Combinatorial Problems

Combinatorial optimization approaches frequently rely on translating problems into either quadratic unconstrained binary optimization (QUBO) or Ising models, providing a standardized format for quantum processing. The QUBO problem, defined as minimizing a quadratic function of binary variables, allows for a common language between classical and quantum solvers, with the equivalent Ising model using spin variables instead. This translation is an important first step, influencing the efficiency of subsequent quantum computations and highlighting the importance of effective problem formulation.

Of these, 67 papers formed the basis for detailed coding, allowing for a subtle assessment of progress beyond simple publication counts. This suggests a need for further investigation into how quantum resources can genuinely accelerate AI algorithms, rather than merely serving as a sophisticated classical simulator. The open problems identified are the steps needed to answer the question with evidence at the highest levels of rigor, rather than relying on preliminary findings.

Study Corpus: AI and Quantum Optimization Review

The analytical core of this review consists of 52 studies assigned to three paradigms of AI and quantum optimization, with an additional four studies serving as structural precursors and one benchmark study included in the coding process. Researchers meticulously coded 57 non-survey studies from full texts, excluding one study from the evidence core due to unavailability of its complete text. This rigorous approach distinguishes between studies where information was absent and those simply not checked.

Detailed analysis revealed that the quantum component demonstrably altered the configuration or operation of a quantum optimizer in 10 of the 24 studies where ablation testing was applied, a metric of material change. However, only 14 of 52 studies reported the number of quantum evaluations or calls to the quantum processing unit, limiting the ability to assess computational cost.

Most current pipelines still submit every eligible subproblem to the quantum processor, a practice that contrasts with the potential for more subtle, adaptive workflows. “Instance-level solver selection already allows ex-ante rejection of quantum execution,” the authors note, pointing to a current capability for optimizing resource allocation. This suggests a need for further development of systems that can dynamically assess the value of quantum computation during the optimization process, rather than relying on pre-determined strategies.

Scaling as a Systems Problem in Quantum Optimization

A detailed analysis of 119 studies reveals a shift in quantum optimization research, moving beyond simply identifying algorithms to address the broader challenge of resource allocation across the entire optimization process. This suggests the field is increasingly recognizing that scaling quantum optimization isn’t just about fitting a problem onto a quantum processor, but intelligently distributing classical and quantum resources. O’Leary et al. demonstrated this approach with an online surrogate learning technique, achieving query-efficient optimization on a 127-qubit Ising problem and transferring learned parameters to even larger systems, primarily reducing the number of quantum evaluations needed.

More like this

Quantum Computing NewsReinforcement learning cuts gate count for 4-qubit states

Quantum Research NewsQuantum memory gets a boost from encoding data more efficiently

Quantum Research NewsAnthropic’s Claude improves quantum optimization designs

Quantum PhysicsSparse Hamiltonians can be learned with slower quantum controls