QML Quarterly / Q2 2026

Five papers on noise, structure and the narrowing path to quantum advantage

QML Quarterly
Quantum Machine Learning
Research
Published

2026-07-27

The most important quantum machine learning result of the last quarter may be that several different paths have been more carefully evaluated.

Noise can make a variational circuit easier to train, but also easier to simulate classically. A quantum processor can reproduce its intended kernel with high fidelity, while that kernel still loses to an ordinary radial-basis-function kernel. A fault-tolerant computer can protect its own calculation, but not necessarily the unknown physical system from which it is trying to learn. A larger quantum state space can increase a model’s capacity, but capacity alone is not evidence of a computational advantage.

Perhaps not a super exciting summary if breakthroughs were expected. I read it differently.

QML has spent years accumulating model proposals whose motivation often begins with the exponential dimension of Hilbert space. The five papers I selected from Q2 2026 ask the harder questions that have to come next. What structure makes that space learnable? Which part of a result survives a realistic noise model? Is the classical baseline genuinely competitive? What physical resource changes the learning problem, rather than merely making the circuit larger?

The quarter did not deliver a general-purpose QML breakthrough, but rather something more useful - more clarity in the question of where an advantage could still hide.

How I made the selection

I treated “last quarter” as the completed calendar quarter from 1 April through 30 June 2026. I included both new preprints and peer-reviewed papers published during that period. Three of the journal papers existed as earlier preprints; their Q2 event is publication, not first disclosure.

For a broad map, I started with Yifan Sun and colleagues’ 47-page survey, Research progress on quantum neural networks and quantum machine learning, whose bibliography contains 276 references. It covers the main QNN families, quantum kernels, training methods, applications and the relationship between barren plateaus and classical simulability.

For higher-recall discovery I also used the Quantum Articles QML archive, which currently contains roughly four thousand records gathered from arXiv, Crossref and other indexes, and the year-by-year Roadmap to QML bibliography. The large archive is useful as a search surface, but its automated category is noisy enough to include adjacent and occasionally irrelevant work. I therefore treated neither index as an authority and checked every selected paper against arXiv or the publisher.

There has not been enough time for citation counts to say much. “Important” here is an editorial judgement based on four tests:

  1. Does the paper change how I would evaluate other QML work?

  2. Does it discuss a well defined problem in QML rather than add another model variant?

  3. Does it combine a strong claim with equally strong assumptions or evidence?

  4. Do its ideas seem reusable beyond the particular experiment?

That process produced this set:

Paper Q2 event What it changes
Noise-induced shallow circuits and the absence of barren plateaus 2 April, Nature Physics Separates trainability from computational usefulness
Benchmarking Quantum Kernel Support Vector Machines Against Classical Baselines on Tabular Data 20 April, arXiv Shows that faithful hardware execution does not repair a poor inductive bias
Provable and scalable quantum Gaussian processes for quantum learning 30 April, arXiv Replaces generic expressivity with physics-informed Bayesian structure
Noisy quantum learning theory 29 May, Nature Communications Moves noise from the processor into the learner’s interface with nature
Multiple photons enhance quantum machine learning 19 June, npj Quantum Information Identifies photon number as a concrete, testable learning resource

1. A trainable noisy circuit may only be an effectively shallow circuit

One of the standard problems in variational QML is the barren plateau. As a circuit grows, the gradients of its loss can become exponentially small, leaving the optimizer with an almost flat landscape. A result that removes a barren plateau therefore sounds like good news.

Antonio Anna Mele and colleagues show why that conclusion can be too quick.

Their paper studies random quantum circuits under local, uncorrected noise. The noise may be unital, such as depolarizing noise, or non-unital, such as amplitude damping. The latter matters because real devices are not always well represented by a noise model that simply pushes states towards the maximally mixed state.

For estimating observable expectation values, the authors prove that the influence of earlier gates decays exponentially with their distance from the output. With high probability over the circuit ensemble, only the final \(O(\log n)\) layers materially affect the result. An arbitrarily deep noisy circuit therefore behaves, for this task, like a logarithmically shallow one.

Non-unital noise produces a superficially surprising second result. For cost functions built from local observables, the gradients do not exhibit the usual barren plateau at arbitrary depth. But this does not make the whole circuit trainable. Parameters far from the output have negligible influence; only the effective final layers matter.

The same effective shallowness gives the authors a classical estimation algorithm. For constant additive accuracy, local observable expectation values of typical noisy circuits can be estimated efficiently classically, in any circuit architecture. In one-dimensional geometrically local circuits, their stronger algorithm also handles inverse-polynomial accuracy in polynomial time.

The useful distinction is:

A loss landscape can have visible gradients because noise has erased most of the computation that those gradients might have trained.

In this setting, noise fixes one optimization symptom by reducing the circuit’s effective computational depth. Trainability and quantum usefulness move in opposite directions.

The scope matters. These are average-case statements for well-motivated random-circuit ensembles, not worst-case theorems covering every hand-designed circuit. The main task is expectation-value estimation, not sampling. Special circuits can exploit non-unital noise, and the paper explicitly leaves more structured circuits, correlated noise and measurement-and-feedforward architectures open.

Even with those limits, this is an important filter for future QML claims. A paper that reports healthier gradients under noise now has another obligation: show that the surviving model is doing something a shallow classical calculation cannot.

2. The quantum kernel reached the hardware and still lost to the baseline

Quantum-kernel papers often leave two questions tangled together.

The first is whether current hardware can estimate the intended kernel accurately. The second is whether the kernel is useful for the learning problem. If a quantum model underperforms in simulation, hardware noise is not the explanation.

Siavash Kakavand, Christoph Strohmeyer and Michael Schlotter make an attempt at pulling them apart with this paper.

Their preprint benchmarks four quantum feature maps and three classical SVM kernels on nine binary tabular datasets. The study includes 970 experiments, nested cross-validation, statistical tests, learning curves, spectral analysis, seed-sensitivity checks and six small hardware experiments on IBM’s 156-qubit ibm_fez Heron processor.

The headline is negative. None of the 29 quantum–classical pairwise comparisons in the main analysis is statistically significant at \(\alpha=0.05\). A larger seed analysis finds the classical kernel better on six of seven tested datasets. One dataset, Haberman survival, shows a modest and repeatable quantum preference, but not the kind of broad result on which to build an advantage claim.

However, the most interesting part are not the benchmark numbers. It is the attempt to explain it.

The quantum kernel matrices tended towards one of two spectral extremes. Some looked close to the identity, making most examples almost equally dissimilar. Others approached rank one, collapsing distinctions between examples. The tuned classical RBF kernel occupied an intermediate region - enough effective dimensions to separate the data, but enough concentration to capture shared structure.

This turns a benchmark failure into a design criterion. The objective is not to maximize the dimension of the quantum feature space. It is to produce a kernel spectrum matched to the data.

The hardware runs sharpen that point. Across six experiments using three to five qubits, the hardware kernel correlated with the ideal kernel at \(r \geq 0.976\). At this scale, the processor did a faithful job of producing the object it was asked to produce. The learning problem was that the object itself was inferior to the classical baseline.

Training the quantum kernel through kernel-target alignment produced the one nearly competitive result: balanced accuracy of 0.968 on the breast-cancer dataset, compared with a classical ceiling of 0.976. But in this implementation the authors estimate roughly 2000 times the computational overhead, report convergence in only 13.6% of configurations, and find considerable parameter instability.

There are good reasons not to universalize the result. This is a preprint. The datasets are standard, small-to-medium tabular problems rather than domains designed around quantum structure. The hardware validation uses only three to five qubits and 60 examples per experiment. With five paired folds, the main two-sided Wilcoxon test cannot attain \(p<0.05\)—its minimum is 0.0625—so “none significant” is not decisive by itself. The later seed analysis reuses the same datasets and only partly resolves that limitation. Four feature maps do not represent every possible quantum kernel.

Still, this is exactly the kind of negative result the field needs. It does not merely say that current quantum kernels lose. It suggests why they lose, verifies that the conclusion reaches the tested hardware, and sets the bar higher for what needs to be proven in the future.

3. Quantum Gaussian processes put the prior before the circuit

Much of QML borrows the architecture-first habit of deep learning: choose a flexible parameterized circuit and ask optimization to discover a useful representation.

Jonas Jäger and colleagues reverse the order. Their quantum Gaussian process framework begins with what is already known about an unknown quantum transformation—its symmetry group, dynamics and relevant observables—and turns that structure into a Bayesian prior.

A Gaussian process is a distribution over functions. Its kernel encodes which function values should be correlated before the learner sees observations. After data arrive, the posterior gives both a prediction and uncertainty. That uncertainty can drive active learning or Bayesian optimization, directing scarce queries towards the places where they are most useful.

In the quantum version, the function maps an input quantum state to an expectation value after an unknown unitary process. The main theoretical result identifies conditions under which these random functions become Gaussian processes and constructs a scalable family for matchgate, or free-fermionic, evolutions. Unlike earlier examples restricted to a logarithmic-size subset of qubits, the unknown matchgate transformation can act non-trivially on the full system.

The paper then uses the framework for several numerical demonstrations. A correctly chosen matchgate prior extrapolates a signal far beyond a narrow training interval, while a generic unitary prior loses the long-range correlations. The authors reconstruct and classify many-body phase diagrams with one or two carefully placed training states, including simulations at 50 and 100 qubits. In a 55-qubit quantum-sensing example, a QGP-guided Bayesian optimizer locates the most sensitive probe state in seven objective evaluations.

The result I find most valuable is not any one of those numbers. It is the controlled comparison between the right and wrong priors. The model generalizes because it is told which physical transformations and observables matter. The prior is not decorative background knowledge; it is the learning mechanism.

This paper also needs careful wording. “Scalable” means that the relevant quantities can be estimated with polynomially many measurements under the stated matchgate and observable conditions. It is not a claim of a general quantum computational advantage. Matchgate dynamics are themselves a famously structured and often classically tractable family. The demonstrations are simulations rather than hardware experiments, and some examples extend beyond the regime directly covered by the Gaussian-process theorem.

QGPs look most compelling as a way to learn from quantum data when experiments are expensive, uncertainty is useful, and physics supplies a real inductive bias. That may be a narrower target than generic classification of classical data. It is also a more credible one.

4. Fault tolerance protection of the thing being learned

Discussions of noisy QML often assume a clean future transition: today’s NISQ devices are noisy, tomorrow’s fault-tolerant machines will run the ideal algorithms.

Jordan Cotler, Weiyuan Gong and Ishaan Kannan point out a missing interface.

A fault-tolerant quantum computer can encode and protect its internal logical qubits. But a learning experiment may query an unknown material, state or process supplied by nature. That external system does not arrive inside the computer’s error-correcting code. Coupling it to the protected computation can introduce errors precisely where the unknown quantum information enters the learner.

The authors formalize this with a complexity class they call NBQP, or noisy BQP. The computer is powerful enough for fault-tolerant computation, but its access to an unknown quantum oracle remains noisy.

This changes several ideal learning separations. For generic purity testing, an ideal two-copy quantum protocol can give an exponential improvement over single-copy measurements. The paper shows that a single layer of local depolarizing noise can destroy that advantage. The Bell measurements and SWAP-like primitives that create the ideal separation depend on high-weight quantum correlations, which are also especially vulnerable to local noise.

The paper does not conclude that every advantage disappears. It proves a superpolynomial oracle separation between NISQ and noisy fault-tolerant learners, so fault tolerance still matters. It also finds a noise-dependent improvement for Pauli shadow tomography and constructs a toy holographic setting in which latent error-correcting structure protects enough information to restore the purity-testing advantage.

That combination is the point. The surviving gains are conditional on the noise rate, memory, task and physical structure. A fault-tolerant processor is necessary for some of them, but it is not sufficient by itself.

This is a theory paper. Its strongest separations live in oracle models; the restored purity example uses a HaPPY tensor network motivated by holography; and depolarizing noise is an analytical model rather than a complete account of an experimental interface.

It nevertheless changes the question I would ask of fault-tolerant QML. “Can the computer run the learning algorithm?” is incomplete. The relevant question is, “Can the entire path from the unknown system, through the sensor and into the logical computation, preserve the information on which the advantage depends?”

Mele and colleagues show noise truncating a variational computation. Cotler and colleagues show noise damaging the coherent access assumed by the learning problem itself. These are different failure modes, and error correction only directly addresses one of them.

5. In photonic QML, photon number becomes a learning resource

The final paper offers a more constructive ending.

Yong Wang, Zhenghao Yin, Tobias Haug and colleagues ask a platform-specific question: what changes when a photonic learning model uses multiple photons instead of one?

For parameterized linear-optical circuits, they measure learning capacity using the rank of the data quantum Fisher information matrix. Informally, this rank counts the independent parameter directions through which the model can respond to the training data. They prove that the capacity grows polynomially with photon number, giving a multi-photon model more learnable directions within the same optical architecture.

The authors connect that geometric result to generalization. In a unitary-learning task, their theory predicts that the critical number of training states falls as the photon count rises. On a programmable integrated photonic circuit, two-photon inputs learned a \(5 \times 5\) unitary from two training states; the single-photon version required four.

They also study metric learning on a seven-class vowel dataset encoded into six optical modes. Simulations compare one through five photons over 100 random initializations. The hardware experiment compares one- and two-photon inputs under the same circuit and training procedure. The two-photon model reaches lower training and test loss and produces clearer within-class blocks in the learned distance matrix.

This is not a proof that photonic QML beats the best classical learner. The comparison is primarily between single- and multi-photon versions of the same linear-optical model. The authors explicitly distinguish their learning-capacity result from the classical hardness of boson sampling, and note that approximate classical simulation remains relevant. The experiment is small, uses product-state inputs and compares at most two photons on hardware. The publisher also labels the available version as an unedited early-access manuscript.

What makes the result important is its specificity. “More quantum” becomes an experimentally controlled resource statement:

Holding the architecture and learning protocol fixed, increasing photon number enlarges the accessible learning directions and reduces the data needed in the tested tasks.

That is a much better foundation for scaling than an appeal to Hilbert-space dimension alone. It suggests a programme in which QML capacity is tied to physical resources—photon number, modes, detectors and state preparation—and then tested against data efficiency, loss and classical cost.

What the five papers say together

Across very different methods, the same distinction keeps returning: representation size is not the scarce resource. Useful structure is.

The noisy-circuit paper shows that visible gradients do not guarantee a deep computation. The kernel benchmark shows that accurate quantum execution does not guarantee a useful feature geometry. The Gaussian-process paper obtains generalization by putting symmetry into the prior. Noisy quantum learning theory shows that an ideal separation survives only when the physical information is robust at the learner’s boundary. The photonic experiment ties increased capacity to a concrete resource and a measured reduction in training data.

I take three lessons from the quarter.

First, trainability is not a proof of advantage. A model can be easy to optimize because noise or architecture has made it effectively simple.

Second, hardware fidelity and model quality are separate axes. Better devices will make it possible to run larger versions of an algorithm. They will not supply the missing inductive bias.

Third, quantum data currently offers a more natural target than generic classical data. The most convincing positive stories in this set concern learning unknown quantum dynamics, optimizing a sensing experiment and exploiting a native photonic resource. The weakest setting is the familiar one in which small classical tables are compressed into a few qubits and sent through a generic feature map.

That last point is an inference from this selection, not a theorem about the whole field. But it is consistent with the practical accounting. Classical data has to be encoded, quantum outputs have to be estimated from shots, and the whole pipeline must beat mature classical models. When the data and the phenomenon are already quantum, the device can contribute information that is not merely an expensive re-expression of a classical input.

What I would watch in Q3

The next useful results will not necessarily have higher accuracy. I would look for four kinds of follow-up:

  1. Kernel designs tested before deployment. hypothesis should be evaluated across more datasets and against problem-informed or symmetry-aware feature maps.
  2. Noise theory connected to device traces. Both noise papers use deliberately general or analytically clean models. The next step is to identify the same mechanisms in calibrated, time-dependent hardware noise.
  3. Quantum Gaussian processes beyond matchgates. The framework becomes more consequential if other physically important dynamics yield estimable kernels, especially on hardware and in regimes that are not already easy to simulate classically.
  4. Photonic scaling with a full cost evaluation. More photons should be tested alongside source rates, loss, detector requirements, shot counts, training time and competitive classical baselines.

More generally, I would like QML papers to report four budgets as routinely as they report accuracy: data-loading cost, circuit or query count, measurement shots and classical-comparison cost. An advantage that appears only after one of those columns is omitted is unlikely to survive contact with an application.

The quarter in one sentence

Q2 2026 did not show that quantum machine learning works in general. It made “in general” a less acceptable way to talk about the field.

The promising results here have come with nouns attached: matchgate symmetries, local observables, Pauli tomography, photon number, kernel spectra, sensing queries. The negative results come with equally useful boundaries.

That is progress. A maturing field should replace broad possibility claims with conditional statements that can be proved, measured and eventually engineered.

The path to a QML advantage looks narrower and less full of weeds after this quarter.

Papers and bibliography