Language model writes quantum circuit angles—but only when their meaning is clear

Writing parameters instead of training them

Variational quantum algorithms use circuits with adjustable gate angles. A classical optimizer normally changes those angles repeatedly, evaluating how well the circuit performs after each change. On quantum hardware, those evaluations require running the device.

In a November 8 arXiv preprint, Debarshi Kundu, Swaroop Ghosh and Vasant G Honavar of Pennsylvania State University ask whether a large language model can supply useful angles directly. They call this weight-space programming: writing a model’s numerical parameters rather than finding them through iterative training. The supplied record does not establish peer review.

The central finding is conditional. The language model often approached the study’s optimizer when it could reason about what individual parameters did. Its performance deteriorated when the relationship between angles and outputs became harder to interpret.

A test without performance feedback

The researchers report 139 requests to one model, identified in the paper as GPT-6 Astra. Each request supplied a problem, gate conventions and either a fixed circuit or rules for designing one. The returned angles were simulated exactly as written, without sending back accuracy, energy or other performance scores.

All experiments used classical, noise-free simulations of circuits containing at most eight qubits—not quantum hardware. The simulator calculated exact expectations rather than estimating them from repeated measurements.

The tasks covered handwritten-digit classification, energy estimation for small physical systems, and MaxCut, a graph problem that divides vertices into two groups to maximize the weight of edges crossing between them. Comparisons used random parameters, a classical optimizer and the same optimizer initialized with the language model’s parameters. For classification, the main optimizer comparison used the same labelled examples shown to the model.

Strong results on structured problems

The energy experiments used the variational quantum eigensolver, or VQE, which searches for a system’s lowest-energy state. Across eleven problem–circuit combinations, the model’s results were within 0.007 of the optimizer on a normalized energy scale.

Scores ranged from about 0.95 to 1.00. Here, 1 means the exact ground-state energy and 0 means the highest energy; these are not percentages of state fidelity. For the specified two-qubit hydrogen-molecule problem, the model derived the energy’s dependence on a single angle and supplied its exact minimizing value.

For MaxCut, the quantum approximate optimization algorithm, or QAOA, alternates operations representing the objective with operations that mix states. At one layer, the model analytically selected angles that matched the best values found across the six tested graphs. This means optimal angles for those shallow circuits, not exact solutions to every graph’s MaxCut problem.

The parameters also helped subsequent optimization. In both VQE and QAOA, the median number of evaluations needed to reach the study’s target was one: the initial parameters were often already within tolerance. Mean counts were higher—48 for VQE and 36 for QAOA—compared with 607 and 263 from random starts.

Digits reveal the boundary

For distinguishing handwritten 3s from 8s, fixed circuits with 12–18 adjustable angles achieved roughly 93–96% test accuracy, exceeding the optimizer trained on the same ten examples.

Allowed to design circuits, the model repeatedly chose rotations whose angles accumulated weighted pixel values. The unconstrained binary designs were effectively linear classifiers expressed as quantum circuits. Across all ten digits, free-design accuracy averaged about 65%, below roughly 76% for ordinary logistic regression using the same examples.

Deeper fixed circuits exposed a limitation. Binary accuracy fell to about 81% at 30 parameters and 70% at 54, while the optimizer improved. Ten-class performance stayed near chance across the tested fixed circuits, even as optimization reached 57% with 54 parameters. The detailed results therefore do not support a blanket claim that every circuit beyond 30 parameters fell to chance.

Promising initialization, not a general replacement

Convention errors were another weakness: six of 42 QAOA replies reversed mixing-angle signs and performed poorly. Correcting those signs improved results, but the authors counted the original outputs as the findings.

The study tests one model, familiar problems and modest optimizer baselines. Prior exposure during model training may contribute, and performance on unfamiliar systems remains untested. Fewer circuit evaluations also do not establish a wall-clock speedup: model requests took tens to hundreds of seconds, while these small simulations were fast.

The evidence supports a possible source of useful initial parameters for interpretable circuits—not a general replacement for optimization or a demonstration of quantum advantage.