There is a puzzle in which you fill a 3-by-3 grid with +1 or -1 so that the product of every row is +1 and the product of every column is -1. Try it with pencil and paper and you will find that no solution exists. The row conditions require an even number of -1 entries overall, while the column conditions require an odd number. Google Quantum AI researchers turned this puzzle into a game that exploits quantum mechanics and ran it on their superconducting quantum processor, Willow. They recorded a win rate of 98.30%, far above the 88.89% ceiling for classical strategies.
What makes the difference is a property called quantum contextuality, in which the outcome of a measurement cannot be fixed independently of the other measurements performed alongside it. The study, published in Science on November 19, 2026, used experiments with up to 105 qubits to test whether this property can be observed in large quantum systems. It also examined whether contextuality can serve as a new metric for evaluating the performance and errors of quantum processors.
Solving an Unsolvable Puzzle on a Quantum Computer: A 98.30% Win Rate
The puzzle is recast as a cooperative game for two players. In each round, Alice is assigned a row and Bob a column at random. Without consulting each other, each fills in +1 or -1 for the three cells in their line. Alice must choose values whose product is +1, and Bob must choose values whose product is -1. They win if their entries agree in the cell where the row and column intersect. Whatever answer table they prepare in advance, at least one of the nine possible row-column combinations must fail, so the best classical strategy is capped at 8/9, or 88.89%.
Quantum mechanics can break through this limit. In the paper, submitted on September 16, 2025, accepted on September 4, 2026, and published on November 19 in Science (Vol. 394, pp. 870–874), the research team ran the "Mermin-Peres magic square game" on the real Willow hardware. What Alice and Bob share in advance is not an answer table but two pairs of entangled qubits known as Bell pairs (four qubits in total). Each follows the measurement procedure corresponding to their assigned row or column, measures the two qubits they hold, and obtains a result of +1 or -1.
In an ideal, noise-free quantum system, Alice and Bob could win every combination without communicating after receiving their row and column assignments. On the actual device, the team ran each of the nine combinations 100,000 times and evaluated the row and column product conditions and the agreement at the intersecting cell. The average win rate came to 0.9830 ± 0.0001 (98.30%). The error margin reflects statistical uncertainty based on the binomial distribution.
The key point is that results are not obtained from fixed values pre-assigned to each cell, but depend on the "context" of which row or column is being measured. Thanks to entanglement, Alice and Bob can make their measurement results for the intersecting cell agree with high probability. This confirms, on a quantum processor, correlations that a classical answer table cannot reproduce.
Quantifying Contextuality: 5.590 Against a Classical Limit of 4
Beyond the game's win rate, the team also tested a Kochen-Specker-Bell (KSB) inequality to quantify how much contextuality the quantum processor exhibits. The comparison target is the "noncontextual hidden-variable model," which assumes that measurement outcomes are determined in advance and do not depend on which other measurements are made alongside them. In this model, each of the nine observables is assigned a fixed value of +1 or -1, regardless of which row or column it is measured with.
The evaluation used the following statistic:
Here, and represent the product of measurement outcomes for each row and each column, and the angle brackets denote averages computed from many measurements. Noncontextual models are bound by , whereas in quantum mechanics every row product is +1 and every column product is -1, so the statistic can reach a maximum of 6. This inequality is "state-independent": it does not depend on how the state is prepared, as long as all six terms are evaluated under the same initial state preparation conditions.
The experimental difficulty is that three measurements must be performed in succession on the same two data qubits, with the results remaining usable for subsequent measurements. The team used auxiliary ancilla qubits and implemented quantum nondemolition (QND) measurements, which obtain results via the ancilla rather than reading out the data qubits directly. This made it possible to perform consecutive measurements on the same data qubits.
In the experiment, the team performed three measurements for each of 180 randomly chosen measurement contexts, for a total of 540 QND measurements in sequence. The order of measurements within each context was also randomized, and the whole sequence was repeated 1,000 times. The official Google Research explainer, published on November 19, reports a result of .
The publicly available arXiv preprint, on the other hand, lists a value of . Although the two figures differ, both far exceed the noncontextual limit of 4 and come close to the quantum-mechanical maximum of 6. The gap from the maximum is attributed to errors in the experimental apparatus, and the paper's supplementary materials analyze the contributing factors.
Beating Classical Strategies With Up to 71 Qubits
Confirming contextuality with a small number of qubits does not guarantee it survives at scale, since noise accumulates as qubit counts grow and quantum correlations can be lost. To test whether contextuality persists in larger quantum systems, the team extended the experiment to a multiplayer competition called the "GHZ parity game."
In this game, players share the following entangled quantum state, known as a GHZ state:
Each player is given a single bit, 0 or 1, under the condition that the sum of all input bits is even. The players return a one-bit answer each without communicating, and they win if the parity of the sum of their answers matches the parity of half the sum of the input bits. With inputs and answers , the winning condition is:
Without quantum mechanics, the optimal classical win rate is given by:
With three players it is 75%, and it approaches 50% as the number of players grows. If the players share an ideal GHZ state, however, each can measure their qubit in the X or Y basis depending on their input, and in theory they always win.
The team assigned one qubit to each player and scaled the experiment up to 71 players, that is, 71 qubits. For each qubit count, they prepared 20 input patterns and 50 GHZ state generation patterns, and ran each of the resulting 1,000 configurations 1,000 times. The win rate exceeded the classical limit at every size measured. However, as the number of qubits grew, noise had a larger effect: the win rate moved away from the ideal 100%, and the margin over the classical strategy narrowed.
Note that the argument that more repetitions raise statistical confidence in the gap from the classical limit assumes that errors are not systematically biased toward the opposite-phase GHZ state . Separately from the 71-player win rate, the team also confirmed genuine multipartite entanglement across 45 qubits.
A 105-Qubit Circuit Experiment, With Caveats on Comparison to Classical Computation
Following the game-style tests, the team also ran a structured quantum algorithm on up to 105 qubits. The target was the "Hidden Linear Function (HLF) problem."
In 2018, Bravyi, Gosset, and König proved in a paper in Science that for a certain HLF problem, a constant-depth quantum circuit can solve it more efficiently than classical circuits with restrictions on fan-in or depth. However, the quantum-classical gap shown by that theorem applies to computational models with restricted circuit depth and gate structure, and it does not imply an advantage over every classical algorithm.
The HLF problem used in this experiment is defined by a symmetric binary matrix representing a quadratic form modulo 4. The task is to find bit strings satisfying a prescribed linear relation on the binary null space of that matrix. The quantum circuit had a shallow structure: Hadamard gate initialization, S gates determined by the matrix, four layers of CZ gates, a final Hadamard layer, and measurement of all qubits.
The team generated 1,000 random matrices for each size from 9 to 105 qubits and performed 100 measurements for each. Whether an output bit string was a correct solution was verified using the classical Clifford circuit simulator Stim. The success rate was nearly 100% for small circuits but fell to roughly 60% at the largest size of 105 qubits.
The experiment also evaluated an "effective depth" that takes the success rate into account. This metric reflects the number of runs needed to obtain a correct solution, and for this circuit it is calculated as four layers divided by the success rate. Whereas an ideal noise-free circuit needs four layers, the 105-qubit experiment rose to the equivalent of about six layers.
When interpreting the 105-qubit result, however, one must distinguish theoretical proof from finite-size experiment. The logarithmic depth lower bound for classical circuits implied by the 2018 theorem becomes meaningful only at extremely large scales, on the order of qubits, and cannot be applied directly to 105 qubits.
To compare at scales closer to current quantum processors, the authors therefore adopted a conjectured value: that a classical circuit using two-input, one-output gates needs a depth of to find the exact solution, where is the number of relevant inputs. The paper itself distinguishes this comparison from a rigorous theorem, and the 105-qubit experiment does not constitute an unconditional proof of quantum advantage on a practical computational problem.
Detecting Errors Random Circuits Miss: A New Hardware Evaluation Metric
Another outcome of the experiments is the potential to use quantum contextuality for hardware diagnostics. In conventional benchmarking with random quantum circuits, errors are averaged out across the whole circuit, which can make subtle coherent errors, such as slight rotations of the quantum state away from its target, hard to see. In the structured tasks used here, by contrast, the measurement results of many qubits must satisfy specific relationships exactly, which brought to the surface errors that are hard to capture with random circuits.
The team proposes using contextuality not merely as a property demonstrating quantum computational advantage but as a metric for quantifying the errors of a quantum processor. The experiments are summarized below.
| Experiment | Scale | Metric | Classical / noncontextual baseline | Ideal quantum maximum | Experimental result |
|---|---|---|---|---|---|
| Magic square game | 4 qubits | Average win rate | 88.89% | 100% | 98.30% ± 0.01 points |
| KSB inequality test | 2 data qubits + 1 ancilla | 4 or less | 6 | 5.590 ± 0.008 (Google's official explainer) | |
| GHZ parity game | Up to 71 qubits | Win rate | 100% | Exceeded classical limit at all sizes | |
| HLF circuit experiment | 9–105 qubits | Success rate, effective circuit depth | (conjectured classical circuit depth) | 100% success, 4 layers | About 60% success and about 6 effective layers at the largest size |
The distinguishing feature of these benchmarks is that they show not only whether a classical limit has been exceeded but also how far the actual measured values fall short of the ideal quantum upper bounds, such as or a 100% win rate. Because the theoretical upper bounds do not change as quantum processors improve, measuring that gap makes it possible to gauge hardware improvement. A Perspective article by Rui Soares Barbosa, published alongside the paper in Science, likewise assesses these experiments as candidates for new hardware benchmarks and positions them as a means of probing the boundary between quantum and classical systems.
To ensure reproducibility, the team has released the experimental plot data on Zenodo and provides the code for circuit generation and analysis through ReCirq. The study does not, however, include independent replication on quantum processors owned by other research groups, so it remains unconfirmed whether the same diagnostic performance would be obtained on different hardware.
Looking ahead to applications such as molecular simulation in chemistry and materials science, the team points to the possibility of using contextuality as a way to evaluate the execution accuracy of structured quantum algorithms. The experiments did not perform any practical molecular simulation, but the ability to detect coherent errors that are hard to capture with random-circuit evaluation widens the range of diagnostic methods available for future quantum processor development.
