Willow wins 98.3% in quantum magic-square game, Google says


Google Quantum AI says its processor beat the 88.9% ceiling for classical strategies in a game where players cannot communicate. Other tests explored how far the effect could scale and whether it could expose hardware errors.
Google Quantum AI and academic collaborators say the Willow quantum processor won a four-qubit magic-square game 98.3% of the time, with a reported uncertainty of 0.01 percentage points. The result exceeded the best classical strategy allowed by the game’s rules, offering a way to test quantum hardware against a defined limit. A qubit is a unit of quantum information.
The team published its study in Science after making the results public in a preprint on arXiv in December 2025, Numerama said. The researchers also tested a larger game and a mathematical task, though the classical limit for the latter at the size tested remains a conjecture.
In the magic-square game, a referee gives one player a row and another a column from a three-by-three grid. Each independently fills the assigned spaces with +1 or −1. The row must have an even number of minus signs and the column an odd number; the players win if their answers match where the row and column meet. They cannot communicate or know each other’s assignment.
A fixed grid of answers cannot satisfy both rules: counting minus signs across the rows gives an even total, while counting them down the columns gives an odd total. A classical strategy can therefore win at most eight of the nine equally likely assignments.
The experiment instead gave each player one qubit from each of two entangled pairs. Entanglement links the results of measurements on the qubits. Each player measured their qubits according to the assigned row or column; in an ideal system, their answers would always agree at the intersection. Google ran each assignment 100,000 times.

The study examined quantum contextuality: the result of a quantum measurement can depend on which other compatible measurements are made alongside it. In a separate test of this property, Google measured a score of 5.59, with an uncertainty of 0.008. A model with fixed, context-independent measurement values could score at most 4; the ideal quantum score is 6.
In a separate game, each player received and returned a binary digit without communicating, and the team had to produce a combined answer that followed a rule. Google said Willow stayed above the classical winning limit with up to 71 players, one per qubit, although the margin narrowed as qubits were added. A separate analysis verified genuine group entanglement across as many as 45 qubits; that measured a different property from the game’s winning rate.
Willow also tackled a two-dimensional hidden-linear-function problem, returning valid answers more than half the time with up to 105 qubits. The chip had to produce a string of binary digits satisfying a set of mathematical relationships. Researchers encoded each input table in a grid of qubits, ran four layers of operations linking neighboring qubits and checked the answers using a classical simulation.
A 2018 theorem established a quantum advantage for sufficiently large versions of that problem under restrictions on the classical circuit. Unlike the games’ winning ceilings, the team’s estimate of the classical requirement at the experimental scale remains unproven. Establishing a bound at that scale would let researchers determine how much of the predicted advantage survives hardware errors.
The 2018 theorem compares a constant-depth quantum circuit with a classical probabilistic circuit whose gates have bounded fan-in and whose required depth grows with the problem size. According to the Google study, the earlier proof gives a meaningful comparison at tens of thousands of qubits; the classical bound for the 105-qubit experiment remains a conjecture. Numerama reports that Google acknowledged the work has no immediate application.
Analysis
Google said the experiments exposed subtle errors that can shift a quantum state slightly off target and can be averaged away in tests using random operations. The researchers propose using the games and mathematical task to track improvements in future hardware.
The Polish actor agreed to have his new recordings treated to sound more like his earlier performance after a stroke changed his voice, Game Developer said on 13 November. The expansion is planned for 2027.
2 min

Minister Philippe Baptiste presented the strategy on 16 November 2026, setting a goal of having the technology needed for a European fusion pilot plant by 2035.
3 min

The company says it will not make the proposed investment in Texas without an agreement and is considering sites elsewhere. The state and a local school board must approve the tax break.
3 min