Key facts
- Google said Willow won the four-qubit magic square game 98.3% of the time, against an 88.9% classical ceiling.
- The 98.3% result came across 900,000 trials, and hardware noise kept it below the quantum ideal, the researchers said.
- Google reported a contextuality score of 5.590 with uncertainty of 0.008; classical models score at most 4.
- Google said up to 71 entangled qubits stayed above the classical winning rate in a GHZ parity game.
- The 105-qubit hidden linear function task reached roughly 60% success, Google said.
- Google's corporate blog said contextuality testing could help engineers catch subtle hardware errors.
Google said its Willow quantum processor won a mathematical coordination game 98.3% of the time, above the 88.9% maximum attainable by classical strategies, in a Science study published Thursday. The experiment tests quantum contextuality, in which measurement outcomes depend on what else is measured alongside them. Google researchers said "you cannot assign fixed, pre-existing values to everything you might measure."
The first test used the Mermin-Peres magic square game, which asks two players to fill intersecting parts of a three-by-three grid without communicating. A prearranged classical strategy can succeed on no more than eight of nine assignments, or 88.9%. Entangled qubits can theoretically win every round, while Google's physical processor won 98.3% across 900,000 trials, according to the researchers. They wrote that "our qubits coordinated better than any pre-written table allows," and said unavoidable hardware noise prevented the result from reaching the quantum ideal.
Google also measured contextuality. Any model with predetermined outcomes scores at most 4; quantum theory allows 6. Willow registered 5.590, with a reported uncertainty of 0.008.
A separate GHZ parity game asked up to 71 entangled qubits to deliver coordinated binary answers. Success was measured against the strongest classical winning rate, which approaches 50% with more players; Google said it remained above that limit throughout its tests despite accumulating hardware noise. Researchers also verified genuine entanglement across 45 qubits.
In the hidden linear function task, a shallow quantum circuit had to produce a binary string satisfying mathematical constraints. The processor reached roughly 60% success using 105 qubits. Google said the tasks exposed small, systematic errors that conventional randomized tests can hide, and its corporate blog called contextuality testing "a sensitive new method to catch subtle hardware errors," potentially helping engineers evaluate circuits used for chemistry and materials research.
The work follows Google's October report of a 40-fold error reduction in preparing magic states, a resource needed for fault-tolerant quantum computing. Google said the 105-qubit benchmark uses a conjectured classical bound and does not establish fault-tolerant computing; its largest instance still failed in about four out of 10 runs.
