IBM shows a problem classical machines can't solve — plus a way to check the result

A team at IBM Quantum and the error-mitigation startup Qedma published three physics calculations on Thursday that classical supercomputers could not reproduce, alongside a verification protocol meant to separate the quantum result from the noise that produces it.

Quantum advantage, the term for a calculation a quantum machine can do and a classical one cannot, has been claimed several times in the past few years. Most of those claims have collapsed under scrutiny, as classical methods caught up or the verification failed. Thursday's result is structured differently: the protocol is published alongside the result, the data is posted to an open benchmark, and independent teams on different hardware checked at least one of the three cases.

The three cases are scientific problems, not benchmark trophies. The first is a 2D magnet hit by laser light. On a chain of 74 superconducting qubits, each one a quantum bit encoded in a superconducting circuit, the team observed long-lived oscillations in the magnet's state that Japan's Fugaku supercomputer and an Nvidia GPU both failed to reproduce. The result is described in a preprint posted this week on arXiv.

The second is an information-scrambling problem inside a material that is not uniform. Scrambling, the way information spreads and becomes unreadable in a complex system, matters for designing catalysts and predicting chemical reactions, because it tells researchers how energy and electrons move through messy materials. A leading classical simulation method broke down for some of the parameters the quantum machine ran.

The third is a sampling problem, a test of the machine's raw ability to produce outputs that classical random-number generators cannot fake. It is a standard diagnostic, but the team wrapped the same verification protocol around it as the other two.

What makes the announcement harder to dismiss is the trust protocol. The team ran each calculation on multiple IBM quantum machines, each with different noise levels, then deliberately injected known noise into the simulation. By comparing the real noise to the injected noise, they estimated how much of the result is signal and how much is artifact. They validated the magnet case on a Quantinuum trapped-ion machine, which uses charged atoms held in electromagnetic fields rather than superconducting circuits. The full data set for the magnet problem is on the Quantum Advantage Tracker, an open benchmark.

"It's a different level of rigor than we usually see," said Dominik Hangleiter of ETH Zurich. Hangleiter cautioned that the protocol is not fully hardware-agnostic: it works on superconducting and trapped-ion machines, but the assumptions may not hold for every quantum architecture. "The verification is convincing within the team's protocol, but not yet universal," he said.

Emanuele Dalla Torre of Bar-Ilan University, who also reviewed the work, called the magnet result "fundamental-science level" but still idealized compared to real-world magnets. The oscillation is real, he said, but the simulated magnet is a model, not a material you can hold.

The calculation runs on IBM's Heron quantum processor through the Qiskit Functions Catalog, IBM's quantum software marketplace, with QESEM, Qedma's error-mitigation software, cleaning up the results. IBM and Qedma framed the announcement as the first advantage demonstration on commercially available quantum hardware combined with error-mitigation software. Researchers from RIKEN in Japan and Quantinuum contributed to validation.

The lasting shift is methodological. New Scientist's writeup of the result notes that several past advantage claims have been "dispelled." Thursday's announcement is structured so other teams can re-run the same protocols on the same hardware family and check. Whether the gap to classical holds for harder problems, and on architectures beyond superconducting and trapped-ion qubits, is the next test.