Researchers from IBM Quantum, the Joint Center for Quantum Information and Computer Science (QuICS, affiliated with NIST and the University of Maryland) and MIT have developed a method that uses artificial intelligence agents to discover new ways of producing so-called magic states — a key ingredient for quantum computers that work with error correction. The work, published as a preprint on arXiv, mapped 699 classes of magic-state factories, 564 of which were unknown until now.
Why magic states are so expensive
In fault-tolerant quantum computers, ordinary operations can be executed directly, but non-Clifford operations — required for universal computing — depend on magic states purified through a process known as distillation. Historically, this process was projected to consume up to 90% of a machine's physical qubits and runtime, according to the Quantum Computing Report. Cutting that cost is one of the biggest bottlenecks on the way to useful large-scale quantum computers.
How the AI agents searched
To explore this space, the team combined large language model agents with deterministic solvers in an automated loop: the agents propose symmetry orbits, an integer-programming solver builds candidate circuit matrices, and an independent verification engine validates gate identities and code distances. A reviewing agent examines the logs and steers later rounds toward the most promising geometries.
Among the highlights, the researchers found a factory that produces 287 pure T states from 1,715 inputs — the smallest known to break the sub-linear overhead barrier, with an exponent below 1 (γ = 0.998). Another configuration, with 850 inputs and 128 T states, recorded the lowest exponent reported for distillation protocols with up to 1,000 inputs. The method also produced factories that output entangled states, such as combinations of T, CS and CCZ gates, removing decomposition steps.
A preprint, not a peer-reviewed result
It is important to note the status of the work: this is a preprint, a paper released before peer review. The verified circuits were released in an open repository, with an interactive catalogue of the factories found, so other researchers can check the results and use them in resource estimates.
The advance connects with IBM's roadmap for fault-tolerant quantum computing, which targets the Starling system, with 200 logical qubits, by 2029 — a plan that depends precisely on lowering the cost of magic-state factories.
Image credit: photograph of the IBM Quantum System One in Ehningen, Germany, by IBM Research, licensed under CC BY 2.0 and available on Wikimedia Commons.
Also on Geekiki: Universal Quantum launches iQPU, a quantum chip built to scale.
Sources: preprint “Symmetry and AI-assisted discovery of magic-state factories” on arXiv; Quantum Computing Report; IBM Quantum blog on the path to fault-tolerant quantum computing.
Transparency: this article was produced with the assistance of artificial intelligence, drawing on public posts and internet sources, including the cited preprint and specialist coverage. We recommend consulting the original sources to confirm the technical details. The results described have not yet undergone peer review.
Muse