IBM Quantum and QuICS Deploy AI Agents to Discover Over 500 High

Researchers from IBM Quantum , the Joint Center for Quantum Information and Computer Science (QuICS / NIST / UMD) , and MIT have developed a neuro-symbolic framework that pairs large language model (LLM) agents with deterministic solvers to map uncharted magic-state distillation protocols. Detailed in a preprint titled Symmetry and AI-assisted discovery of magic-state factories , the team uncovered 699 factory classes—including 564 previously unknown protocols—delivering non-Clifford states with record-low physical resource footprints.

Universal fault-tolerant quantum computers running planar CSS or surface codes cannot execute non-Clifford operations natively through transversal gates alone. Standard architectures rely on distillation factories to purify noisy ancilla states via gate teleportation, a routine historically projected to consume up to 90% of a machine’s physical qubits and runtime. While exhaustive classification historically stalled at small input sizes ( n ≤ 54), the researchers made larger searches tractable by proving that achieving code distance d ≥ 3 is mathematically equivalent to selecting nonzero, pairwise distinct syndromes within a binary-matrix representation.

To navigate the expanded space, the team deployed an automated multi-agent loop: language-model agents propose targeted subgroup symmetry orbits, an integer-programming solver constructs candidate circuit matrices, and an independent verification engine validates gate identities and code distances. A reviewing agent evaluates solver logs to redirect subsequent rounds toward productive geometries. In addition to pure- T states, the framework yielded factories outputting entangled states such as CCZ ⊗2 and T ⊗ CCZ ⊗4 , directly lowering the gate depth needed during algorithm synthesis.

To support community resource estimation, the authors released the verified circuits in an open-access repository indexed with Clifford correction rules and matrix export formats.

The research paper is accessible on arXiv here , with executive commentary shared in Jay Gambetta’s Post here , circuit models hosted on the GitHub Repository here , and the interactive Magic State Factory Catalog here ; related QCR reporting on non-Clifford state generation includes Google Quantum AI’s Willow hardware cultivation demonstration here and NVIDIA and Jülich’s exascale cultivation simulation on JUPITER here .