The Signal
Phasecraft AI Agents Cut Costs Across Three Quantum Benchmarks
Researchers report lower measured costs across three quantum-simulation compilation tasks using AI agents paired with frozen verifiers. The results concern benchmark resource counts, not demonstrations on quantum hardware.
Quantum.world Team · Written with AI · fact-checked ·
“In each, the agents improved substantially on published or standard constructions, reducing the cost being optimized by up to 73% at the largest sizes tested and outperforming a state-of-the-art optimizer for fermion-to-qubit encodings.”
Marcos Crichigno and Steve Flammia, the paper’s authors, write.
Phasecraft researchers Marcos Crichigno and Steve Flammia report gains across three quantum-simulation compilation tasks in a paper submitted to arXiv on October 22. The coding agents, built on Claude models, search for programs that construct lower-cost encodings and circuits. Their results are improvements in measured resource counts, not demonstrations on quantum hardware.
Frozen verifiers set the rules
The method pairs an agent’s search with a verifier that checks whether each proposed solution is correct. Before agents compete, a separate agent builds the challenge and evaluator with a researcher; the verifier must pass tests on a solved case, deliberately broken solutions and published results. The researcher approves it, and the evaluator is frozen before submissions begin.
That checking step changed the benchmarks as well as the solutions. During validation, the authors found errors in early versions of two evaluators, including one that could have accepted a circuit implementing the wrong unitary at half the cost. They also found that a published fermionic-encoding benchmark was not optimal under the corrected evaluation.
The authors report 19 accepted agent submissions across the three challenges: 15 for fermion-to-qubit encoding, three for Hamiltonian layout and one for unitary synthesis. Each ran in an isolated sandbox with no network access, one CPU, 512 MB of memory and a 30-second limit per instance. A submission counted only if every instance it claimed passed the frozen verifier.
A 225-site encoding cuts Pauli weight
For fermion-to-qubit encoding, the agents searched for ways to represent a fermionic Hamiltonian as Pauli operators while reducing its total Pauli weight. On a 15-by-15 square lattice, the best result scored 10,749, compared with 17,025 for the Jordan–Wigner encoding—a 37% reduction. It also improved on the paper’s state-of-the-art optimizer comparison by 23% at that size.
Pauli weight counts the operators’ non-identity factors and serves as a cost measure for the encoded Hamiltonian. It is not, by itself, a measurement of end-to-end runtime or the performance of a complete algorithm on a device. The paper also notes that total weight and maximum individual weight can trade off: a single encoding did not hold both records.
SWAP savings depend on the grid
The Hamiltonian-layout challenge measures how many SWAP gates are needed to bring interacting qubits next to one another on a grid. Against a standard swap network, the best submissions cut the count by 90% on 3-by-3 grids and 73% on 8-by-8 grids. At 8-by-8, the best submission averaged 552 SWAPs across nine instances, versus 2,016 for the reference network.
Those savings do not carry over to every circuit metric. The reference network still had the lowest circuit depth on grids from 4-by-4 upward because the agents optimized the number of SWAPs, not depth. The authors also caution that the reference does not use the interaction graph’s structure, leaving room for graph-aware routing methods to improve on it.
CNOT gains come with a narrow test
For unitary synthesis, the agents reduced CNOT counts in phase-polynomial circuits by up to 66% against standard synthesis. At eight qubits, one diagonal target required 23 CNOTs rather than 46; a target with an added linear map required 25 rather than 73. Exhaustive searches established that the submitted circuits were optimal for the tested cases up to four qubits.
The eight-qubit task contains one target per size, so the authors describe its results as per-instance constructions, rather than a general synthesis algorithm. More broadly, their method applies only when a solution can be checked exactly and efficiently on a classical computer and the score is deterministic. The reported gains are specific to Pauli weight, SWAP count and CNOT count under those benchmark conditions.