Agent-assisted experimental strategy transfer across quantum processors

The gist The agent-assisted experimental strategy can be reused and refined across processors to realize a common quantum task, providing a basis for scaling experimental practice beyond individual devices

Agent-guided framework for strategy transfer

The paper presents an agent-guided framework that separates a transferable experimental strategy from its device-specific physical realization In the fixed-rule feedback scheme shown in Fig. 1a, measurements update predefined control variables The conceptual framework links a scientific target to an experimental result through two stages: target-to-task (T2T) and task-to-experiment (T2E) T2E realizes a supplied task specification on physical hardware The agent invokes and coordinates deterministic tools to characterize the processor, synthesize controls, acquire measurements and reconstruct experimental outputs Operators retain responsibility for instrument preparation, physical adjustment, safety and procedural authorization

Development of the initial strategy

The initial strategy P0 was developed under human supervision on the development processor Q0 This involved estimating local Hamiltonian parameters by fitting resonance lines from x- and y-phase excitation spectra An 11-ms radio-frequency (RF) pulse optimized using GRAPE implemented the state-preparation task defined by two Grover iterations The ideal marked-state population was 121/128 ≃ 0.945 This operator-guided development provided the basis for P0, which defined rules for action selection and stopping These rules specified that unreliable spectral fits required further characterization, while insufficient reference data or inconsistent repeated measurements required additional acquisition before task performance could be assessed

Strategy transfer across four processors

The T2E strategy was deployed sequentially to Q1, Q2, Q4 and Q3 At each processor, the agent assessed local device information and calibration results, then invoked the tools needed for characterization and control-pulse optimization Each deployment implemented the state-preparation task defined by two three-qubit Grover iterations, with a processor-specific marked computational-basis state The agent used experimental evidence to decide whether to collect more data, revise the procedure, request operator assistance or stop Across these deployments, the agent used experimental evidence to decide whether to collect more data, revise the procedure, request operator assistance or stop

Agent-guided repeated measurements

After the initial transfer, repeatability was assessed in subsequent experimental sessions The agent used reference checks to guide whether acquisition could proceed These sessions illustrate how reference checks, interruption recovery and repeated measurements supported continued experimentation while revealing variation in task outputs The marked-state populations of the reconstructed dominant eigenstates spanned 0.929, 0.959, 0.936 and 0.910 for Q1, Q2, Q4 and Q3, respectively

Discussion

The study extends cross-device calibration and control research to the realization of a common multiqubit task across distinct processors With the logical task and experimental interface held common, each deployment rebuilt its Hamiltonian model, measurement references and Grover pulse This setting reveals how the experimental procedure changes when new hardware alters both numerical parameters and the applicability of characterization and analysis methods Our study extends cross-device calibration and control research to the realization of a common multiqubit task across distinct processors

Methods

Experiments were performed on five three-qubit liquid-state NMR processors (TriangulumPro, SpinQ Technology Co., Ltd.) The common logical task was three-qubit Grover state preparation with one marked state among eight computational-basis states The Hamiltonian calibration and pulse optimization used the local three-spin Hamiltonian in the frame rotating at the RF carrier frequency GRAPE optimized an 11-ms pulse comprising 440 piecewise-constant segments against the Grover target unitary Um = (DOm) 2H⊗3 The total-variation distance between the reported and ideal population distributions was DTV(p, pideal) = 1/2 X x px − p ideal x, where x runs over the eight computational-basis states and the ideal distribution is defined by Eq. (1)

Agent implementation and transfer constraints

The agent, identified in the Q0 manifest as Codex based on GPT-5, used Python experimental tools and a locally bundled spinqlink instrument interface Transferable context included the three-spin model class, generic experimental procedures, quality criteria, acquisition budgets, stopping rules and deterministic tools Device-specific numerical parameters, spectra, response matrices, reference shapes and waveforms from Q0 were excluded from the prior information supplied for a new processor

Grover task and ideal output

The logical circuit in Fig. 5 starts from 000⟩ and applies a Hadamard gate to each qubit to prepare s⟩ = H⊗3000⟩ = 8−1/2 Px∈[x≤1]3x⟩ The ideal marked-state population was 121/128 ≃ 0.945

Hamiltonian calibration and pulse compilation

The control Hamiltonian was Hc(t)/ħ = 2π[bx(t) P i I(i)x + by(t) P i I(i)y] GRAPE maximized FU = X δ,s wδ,s Tr(Uδ,sU†target) 8, X δ,s wδ,s = 1

Reconstruction and validation

Tomography using seven measurement bases provided observations of all 63 non-identity Pauli components The total-variation distance between the reported and ideal population distributions was DTV(p, pideal) = 1/2 X x px − p ideal x, where x runs over the eight computational-basis states and the ideal distribution is defined by Eq. (1)

Validation policies and run accounting

Closely agreeing purified marked-state populations are classified as satisfactory, without a numerical cutoff All 15 complete measurements were included; each session is summarized by its mean and range (maximum minus minimum) The agent used reference checks to guide whether acquisition could proceed

Appendix S3: Hamiltonian estimation and pulse optimization

The first-order line-centre model used for resolved multiplets was fj(z) = ∆νj + 1/2 X k≠j Jjkzk, zk ∈ (−1, +1) The Q3 line-centre RMS errors ranged from 0.075 to 0.

1. Adapt experimental procedures based on local evidence: The agent uses experimental evidence to select subsequent actions and adapt the experimental procedure, which allows it to handle device drift by deciding whether the available evidence supports proceeding or calls for further investigation.

2. Implement strategy transfer across processors: The system can realize a common task across devices by using an agent-guided framework where uncertain fits or inconsistent references were followed by additional measurements, and the fitting procedure was adapted to the observed spectral structure.

3. Automate complex experimental decision-making: The agent coordinates decisions at a higher level than a conventional controller, while numerical calculations remain assigned to deterministic tools and physical interventions to operators, enabling autonomous selection of characterization procedures based on local conditions.

Progress in quantum hardware requires a growing volume of experiments across diverse devices, straining the capacity of human experts. Automation remains challenging because device differences and drift can render workflows developed on one processor unsuitable for another. Here we present an agent-assisted experimental strategy for conducting experiments across three-qubit nuclear magnetic resonance processors. Developed on one processor and applied sequentially to four others, the strategy uses an agent to assess experimental evidence, select subsequent measurements and adapt procedures, with operators providing instrument preparation and physical intervention. Using three-qubit Grover state preparation as a common task, each deployment combined device-specific calibration, control-pulse optimization and measurement. The experimental strategy was refined through sequential deployment, and all four processors ultimately produced outputs dominated by their respective marked states. The marked-state populations of the reconstructed dominant eigenstates ranged from 0.910 to 0.959, with relative deviations below 3.8% from the ideal value of 0.945. These results demonstrate that an agent-assisted experimental strategy can be reused and refined across processors to realize a common quantum task, providing a basis for scaling experimental practice beyond individual devices.