On superconducting quantum processors, the connection between two neighboring qubits is often a tunable coupler: a small circuit whose magnetic bias sets how strongly the two qubits interact, all the way down to an exact "off" state. This means a single chip can run many different circuits. Any subset of its couplers can be switched off, for example because a coupler has drifted into a noisy frequency band, and the result is a smaller but still valid circuit. We propose training one set of variational circuit parameters that works for all of these coupler subsets without retraining. The idea borrows the "once-for-all" supernet from neural architecture search, where one shared set of weights serves many sub-networks.
The obstacle is known from classical deep learning. When many sub-networks share weights, their gradients conflict. The shared weights then settle on a compromise, and their ranking of sub-networks stops matching how those sub-networks perform when trained alone. Nobody has measured this conflict in quantum circuits. Our central claim is that continuous couplers offer a remedy fixed-gate hardware cannot. Turning a coupler off does not have to be a hard switch; it can be a smooth dial. Training with a slowly moving, continuously scaled coupler mask should reduce gradient conflict between sub-circuits and restore faithful ranking. That would let a chip swap in a working circuit the moment a coupler goes bad.
Core Claim and What Is New
Consider a fixed brickwork circuit on a device with $K$ couplers. Each coupler slot applies the fermionic-simulation gate $\mathrm{fSim}(\theta,\phi)$, a two-qubit gate with a swap angle $\theta$ and a conditional phase $\phi$. Demonstrating a Continuous Set of Two-qubit Gates for Near-term Quantum Algorithms showed that flux-tunable couplers implement this gate natively over the whole $(\theta,\phi)$ range, including the exact zero-coupling point. Because $\mathrm{fSim}(0,0)$ is the identity, a sub-circuit is simply a coupler mask $m\in\{0,1\}^K$. We define the scaled slot
$$U_k(s_k)=\mathrm{fSim}(s_k\theta_k,\ s_k\phi_k),\qquad s_k\in[0,1],$$
where $\theta_k,\phi_k$ are shared trainable angles and $s_k$ is a per-coupler scale. At deployment, $s_k=m_k$. We measure interference between two masks $m,m'$ as the cosine between their loss gradients with respect to all shared parameters $w$:
$$c(m,m')=\cos\big(\nabla_w L(w;m),\ \nabla_w L(w;m')\big).$$
This follows the classical definition in Analyzing and Mitigating Interference in Neural Architecture Search. That paper found conflict grows with the number of operators by which two sub-networks differ.
The new ingredient is coupling-dilation training. The training mask moves as a slow random walk, flipping one coupler at a time. Each flip is not instantaneous: the scale $s_k$ ramps continuously between 0 and 1 over many steps, so the shared parameters follow a smooth path from one sub-circuit to its neighbor. We predict three things. First, $c(m,m')$ decays with mask Hamming distance in quantum circuits, as it does in classical supernets. Second, continuous ramping flattens that decay compared with hard flips at the same walk speed. Third, flatter decay translates into higher rank correlation between supernet energies and standalone energies.
Relation to Prior Work
Quantum supernets exist. Quantum circuit architecture search for variational quantum algorithms shares one parameter pool across sub-circuits built from discrete rotations and CNOTs, samples sub-circuits uniformly, and retrains the winner. Differentiable quantum architecture search relaxes gate choice into learned probabilities over a discrete gate pool. Neither measures gradient conflict or rank correlation against standalone training. Neither treats coupling as a continuous physical quantity. Both aim to find one good circuit. We instead aim to supply parameters for many circuits, each of which may be forced on us by the hardware.
Three works supply our methods. The MAGIC-T one-operator random walk from Analyzing and Mitigating Interference in Neural Architecture Search provides our walk. The epoch-level switching of Shipard et al. and the delayed small-subnet introduction of Annavajjala et al. are the temporal-smoothing competitors. Our addition is a smooth path through parameter space between neighboring masks, which exists only because the gate family is continuous and contains the identity.
Two further works are heavier alternatives we do not adopt as the core. Mixture-of-Supernets routes expert weights by architecture. FlexRank forces smaller models to be nested inside larger ones. The Mixture-of-Supernets router appears as an optional add-on.
Two tunable-coupler papers motivate the deployment setting. Strong Quantum Computational Advantage Using a Superconducting Quantum Processor shows that 110 couplers on a 66-qubit grid can each be set independently. Foxen et al. report parameter bands with elevated error caused by microscopic material defects called two-level systems, which is exactly what makes per-coupler shutdown a practical need. Neural Architecture Search for Quantum Autoencoders found that excess entanglement hurts trainability, which suggests many masked sub-circuits may be competitive rather than merely tolerable.
Research Program
The first phase is a measurement study. We use noisy statevector simulation with gate errors set to the reported fSim Pauli error of $3.8\times10^{-3}$ from Foxen et al. We add extra dephasing on active couplers, mirroring the $T_2^*$ drop from about 10 µs to about 4 µs reported under flux modulation in Parametrically Activated Entangling Gates Using Transmon Qubits. On this model we train naive uniform-sampling supernets and map $c(m,m')$ against Hamming distance. This produces the first curve of quantum weight-sharing interference, and it tests whether the classical correlation with topological distance carries over.
The second phase compares training schedules under equal budgets: hard flips versus continuous ramps, and several walk speeds. The third phase evaluates deployment. A held-out set of masks is drawn from realistic failure patterns, meaning random subsets of couplers declared unusable, and each is run with the supernet's parameters directly, with no retraining. If ranking still lags, we add a small classical network that maps the mask to angle corrections, in the style of Mixture-of-Supernets, and report its extra cost separately.
Baselines
The do-nothing reference trains one dense circuit with all couplers on. At deployment it simply zeroes the failed couplers and keeps the remaining parameters. This is what a practitioner would do by default. Any method that cannot beat it on held-out masks has no reason to exist.
The upper reference is standalone training of each evaluation mask from scratch, with the same optimizer and step count per circuit. It defines the "gap to standalone" and supplies the ground truth for rank correlation. The naive supernet is uniform mask sampling in the style of Du et al., run with the same total number of circuit evaluations as our method.
The strongest existing competitor is MAGIC-T's one-flip random walk with hard switching, combined with epoch-level switching as in Shipard et al. Both are tuned over walk speed and switching interval with the same search budget we give our ramp schedule. This comparison is also the key ablation. It differs from our method only in whether the flip is a hard switch or a continuous ramp, so it isolates the contribution of continuous coupling.
A further baseline trains a single dense circuit with an L1 penalty that drives coupler angles toward zero. This tests whether sparsity alone achieves robustness.
For reported numbers, the only interference-mitigation figure in the file is classical: MAGIC-AT raised Kendall's tau from 0.36 to 0.57 on a BERT search space. No reported result was found on Lacuna for Kendall's tau of any quantum supernet. We therefore establish the quantum baselines ourselves. On task energies, Du et al. report the hardware-efficient ansatz at −0.61 Ha and their searched circuit at −0.96 Ha on IBM hardware, against an exact −1.136 Ha for H₂. These are hardware numbers and serve only as context for our simulations.
Beating the strongest baseline means two things. The first is a Kendall tau on held-out masks higher than hard-flip MAGIC-T with non-overlapping seed confidence intervals. The second is a smaller median gap to standalone energy, both at equal circuit-evaluation budget.
Why not just…?
Why not just retrain for each coupler failure? This gives the best circuits, and we include it as the upper reference. The cost is a full variational optimization every time the set of usable couplers changes, which on real hardware means many shots per step. The proposal is worthwhile only if the zero-retraining gap is small. If a short fine-tune of 5–10% of standalone steps, starting from the dense circuit, closes the gap equally well, that is a real threat. We run exactly this fine-tune as an additional arm and will report it honestly.
Why not just use a hardware-efficient ansatz with fixed CZ gates? Fixed gates remove continuous coupling but not the failed-coupler problem, and Du et al. found this ansatz lags a searched circuit on H₂. We include it with fixed-angle gates in the same masked-deployment protocol.
Why not just train one sparse circuit with an L1 penalty on coupler angles? It yields one good sparse circuit. It gives no guarantee for a different circuit forced by a failed coupler it chose to keep. Our held-out-mask evaluation tests exactly this. If L1 turns out to be robust to arbitrary masks, the supernet machinery is unnecessary and we will say so.
Why not just use MAGIC-T or epoch-level switching as they are? They are the direct competitors, and the ramp-versus-hard ablation decides the question. If ramping gives no gain in either gradient cosine or tau, the continuous-coupling ingredient fails.
Proposed Experiments
We use three benchmarks from the file. The first is H₂ ground-state energy (exact −1.136 Ha). The second is the 6-qubit transverse-field Ising chain (exact −7.7274), where the predictor-based search of Neural predictor based quantum architecture search matched the exact energy. The third is Fashion-MNIST T-shirt-versus-dress classification, where that same work reports 0.924 validation accuracy versus 0.888 for the hardware-efficient ansatz. For each task we build 6–10 qubit brickwork circuits on a grid layout modeled on the Zuchongzhi grid.
We evaluate on 200 held-out masks at several failure rates. Metrics are the gradient-cosine-versus-Hamming-distance curve, Kendall tau against standalone training, median energy or accuracy gap to standalone, and total circuit evaluations.
The decisive check has two parts. First, the ramped schedule must show a flatter cosine-distance curve and a higher held-out tau than hard-flip MAGIC-T at matched budget. Second, it must beat both do-nothing zeroing and the short fine-tune arm on median gap. If only the cosine curve improves but tau does not, then gradient conflict is not what limits quantum supernets, which is itself a publishable negative result.
References
- Demonstrating a Continuous Set of Two-qubit Gates for Near-term Quantum Algorithms
- Analyzing and Mitigating Interference in Neural Architecture Search
- Quantum circuit architecture search for variational quantum algorithms
- Differentiable quantum architecture search
- Does Interference Exist When Training a Once-For-All Network?
- DεpS: Delayed ε-Shrinking for Faster Once-for-All Training
- Mixture-of-Supernets: Improving Weight-Sharing Supernet Training with Architecture-Routed Mixture-of-Experts
- FlexRank: Nested Low-Rank Knowledge Decomposition for Adaptive Model Deployment
- Strong Quantum Computational Advantage Using a Superconducting Quantum Processor
- Neural Architecture Search for Quantum Autoencoders
- Parametrically Activated Entangling Gates Using Transmon Qubits
- Neural predictor based quantum architecture search