Skip to content
Quantum.world

The Signal

Batched optimizer reaches quantum training loss in fewer jobs

A preprint reports that a batched parameter-shift optimizer matched a stochastic optimizer’s final training loss using 4 to 20 of its 50 hardware jobs. The comparison used IBM’s ibm_fez processor and a small quantum neural network classifier.

Batched optimizer reaches quantum training loss in fewer jobs

Filling a job before it is sent

A preprint submitted on November 12, 2026, reports that a batched parameter-shift optimizer reached a stochastic optimizer’s final training loss using fewer hardware jobs. The paper by Muhammad Faryad and Rabab Alomairy studies a quantum neural network classifier, comparing optimizers given the same job count and shots per job.

Source arxiv.org

The work focuses on the repeated submissions needed when a classical optimizer trains a model on a queued quantum processor. Each job involves sending circuits to the processor and waiting for their results before deciding what to try next. The authors frame the job, rather than the individual measurement shot, as a unit of training cost.

Their approach puts a batch of parameter settings into a job before submission. For circuits built from Pauli rotations, the parameter-shift rule supplies a stencil of parameter variations that can estimate the gradient; the optimizer uses that gradient to update the model with Adam. The comparison method, simultaneous perturbation stochastic approximation, or SPSA, evaluates two perturbed parameter settings to estimate a direction for its update.

Simulation and hardware results

In simulations, the batched parameter-shift method needed between a fifth and a half of the jobs SPSA needed for models with 24 to 48 parameters, the authors report. The advantage held across two simulated noise levels. The simulations used 100 jobs of about 22,500 shots each and compared the optimizers’ training progress.

The hardware comparison used IBM’s ibm_fez processor in three interleaved runs: two with six qubits and one with eight. Each optimizer received 50 training jobs, with about 2.2 × 104 shots per job. In the three runs, the batched method reached SPSA’s final training loss after 4, 20 and 8 of SPSA’s 50 jobs, respectively, according to the paper.

For the eight-qubit run, the reported test accuracy was 0.79 for the batched optimizer and 0.71 for SPSA. The authors say the difference was not significant in the six-qubit runs. They also caution that the pooled test comparison should be read carefully because all runs used the same 300 test points and the result was carried by the eight-qubit run.

Scope of the comparison

The classifier was a controlled training task, not a test of whether a quantum model outperformed a classical one. Its six-qubit circuit had 36 trainable angles and 10 CZ gates; the eight-qubit circuit had 48 trainable angles and 14 CZ gates. The paper says that at six qubits, classical logistic regression and an RBF support-vector machine reached about 0.8 test accuracy.

The hardware tests used IBM’s job mode on the Open Plan, without sessions or batches. Each run took 102 jobs and 964 metered seconds, including training and additional read-out and drift-check jobs. The authors say the parameter-shift method’s final loss was reached at a metered cost per job within one second of SPSA’s; the reported job savings therefore came from needing fewer jobs, not from a shorter metered time for each one.

The paper is a preprint. It says the hardware job identifiers, raw shot records and verification notebook are available with the associated code and data.

Sources

  1. Training quantum neural networks one job at a time
  2. arXiv
  3. arXiv (full text)

Related stories

Explore

More articles