A study of a four-qubit variational quantum classifier evaluated on two IBM Heron r2 backends reveals that post-measurement readout strategies and independently trained model variations heavily influence real hardware performance. Researchers analyzed the quantum classifier across a three-class traffic-state classification task to understand deployment behavior on physical processors.
In a direct single-seed comparison, readout design produced divergent performance profiles between idealized software simulations and real quantum hardware. A four-channel linear-projection readout achieved a higher clean balanced accuracy of 94.4 percent compared to 78.9 percent for a three-channel direct-argmax readout, yet their respective hardware recoveries on the ibm_kingston processor were measured at 53.0 percent and 94.4 percent.
To contextualize these accuracy shifts, the researchers implemented a matched synthetic-noise control experiment. This control confirmed that prediction stability under simulated disturbance does not necessarily translate into higher absolute accuracy when running on noisy quantum hardware.
Across five independently trained direct-argmax models, hardware sensitivity varied significantly across classification categories. The clearest divergence between backends emerged in the moderate traffic class, where the ibm_marrakesh backend outperformed ibm_kingston. Conversely, differences in the congested class were smaller and dependent on measurement shot counts, while recall for the free-flow class remained stable across both systems.
While disagreement between shots decreased steadily as shot counts increased, the underlying hardware discrepancies between the two processors did not follow this monotonic trend.
Register-level assignment-matrix indicators displayed directional alignment with observed hardware variations, although the authors noted that this correlation does not establish a direct causal mechanism. Furthermore, applying temperature scaling reduced expected calibration error, leaving an empirical rather than structural residual error near 0.17.
Together, the findings highlight the critical importance of evaluating readout mechanisms, class-resolved metrics, decision margins, and multi-seed replication when benchmarking variational quantum algorithms on physical hardware.
Subject of Research: Hardware evaluation and readout sensitivity of variational quantum classifiers
Article Title: Post-measurement readout and class-resolved hardware sensitivity of a variational quantum classifier on two IBM Heron r2 backends
Article References: Yi, H. (2026). Post-measurement readout and class-resolved hardware sensitivity of a variational quantum classifier on two IBM Heron r2 backends. EPJ Quantum Technology. https://doi.org/10.1140/epjqt/s40507-026-00570-3
Image Credits: AI Generated
DOI: 10.1140/epjqt/s40507-026-00570-3
Keywords: variational quantum classifiers, quantum machine learning, IBM Heron r2, ibm_kingston, ibm_marrakesh, post-measurement readout, traffic-state classification, hardware sensitivity, direct-argmax readout, linear-projection readout, quantum noise, finite-shot disagreement, decision margins, temperature scaling, expected calibration error, quantum hardware benchmarking, quantum computing
Cite Scienmag News
Katie Riggs. (November 20, 2026). Readout Methods and Noise Shape Variational Quantum Classifier Performance on IBM Chips. Scienmag. https://scienmag.com/readout-methods-and-noise-shape-variational-quantum-classifier-performance-on-ibm-chips/
Katie Riggs. "Readout Methods and Noise Shape Variational Quantum Classifier Performance on IBM Chips." Scienmag, 20 November 2026, https://scienmag.com/readout-methods-and-noise-shape-variational-quantum-classifier-performance-on-ibm-chips/. Accessed 20 November 2026.
Katie Riggs. "Readout Methods and Noise Shape Variational Quantum Classifier Performance on IBM Chips." Scienmag. November 20, 2026. https://scienmag.com/readout-methods-and-noise-shape-variational-quantum-classifier-performance-on-ibm-chips/








