A new arXiv preprint reports that a properly chosen local measurement basis can classify predefined entanglement block structures in noisy simulations of systems up to 100 qubits, with mean accuracy above 95% in a task with 30 candidate partitions. On hardware, however, the score was 0.82 at 13 qubits and 0.20 at 16.
The problem it tackled
The study was a predefined sorting task. Its simulated data comprised block-partitioned GHZ, W and cluster states, with labels generated from random integer partitions and combined with Bell-pair and product-qubit components. The benchmark covered systems of 20, 30 and so on through 100 qubits, with 30 candidate partitions; random guessing would be about 3.3%.
The measurement shortcut
The researchers measured overlapping four-qubit windows, stepping three qubits at a time, and used the resulting local features as classifier input. In the full reference protocol, all 81 local Pauli bases were retained. It reached exactly 100% test accuracy from 20 through 100 qubits, while the measurement count rose from 567 to 2,673.
The protocol keeps the number of distinct measurement configurations independent of system size, while the classifier input grows only in proportion to the number of qubits. The reduced version retained one properly chosen basis: at 60 qubits it exceeded 97% accuracy and cut settings by a factor of 81. The representative choices were XZYX for GHZ, YXYX for W and ZZZX for cluster states.
Across the noisy simulations, the single-basis protocol had mean accuracy above 95% through 100 qubits.
Where hardware runs out of room
Hardware validation used the Shenglian superconducting processor, an 84-qubit device with 113 tunable couplers. It covered systems from 7 to 16 qubits, with five structural classes and 100 prediction trials at each size, including 20 trials per class. Each prediction trial used 10,000 shots—repeated measurements—to create one empirical local probability distribution for classification.
Accuracy was 0.99 at 7 and 8 qubits and 0.97 at 9. It then fell to 0.82 at 13, 0.73 at 14, 0.52 at 15 and 0.20 at 16. With five classes, 0.20 is the random-guess baseline.
Noise sets the boundary
A separate simulation stress test combined depolarizing and white noise. At the strongest settings, accuracy fell to approximately 0.88 for cluster states, 0.82 for GHZ states and 0.64 for W states.
Hardware calibration reported median errors of 1.0 × 10−3 for single-qubit gates and 7.0 × 10−3 for two-qubit gates. The authors identify accumulated two-qubit-gate errors and decoherence as important limitations as circuit size and depth increase. The hardware classifier was trained on simulation data without hardware retraining.
The 100-qubit finding is therefore a simulation result, not a 100-qubit hardware demonstration. The hardware evidence covered five structural classes from 7 to 16 qubits on the Shenglian processor, while the numerical work used constructed GHZ-, W- and cluster-block ensembles. The findings concern recognition within those predefined tasks; they do not establish arbitrary entanglement certification or full-state reconstruction.
The document is marked arXiv:2608.20170v1 and dated 20 Aug 2026.
Paper data and sources
Original title: Large Scale Entanglement Structure Detection in 100-Qubit Systems via Local Joint Measurements
Authors: Rui Li, Yuhang Wang, Chunxiao Du et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text