Four-test scalable suite scoring pre-fault-tolerant devices on partial-Clifford random circuits, GHZ-state entanglement, Ising-dynamics simulation reach, and quantum neural network classification accuracy.
QuSquare is a proposed benchmark suite for pre-fault-tolerant quantum devices from Aguirre, Peña & Sanz of the NQUIRE group at the University of the Basque Country. It turns the group’s earlier review of what a quantum benchmark should satisfy into four concrete, scalable tests and deliberately declines to compress a device into a single aggregate number. The name is written “QuSquare” throughout, with no stated expansion.
How it works
The specification defines four tests, plus rules on compilation, error-mitigation disclosure, and reporting:
- Partial Clifford Randomized: whole-device random circuits; scores the largest circuit fraction μ that keeps infidelity at or below 0.1.
- Multipartite Entanglement: scores the largest GHZ state prepared with fidelity above 1/2, the standard witness of genuine multipartite entanglement.
- TFIM Hamiltonian Simulation: quantum-signal-processing simulation of transverse-field Ising dynamics; scores the longest evolution time reached within a bounded magnetization error.
- Data Re-uploading QNN: binary classification with data re-uploading quantum neural networks on four mandatory synthetic datasets (circles, spiral, sinus, corners); scores average test accuracy Acc(N, L).
The first two are whole-device tests in the tradition of Quantum Volume and GHZ-state fidelity; the latter two are application-style workloads closer to suites like SupermarQ, giving the suite a foot on both levels.
Strengths and limitations
The paper is unusually explicit about methodology: each test carries a defined figure of merit, the metrics are analyzed against stated quality attributes, and scalability is a design goal throughout. Against that, QuSquare is so far only a specification. The paper reports no measurements on real hardware, the four constituent tests exist only within the suite rather than as independently established benchmarks, and the lack of an aggregate score rules out simple rankings by design. The reference implementation is public but lacked a license file as of its February 2026 update.
As of mid-2026 the suite remains a proposal: the December 2025 preprint is v1-only with no journal version, and while the authors maintain the code and published the mandatory QNN datasets on Zenodo, no hardware campaign or independent adoption has surfaced yet.