The Signal
PRISM preprint reports up to 30.8% fewer routing SWAPs
A preprint introduces PRISM, a compiler that delays choosing among equivalent dual-rail gate implementations until routing. Its authors report fewer routing SWAPs and lower physical depth than a greedy baseline across simulated workloads.

Keeping gate choices open
A compiler for dual-rail cavity-qubit circuits reduced routing SWAPs by 20.4–30.8% in evaluations, according to a preprint posted on arXiv by its authors. The work introduces PRISM, which delays choosing among equivalent physical implementations of a logical gate until the router finds a layout where one can run.
The authors are Evan McKinney, Kun Liu, Harshvardhan K. Babla, Robert J. Schoelkopf and Yongshan Ding. Their paper, posted as an arXiv preprint on November 17, 2026, describes PRISM as a compiler that carries alternatives across the boundary between logical gates and physical instructions.
Routing against a greedy baseline
For a logical controlled rotation in the dual-rail cavity model, the paper defines four loss-admissible physical implementations, one for each pairing of the two rails belonging to each qubit. Which pairing can run depends on the current placement of the rails on the device graph. PRISM retains these choices during routing and commits to one when it is executable under the current layout.
In tests spanning 25 workloads with 7 to 25 logical qubits on three device graphs, the authors report 20.4–30.8% fewer routing SWAPs and 8.8–10.1% lower physical depth than Greedy. That baseline uses the same boundary representation and physical router, but fixes each gate’s implementation at the start of a routing pass. The paper also compares PRISM with Logical-SABRE, which routes on a fixed logical coupling graph before lowering gates to physical instructions.
The authors say the SWAP reductions held through 32 logical qubits on grid and hexagonal-lattice graphs. The workload set included algorithm benchmarks and circuits drawn from QASMBench; the paper describes the three device graphs as grid, hexagonal lattice and heavy-hex.
Modeled hardware results and limits
A separate analysis using modeled dual-rail cavity hardware estimated 44.3% fewer attempts per retained shot at 16 logical qubits, as a geometric mean across workloads and device graphs. This is a model-based estimate, not a report of a hardware demonstration. The paper also evaluates three-round Steane-memory circuits under skip-gate leakage: it reports 13.0–19.6% lower decoded logical error with flags at readout and 22.6–27.7% lower error with compiled erasure checks.
The compiler’s evaluated model has boundaries. The paper says teleportation, inter-module communication, measurement and feedforward, ancilla-mediated embeddings and microwave links between modules are outside its model. It also says PRISM’s realization set is limited to circuits that keep the same wire mapping and logical frame and can be embedded without SWAPs inside the realization.
The authors say PRISM is open source under the Apache-2.0 license and installs as the Qiskit package prism-qc. They state that the repository includes code, scripts and data to reproduce the paper’s figures and numbers.