A recent paper posted to arXiv on October 23, *Entangled logical qubits on a superconducting quantum processor* (arXiv:2610.32347, published October 26), presents an intermediate but striking result: on a 120-qubit processor where components numbered 55, 62, and 95 were dragging down the whole chip, removing them from the error-correction layout and rewiring the measurement circuits cut the per-round logical error rate of a distance-5 surface-code patch from 1.9% to 0.9%.
The experiment and hardware
- Authors: Robin Harper, Xanda Kolesnikow, and Campbell McLauchlan (University of Sydney School of Physics), plus Georgia Nixon and Stephen Bartlett (Sydney) and Benjamin J. Brown (IBM Quantum / IBM Denmark). Funded by IARPA's "Entangled Logical Qubits" program, agreement W911NF-23-2-0223.
- The machine, ibm_phoenix, is a 120-qubit superconducting processor on IBM's Nighthawk r2 square lattice, released to the cloud in late August. According to Quantum Computing Report, these multi-logical-qubit benchmarks were delivered 54 days after it joined the fleet.
- A distance-5 rotated surface-code patch uses 49 physical qubits; distance-3 needs only 17.
- Ancilla reassignment around defective components. A bad ancilla (readout error or faulty coupler) contaminates every stabilizer round it participates in. The team skipped the affected four-body parity check in one round, then recovered its value in the next round using two two-body gauge measurements on neighboring ancillas. The worst component is never used at all.
- 50 controlled comparisons. Ten distance-3 patch positions fit inside a distance-5 patch; each has four individually removable ancillas, giving 40 variants plus 10 untouched originals. At the position labeled q43(L), the original patch erred 2.18% per round; the best removal cut this to 1.25%, while the worst removal raised it to 5.47%. The authors note every position has exactly one removal that improves performance.
- A numbering caveat. Appendix C's main text lists removed components as 55, 62, 95, while the Figure 8 caption says 62, 55, 96; the underperforming parts are "the couplers around 55 and 62" and "the coupler connecting 95 and 96."
- Why scaling didn't pay off yet. Despite nearly triple the qubit count of distance-3, the reassigned distance-5 patch only beat the *median* distance-3 patch and merely matched the best ones. The authors attribute this to device variability: larger patches cover more of the array and are likelier to absorb a defect, whose penalty grows with code distance. Simply stacking code distance did not yet deliver the theoretically promised gains on this machine.
- Two logical qubits entangled. Via lattice surgery (merging and splitting patches along their boundary), two logical qubits were entangled with a logical Bell state fidelity above 93%. The run used 500,000 shots per circuit on October 6, post-selecting only runs where both patches were confirmed in the code space (~40% ± 2%). The bound is given using logical XX and ZZ expectation values only — logical Y-basis readout error was too high to use.
- Three logical qubits. State and gate teleportation achieved fidelities of 78%–93% (85.4%–91.4% for one set, with post-selection keeping only 1.2% of runs).
- 170-round stability. 8 decoding failures in 50 million shots (~2×10⁻⁷ per shot) across 603.5 billion two-qubit gates and 170.8 billion measurements, using a moving syndrome-extraction circuit that swaps data-qubit and ancilla roles every round, so each physical qubit is measured and hardware-reset every two rounds — clearing leakage before it spreads into correlated errors.
- Sources: arXiv:2610.32347 abstract (submitted 2026-10-23, published 10-26); Scirate listing; Quantum Brief 2026-10-27; Quantum Computing Report 2026-10-28; Quantum Zeitgeist weekly 2026-11-02.
- Limitations: the 93% fidelity is a bound from XX and ZZ expectation values only; ~40% of runs were kept by post-selection; the 50-comparison set and 2×10⁻⁷ figure are self-reported with no third-party replication; the three engineering hurdles are media interpretation; the "54 days after cloud release" claim is from Quantum Computing Report and was not verified against IBM's original announcement.
Key points
What became possible at 0.9%
| Metric | Value | Note | |---|---|---| | Distance-5 per-round logical error | 1.9% → 0.9% | After removing 55, 62, 95 | | Best removal at q43(L) | 2.18% → 1.25% | Worst removal rose to 5.47% | | Two-logical-qubit Bell state | > 93% | Bounded by XX and ZZ expectations | | Three-logical-qubit teleportation | 78% – 93% | State and gate classes | | 170-round stability | 8 of 50M shots failed | ~2×10⁻⁷ |
Context: two routes to the same resource budget
Compared with an October 28 QLDPC-surgery paper on trapped ions (two logical qubits packed into one ten-data-qubit code block), this work keeps the surface code's one-patch-per-plot layout and instead digs the bad parts out of the ground — the former saves on encoding rate, the latter on yield. Neither marks a turning point yet: the defect-avoidance approach "cannot scale indefinitely," distance-5 only matches distance-3, and the logical Y basis is unreadable. Quantum Computing Report additionally lists three engineering hurdles — residual crosstalk on denser lattices, two-qubit-gate fidelity under concurrent mid-circuit readout, and sub-microsecond decoder response for real-time feedforward — though these are media interpretation, not numbers from the paper itself.
The real shift is in the question being asked: not how many physical qubits a machine has, but how many clean patches can be carved out of an already-flawed set of qubits. For a machine just two months into cloud availability, that second answer matters more.