← All publication reviewsWeekly Publications

October 24 to 30, 2026 · The weekly view

Logical qubits
get to work

Error-corrected qubits did logic on commercial IBM and Quantinuum machines, decoders for the qubit-saving qLDPC codes got faster in simulation, and one control line carried four gate signals.

116 (64%) algorithms & software52 (29%) hardware & devices63 (35%) error correction & verificationSelected research ↓

This review covers 180 distinct papers that appeared in arXiv’s dated announcement listings during the reporting window. It is a selected research corpus, not a count of all quantum papers published worldwide. The themes overlap: 116 papers concern algorithms or software, 63 error correction or verification, and 52 hardware or devices.

Three results frame the week.

Logical qubits started doing logic on commercial machines, at small size and after discarding many runs.University of Sydney and IBM · UCL and NQCC on Quantinuum H2-2 · Experiments
Decoders for qLDPC codes got faster and more accurate in simulation, and the classical bill is now where much of the work goes.Chinese Academy of Sciences and EIT Ningbo · UCLA · Simulation
The wiring and calibration around the qubits have to scale too, and one control line carried four gate signals without measurable loss.USTC and QuantumCTek · Amazon Braket · Experiments

Where the research connects

Counts from 180 reviewed papers

The overlap between this week’s themes

No publication falls outside these three themes. Areas, including the overlaps, are proportional to publication counts. Each region shows only its combination of labels.

The 52 hardware papers, by platform

Superconducting12
Cross-platform12
Photonics & networking9
Neutral atoms6
Semiconductor & spin6
Trapped ions3
Enabling devices2
Other hardware2

Every paper in the hardware & devices theme, counted once under its primary platform, so the bars add up to 52. Photonics and networking are counted together.

Bars use a common scale from 0 to 16 papers.

The three themes are editorial labels assigned from each paper’s title and abstract. Hardware includes device proposals and simulations as well as experiments. Papers may carry more than one label, so percentages do not add to 100%. The exact overlap is 27 hardware-only papers, 80 algorithms-only, 27 verification/error-correction-only, 10 algorithms-and-hardware-only, 21 algorithms-and-error-correction-only, 10 hardware-and-error-correction-only, and 5 carrying all three labels. These shares describe this week’s corpus, not a shift in worldwide research activity. Papers are assigned to the week by first arXiv announcement: the October 24 00:00 to October 31 00:00 Asia/Jerusalem window, which holds the arXiv listings dated October 26, 27, 28 and 30.

The research footprint

Countries and regions represented in the review

A global view of participation

Hover, tap, or choose a country or region.

Country or region

United States43reviewed publications

Reviewed publications

012.52537.550

Not represented in this review

Affiliation locations; a cross-border collaboration can contribute to several countries. Country or region information is available for 138 of 180 reviewed publications. Map: Natural Earth.

Author affiliations were read from each paper’s arXiv HTML author header. Country sets are fully resolved for 132 papers; six retain unlocated affiliations, and 42 carry no affiliation in that header and are not mapped. Each paper counts once per verified country, producing 182 country-paper attributions across 38 countries. Gray means no verified affiliation in this cohort, not an absence of research. Hardware vendors and acknowledgments are not author affiliations.

Three themes worth reading closely

Selected research, with company context

IBM

NYSE: IBM

Quantinuum

NASDAQ: QNT

The University of Sydney with IBM Quantum, on IBM's ibm_phoenix; University College London and the UK's National Quantum Computing Centre, on Quantinuum's H2-2 (no Quantinuum authors)

1. Logical qubits start talking to each other

A University of Sydney group, with IBM’s Benjamin Brown as a co-author, ran distance-3 and distance-5 surface codes on ibm_phoenix, a 120-qubit IBM processor it describes as a “commercially accessible device.” Skipping weak qubits cut the distance-5 logical error per round from 1.9% to 0.9%. The team then made logical qubits interact. Lattice surgery, the main proposed way to run gates between surface codes, produced a logical Bell state between distance-3 codes with fidelity above 93% and teleportation across three logical qubits reached 78% to 93%. The 93% counts only runs where the codes were prepared correctly, about 40% of them.

On Quantinuum’s H2-2 trapped-ion machine, researchers from UCL and the UK’s National Quantum Computing Centre, with no Quantinuum authors, did the same kind of move for a denser code (arXiv 2610.39067). A [[10,2,3]] code stores two logical qubits in one block, and seven helper qubits measured one of them with 0.72% error in a single round. With three rounds, 957 verified runs showed no errors, which bounds the error below 0.26%. Each setting ran a few hundred to about a thousand times, after discarding runs that failed verification.

Both results move from storing encoded qubits to operating on them, at small size and on commercial hardware. Both also lean on throwing away bad runs. A real computer can’t discard most of its gates. How much of each run survives?

Entangled logical qubits on a superconducting quantum processor ↗

arXiv announcement October 26 · Experiment

Academic research

No company author affiliation · NVIDIA GPU (NASDAQ: NVDA)

Institute of Physics, Chinese Academy of Sciences, Beijing, and EIT Ningbo; University of California, Los Angeles

2. The cheaper codes need a faster referee

Codes like the [[10,2,3]] above belong to the qLDPC family, which packs many logical qubits into fewer physical ones. The price is decoding, the classical computation that reads error signals and decides what to fix. Researchers at the Institute of Physics, Chinese Academy of Sciences, and EIT Ningbo built Neural Triage, a decoder where a fast neural network handles most cases and heavier computation goes only to the rare hard ones. On one large code, in simulation at a 1.5% physical error rate, it cut the average time per shot to 2.37 milliseconds on one NVIDIA GPU, against 815 milliseconds for the baseline on one CPU thread, with lower logical error. The setup is simplified noise, GPU batches against a single CPU thread, and no real-time latency yet, which the authors list as next steps.

Jacob Murri at UCLA showed in simulation that the decoder also has to know how the hardware fails (arXiv 2610.38283). In neutral atoms, losing one atom in a gate often means losing its partner. A decoder that uses that pairing lowered logical error “by as much as ∼25×” against one that ignores loss, on IBM’s [[144,12,12]] gross code.

The qubit savings of qLDPC codes are real on paper, and they get paid for in classical compute. Which decoder, on what hardware, and how fast in real time?

Fast and accurate neural triage decoding of ultra-high-rate quantum LDPC codes ↗

arXiv announcement October 30 · Simulation

QuantumCTek

SHA: 688027

Amazon

NASDAQ: AMZN

University of Science and Technology of China and QuantumCTek, Hefei; Amazon Braket and AWS Quantum Technologies, with runs on Rigetti (NASDAQ: RGTI) and IQM (NASDAQ: IQMX) processors

3. The plumbing has to scale too

USTC and QuantumCTek, with Jian-Wei Pan’s group, attacked the wiring. The paper notes that leading superconducting systems use about three control lines per qubit, which means nearly a million lines at a few hundred thousand qubits, and more heat than a dilution refrigerator can remove. Their fix sends several gate signals down one line at different frequencies. In a 1-to-4 demonstration, parallel two-qubit gates had 0.64% error and single-qubit gates 0.106%, “without statistically significant multiplexing-induced degradation.” The “one to two orders of magnitude” fewer lines is a projection.

Drift is the other cost. Amazon Braket proposed a cheap screening test that triggers full recalibration only when something has changed (arXiv 2610.34003). It detected changes more than 4× as often as randomized benchmarking at equal cost, in theory and in runs on Rigetti’s Ankaa-3 and IQM’s Garnet and Emerald.

Every minute spent checking a machine is a minute not sold to a customer. Does the wiring count per qubit fall as the machine grows?

A Control Line Multiplexing Framework for Scalable Quantum Computing ↗

arXiv announcement October 27 · Experiment

Also this week

Princeton · Neutral atoms
A cesium atom array inside cryogenic shields reached 99.970% single-qubit and 99.76% two-qubit gate fidelity, a universal gate set at low temperature, in experiment.
arXiv 2610.40329 ↗
Princeton · Superconducting
Sudden drops in transmon lifetimes, lasting minutes to hours, traced in experiment to one or a few defects in the chip materials.
arXiv 2610.34485 ↗
USC and MIT on IBM · Error correction
An experiment on IBM superconducting hardware observed the noise level at which scrambled information stops being recoverable.
arXiv 2610.38148 ↗
Alice & Bob · Resource estimate
A resource estimate puts FeMoco-54 on 7,164 cat qubits in 23.5 hours, assuming negligible bit flips, and about 13,000 if the bit-flip rate is relaxed to 10<sup>-11</sup>.
arXiv 2610.37840 ↗
Alice & Bob · Resource estimate
A companion estimate compiles RSA-2048 on 22,593 cat qubits in 3.14 days with no bit flips, or about 45,000 with an added bit-flip code.
arXiv 2610.37839 ↗
IonQ · Algorithms
New block encodings cut the leading T cost 15× to 17× for a 120 by 120 frustrated magnet model, in theory.
arXiv 2610.32946 ↗
Google and Macquarie · Chemistry
A proof that first-quantized chemistry simulations reach the basis-set limit at the expected 1/N rate in most relevant cases.
arXiv 2610.35205 ↗
Yokohama, Tokyo and RIKEN · Distributed
An estimate that one 1 MHz, 90%-fidelity microwave link per node substantially cuts qubits per node, at about ten times the run time.
arXiv 2610.34143 ↗
Innsbruck and Parity on QuEra · Neutral atoms
Atoms formed a classical error-correcting code that recovered exact optimization answers in at least 8% of cases on average, with up to 107 atoms; bit-flip protection only.
arXiv 2610.39670 ↗
Strathclyde and Newcastle on D-Wave · Annealing
In experiment, fast anneals on D-Wave hardware showed the expected coherent domain-wall pattern, proposed as a benchmark.
arXiv 2610.40504 ↗
Oak Ridge on IBM · Software
QQLink, a framework for workflows across several quantum processors, ran in experiment on two IBM Heron chips and found losses beyond calibration models.
arXiv 2610.34486 ↗
Codes and gates
Logical operations identified from syndrome histories with over 94% accuracy on bivariate bicycle codes, in simulation.
arXiv 2610.40052 ↗

What I would watch next

The week’s common thread was cost. I want to see every reported logical gate come with its yield, its decoder time and its control wiring. Watch for lattice surgery on IBM hardware with fewer discarded runs, for qLDPC decoders timed in real time on the machine, and for multiplexed control beyond four signals per line. For background, see What is a logical qubit? and What is a decoder?

Where the research connects

The overlap between this week’s themes

The 52 hardware papers, by platform

The research footprint

A global view of participation

Three themes worth reading closely

1. Logical qubits start talking to each other

2. The cheaper codes need a faster referee

3. The plumbing has to scale too

Also this week

A cesium atom array inside cryogenic shields reached 99.970% single-qubit and 99.76% two-qubit gate fidelity, a universal gate set at low temperature, in experiment.

Sudden drops in transmon lifetimes, lasting minutes to hours, traced in experiment to one or a few defects in the chip materials.

An experiment on IBM superconducting hardware observed the noise level at which scrambled information stops being recoverable.

A resource estimate puts FeMoco-54 on 7,164 cat qubits in 23.5 hours, assuming negligible bit flips, and about 13,000 if the bit-flip rate is relaxed to 10<sup>-11</sup>.

A companion estimate compiles RSA-2048 on 22,593 cat qubits in 3.14 days with no bit flips, or about 45,000 with an added bit-flip code.

New block encodings cut the leading T cost 15× to 17× for a 120 by 120 frustrated magnet model, in theory.

A proof that first-quantized chemistry simulations reach the basis-set limit at the expected 1/N rate in most relevant cases.

An estimate that one 1 MHz, 90%-fidelity microwave link per node substantially cuts qubits per node, at about ten times the run time.

Atoms formed a classical error-correcting code that recovered exact optimization answers in at least 8% of cases on average, with up to 107 atoms; bit-flip protection only.

In experiment, fast anneals on D-Wave hardware showed the expected coherent domain-wall pattern, proposed as a benchmark.

QQLink, a framework for workflows across several quantum processors, ran in experiment on two IBM Heron chips and found losses beyond calibration models.

Logical operations identified from syndrome histories with over 94% accuracy on bivariate bicycle codes, in simulation.

What I would watch next