One ion, no measurements: USTC reports break-even for autonomous error correction, but the baseline is contested

Status: Disputed (peer-reviewed, but another group contests the break-even baseline) · AI-assisted reporting, passed the newsroom’s automated quality gates (not reviewed by a human editor)

Key Findings

A team led by researchers at the University of Science and Technology of China (USTC) reports that a qubit stored inside a single calcium ion can have its phase errors corrected without anyone measuring it. The protected qubit lasted longer than the best unprotected qubit the team could store in the same energy levels of the same ion, which is how the authors define the “break-even” mark. The paper was peer-reviewed and published in Nature Physics on 23 October 2026 [S1]. Its preprint has been public since April 2025 [S2].

What is established: the paper passed peer review [S1]. The authors have published the data and fitting code behind their figures [S3]. A separate group at MIT and MIT Lincoln Laboratory published a closely related single-ion experiment in the same journal in July 2026 [S4].

What is contested: whether this comparison really counts as break-even. A rival group in Shenzhen says that no autonomous error-correction experiment so far has “unambiguously reached the break-even point”, and that a different qubit in calcium is less sensitive to magnetic-field noise [S6][S7]. That group is a competitor with its own claim, and its paper is a preprint, so read the point as a serious objection and not a verdict.

What is the authors’ claim, not yet checked by others: the specific lifetimes and the size of the gain. We found no independent replication of this result.

How It Works

The qubit stores information in the phase between two quantum states, which works a bit like the timing between two clocks. Tiny wobbles in the magnetic field around the ion shift that phase. This is called dephasing, and in this setup it is the main way information is lost.

Most error-correction schemes spread one logical qubit across many physical qubits. They then measure helper qubits to find the error and apply a fix. This team took a different route [S1][S2]:

1. More levels inside one atom. A calcium-40 ion has a cluster of closely spaced energy levels called the D5/2 sublevels. The team stored one logical qubit across four of them, using a specific weighted mix of states. The mix is chosen so that the most common dephasing error pushes the state into a separate, recognisable “error space” instead of quietly scrambling the data [S2].
2. Undo the error by adding a vibration. A tuned set of laser tones moves any state in the error space back to the right data state. That fix has a cost: it adds one quantum of vibration (a phonon) to the motion the ions share [S2].
3. Drain the vibration with a second ion. A second calcium ion sits next to the first and is laser-cooled. Cooling it removes the vibration from their shared motion, so the error leaves the system as heat. Nothing is measured and no decision is made based on a result. The authors call this autonomous error correction [S1][S2].

One full cycle took 3.2 milliseconds in ordinary lab conditions. About 0.62 ms of that is the correction step and the rest is cooling and idle time [S2]. The code fixes only first-order dephasing errors. A larger double error turns into a logical error that the code cannot fix [S2].

“Break-even” is measured with Λ, the protected qubit’s lifetime divided by the lifetime of the best unprotected qubit. Here that comparison qubit is a pair of D5/2 levels in the same ion [S2]. The independent Shenzhen group notes that a qubit stored between the ground state and the long-lived D5/2 state is less sensitive to magnetic-field noise [S6]. A 2013 Innsbruck review of trapped-ion processors says the same of that transition, though such an optical qubit has its own limits from laser phase noise [S7].

• With added noise: the team deliberately injected random low-frequency magnetic noise. The protected qubit’s fitted lifetime was about 11.7 ms, compared with about 1 ms for the best physical qubit and about 0.8 ms for the encoded qubit with correction switched off [S3]. The preprint reported 11.6 ± 1.9 ms versus 0.9 ± 0.1 ms, a factor the authors give as Λ = 12.9 ± 2.6 [S2].
• With ordinary lab noise: the gain is smaller. The preprint reported 60.9 ± 33.0 ms against 26.0 ± 5.4 ms (Λ = 2.3 ± 1.4) and attributed the “large uncertainty” to fluctuating lab noise [S2]. The figure data released with the published paper give about 42.7 ± 9.0 ms against 29.4 ± 2.6 ms [S3]. That is roughly a 1.5× gain by our own calculation, not a number the authors state. At one fixed time point, the protected qubit scored higher than the physical qubit in 11 of 15 back-to-back trials [S3].

Implications

Measurement-free correction matters because measuring qubits during a computation is slow and itself causes errors. Using extra levels inside one atom could also mean fewer physical qubits per logical qubit. The authors frame both points as a route to “hardware-efficient” quantum memories [S1].

The approach is getting independent support. DeBry and colleagues at MIT, MIT Lincoln Laboratory and Caltech used a different single-ion code with their own autonomous scheme. They report up to 2.2× fewer errors and up to 1.5× longer lifetime than an unencoded qubit [S4]. That supports the general idea. It is not a replication of the USTC numbers.

Keep the scale in mind:

• This is one logical qubit acting as a memory for milliseconds. No logical gates were performed and no computation was run [S1][S2].
• Break-even is measured against the best qubit stored in the same D5/2 levels, not the best trapped-ion qubit anywhere. The paper itself cites a single-ion qubit whose coherence was estimated to exceed an hour [S1]. The Shenzhen group makes the same objection about both this experiment and the MIT one [S6].
• The large 12× gain appears only under noise the team added on purpose, of the kind this code is built to fix. For that run they also shortened the idle time between cycles to 0.12 ms [S2]. A news write-up of the work made the same distinction [S5].

Open Questions

• Replication. No other group has reproduced this specific protocol, as far as our search found. We also found no Nature Physics News & Views piece, no press release and no mainstream science coverage; the one independent engagement is the Shenzhen preprint [S6].
• The break-even baseline. Whether the D5/2-level comparison is the right yardstick is open. The USTC authors state their definition [S2]; the Shenzhen preprint disputes its adequacy [S6]. We found no USTC reply.
• Ordinary-noise margin. The lab-noise result is modest. It also changed between the preprint (Λ ≈ 2.3) and the published figure data (≈1.45 by our calculation) [S2][S3]. We could not read the published main text, which is paywalled, so we cannot say how the authors explain the change.
• Data selection. For the ordinary-noise runs, the authors kept only data where the magnetic-field calibration drifted by less than 0.9 nT. They estimate this has a negligible effect [S2]. Readers should know it happened.
• Other error types. The code targets dephasing only. Other errors, such as spin flips and decay of the D5/2 levels, are not corrected by this scheme [S2].
• Scaling. The authors list next steps they have not yet shown. These are using more sublevels to guard against more error types, entangling logical qubits held in different ions, and building concatenated codes on top [S2].

References

1. [S1] Li, Y., Mei, Q., Jie, Q.-X. et al. “Beating the break-even point with autonomous quantum error correction.” Nature Physics (published 23 October 2026). https://www.nature.com/articles/s41567-026-03486-y. Abstract, Extended Data captions and article information read; full text paywalled.
2. [S2] Same authors. arXiv:2504.16746v1 (23 April 2025). https://arxiv.org/abs/2504.16746. Results, Fig. 4 and Methods.
3. [S3] Same authors. Source data and code for the published figures, Zenodo (20 June 2026), CC-BY-4.0. https://doi.org/10.5281/zenodo.20768851. Fig. 4a, Fig. 4a inset and Fig. 4b CSV fit blocks.
4. [S4] DeBry, K. et al. “Error correction of a logical qubit encoded in a single atomic ion.” Nature Physics 22, 1446–1452 (published 13 July 2026). https://www.nature.com/articles/s41567-026-03315-2.
5. [S5] TSN. “USTC Qubit Beats Break-Even in One Trapped Ion.” (24 October 2026). https://tsnmedia.org/ustc-trapped-ion-break-even/. Secondary coverage based on the preprint.
6. [S6] Ni et al. (International Quantum Academy, Shenzhen). “Autonomous quantum error correction beyond break-even and its metrological application.” arXiv:2509.26042 (30 September 2025). https://arxiv.org/abs/2509.26042. Independent preprint from a competing group; not peer reviewed.
7. [S7] Schindler, P. et al. “A quantum information processor with trapped ions.” arXiv:1308.3096 (journal version New J. Phys. 15, 123012, 2013). https://arxiv.org/abs/1308.3096. Background on qubit choices in calcium-40.