A research team from IBM, Rutgers University in the US and other institutions has used up to 100 qubits, with repeated measurement and reset, to observe a phase transition from a state dominated by quantum chaos to one that can be controlled. The paper, by Bibek Pokharel, Haining Pan and colleagues, was published in the peer-reviewed journal Nature Physics on October 9, 2026. The result combines roughly 5,000 two-qubit operations with roughly 5,000 measurement-and-reset operations, and shows that intervening in a quantum system mid-computation can be used to test physical phenomena even at the 100-qubit scale. However, resetting every qubit to a known state is a different goal from quantum error correction, which must protect unknown quantum information.
The "roughly 5,000" figure counts operations in the circuit, not processing speed
The experiment ran on ibm_fez, an IBM Heron r2 processor with 156 physical qubits. The team selected up to 100 qubits connected in a line. In other words, 100 qubits is the number used in the experiment; IBM did not develop a new 100-qubit chip.
The circuit probabilistically chose between two kinds of operations: one that mixes information between neighboring qubits, and one that measures a qubit and returns it to the "0" state.
The measurements here are not the ones taken at the end of a circuit to read out the final result. While processing continues on other qubits, some qubits are measured, and the next operation depends on the outcome. What the team tested was whether this kind of mid-circuit measurement and control could run alongside a large number of quantum operations.
According to Appendix II of the published preprint, a circuit using 100 qubits, with the probability of choosing the control operation set near 0.5, contains about 5,000 two-qubit operations and about 5,000 conditional resets based on measurements. Near the critical point of the phase transition, a 100-qubit circuit runs a total of about 10,000 two-qubit operations plus measurement-and-reset operations.
| Operation | Approx. count in the 100-qubit circuit | What it does |
|---|---|---|
| Two-qubit operations | About 5,000 | Mix information between qubits |
| Measurement and reset | About 5,000 | Measure a qubit and return it to 0 based on the result |
| Total | About 10,000 | Sum of the two types above |
This total is the sum of the operation counts for the same conditions given in Appendix II of the preprint. Measurement and reset are counted as a single paired process.
It does not count every instruction, such as single-qubit gates, and it is not a figure for operations per second. Nor does it mean that error correction on logical qubits succeeded 5,000 times.
Measuring mid-circuit and switching operations based on the result demands more sophisticated hardware control than a circuit that simply runs quantum gates in sequence. The system must read out a measurement and carry out the corresponding operation while preserving the states of the qubits that were not measured. In Rutgers University's announcement, IBM's Maika Takita also stressed how difficult it is to repeat measurements while the rest of the processor keeps running.
At 50%, the balance between quantum chaos and control flips
Operations that mix information between qubits generate quantum chaos in the circuit. The chaos here is not a hardware fault or an accidental computational error. It refers to the complex spreading of quantum information, which the experiment creates deliberately.
The team asked how far measurement and control can return that spreading information to an ordered state.
In the control operation, a target qubit is measured, and if the result is "1," its state is flipped back to "0." If this is done often enough, the whole quantum system can be driven toward the all-zero state.
On the other hand, when the operations that mix information between neighboring qubits dominate, the system moves away from that simple state.
The probability "p" of the control operation determines which of the two operations is chosen.
The circuit has a mechanism that explains why p = 0.5 is the dividing line. Each mixing operation moves the position of the next operation one step forward, and each reset moves it one step back. The position moves randomly, but when p is 0.5, forward and backward moves are equally likely and neither is favored. This balance point separates the state in which chaotic behavior persists from the state in which control produces order.
So the 50% figure represents the probability of choosing the control operation in this model. It is not a success rate for quantum error correction or an acceptable error rate.
Behind this mechanism is the Bernoulli map, a simple mathematical model of classical chaos.
If you repeatedly double a number between 0 and 1 and drop the integer part, tiny differences in the initial value grow larger over time. Combine this with a control operation that halves the number and moves it toward 0, and the two compete.
The team extended this mechanism to a quantum system and implemented it as a qubit circuit.
To look for a phase transition, they used indicators such as the average magnetization, which shows whether qubits lean toward 0 or 1.
They varied the number of qubits from 10 to 100 and, after organizing the data to account for differences in system size, found that it approached a common curve predicted by theory.
What matters as a physics experiment is that the team did not merely see measurement results change sharply; they confirmed that theory-consistent behavior appears across different qubit numbers.
Note that the control-driven phase transition near p = 0.5 is a different phenomenon from the "measurement-induced phase transition," in which the way quantum entanglement spreads changes.
In the phase diagram from theory and numerical calculations in the published preprint, the measurement-induced phase transition appears at around p = 0.3. At that stage, chaotic behavior remains even though the nature of the entanglement has changed.
Increasing the share of control operations further moves the system into a controlled state with almost no entanglement. A decrease in entanglement does not mean the same thing as having steered the system into a target state.
Checking that the change is not an artifact of noise, using fluctuations in measurement results
Even if the average magnetization matches theoretical predictions, that alone does not confirm a distinctly quantum phenomenon.
That is because, even if a quantum system were heavily affected by noise and behaved like a classical stochastic process, the averages alone could look similar.
So in addition to the average magnetization, the team looked at the fluctuations in magnetization that arise from measurement outcomes.
For comparison, they used a simulation based on matrix product states, which compute quantum states, and a classical model that assumes the phase information of the quantum state is completely lost after each gate.
They calculated how much the magnetization varies when the same circuit is run repeatedly, and then compared the distributions of measurement results across different circuits.
In effect, they tested whether the spread of results could be reproduced, not just whether the averages matched.
According to the preprint's analysis, on the chaotic side where p is below 0.5, the distributions obtained in the experiment matched the quantum model better than the classical model.
This suggests that it is difficult to explain the observed behavior simply as the result of noise destroying the phase information of the quantum state.
However, this conclusion holds within the range of models used for comparison. It does not prove that no classical computational method could reproduce the results.
In the region where the share of control operations is increased, the limits of the hardware also became apparent.
Ideal calculations predict a sharp peak where the magnetization fluctuation drops to zero. The actual quantum processor, however, showed no such peak.
The team cites device noise, especially qubit readout errors, as the main cause.
In the region where p exceeds 0.5, the influence of noise grows, and neither the quantum model nor the classical model fully reproduced the distributions obtained in the experiment.
The number of qubits handled on real hardware must also be distinguished from the scale of the classical simulations.
In Figure 2 of the preprint, the hardware experiments went up to 100 qubits, while the simulations using matrix product states went up to 40 qubits. For some observables, however, the team compared results up to 100 qubits using a statistical mechanics model.
It therefore cannot be said that "because it uses 100 qubits, classical computers cannot verify it."
How hard a calculation is for a classical computer depends on which quantum states or observed results one tries to reproduce, and to what precision.
Applying this to quantum error correction requires measuring without losing information
The goal in this experiment was to return qubits to the known "0" state.
In contrast, what quantum error correction must protect is the unknown quantum information used in a computation.
If you measure all qubits and reset them to 0, you can certainly create an ordered state. But in the process, the quantum information that should be preserved is lost as well.
Understanding this difference in purpose is important when considering how the result relates to fault-tolerant quantum computing.
In quantum error correction, information is distributed and encoded across multiple physical qubits to form a "logical qubit." Information for detecting errors that have occurred is then measured, without directly reading out the stored quantum information.
IBM's educational material also explains a method of transferring information about errors to an auxiliary quantum system and measuring that auxiliary system.
One must also deal with the possibility that measurement itself introduces new errors, and with the problem of errors spreading to other qubits through quantum gates.
What this experiment shares with quantum error correction is the technical element of measuring mid-computation and repeating processing according to the results.
This study does not show that logical qubits were protected or that error rates were reduced.
Still, it is significant that large numbers of measurements and feedback operations were coordinated on a real quantum processor, yielding statistical data usable for testing a physical phenomenon. The repeated measure-and-control operation that quantum error correction also requires was executed in a large circuit.
Note that the results shown in Figure 2 of the preprint were obtained without error mitigation, which corrects for noise after measurement. This does not mean, however, that every individual quantum gate and measurement was executed without error.
The work was released as a preprint on September 22, 2025, and was also presented at the American Physical Society (APS) meeting in March 2026. Its publication in Nature Physics means these results have now been formally published in a peer-reviewed journal.
The numerical data used in the figures can be accessed from the Nature Physics page, while the full experimental and numerical data and the code are available from the authors on request.
To develop this kind of repeated control technique into practical quantum error correction, researchers will need to show that they can keep measuring while maintaining encoded quantum information, and suppress logical error rates, including the errors generated by the measurement and control themselves.
Can the sheer number of operations, around 10,000, be turned into the ability to keep long computations running correctly while protecting quantum information? That is the question this research leaves for next.