IBM and Rutgers Controlled Quantum Chaos on 100 Qubits With Thousands of Mid-Circuit Resets. The Experiment Was Not Quantum Error Correction.

October 9, 2026 – Nature Physics published an experiment in which researchers from IBM, Rutgers University and collaborating institutions controlled quantum chaos on a chain of up to 100 superconducting qubits using mid-circuit measurements and resets. The team ran its circuits on ibm_fez, a 156-qubit IBM Heron r2 processor, and applied no error mitigation to its steady-state results.

Bibek Pokharel of IBM and Haining Pan, then a postdoctoral researcher at Rutgers, were co-first authors. Jedediah Pixley of Rutgers and Maika Takita of IBM jointly supervised the work. The authors reported a sharp transition between chaotic and controlled behavior at the point where resets and scrambling operations occur equally often.

Each circuit applied local operations one at a time along the chain. With probability 1 − p, an operation scrambled two neighboring qubits with a randomly chosen two-qubit gate. With probability p, it measured a single qubit mid-circuit and flipped it back to zero if the result was one.

The position of each operation followed a random walk, moving one site to the right after a scrambling step and one site to the left after a reset. In the classical limit, the circuit reproduces the Bernoulli map, a textbook chaotic system that doubles a number and keeps its fractional part, competing with a control map that halves the number.

When resets were rare, scrambling dominated and the average magnetization of the chain stayed near zero. When resets were frequent, they drove the chain toward the state in which every qubit reads zero. The authors measured chains of 10 to 100 qubits and placed the transition at p = 0.495.

The correlation-length exponent, which sets how quickly the transition sharpens as chains get longer, came out at 1.05. Noiseless matrix product state (MPS) simulations, a classical method that stays efficient while entanglement is limited, gave 0.4947 and 1.05 for chains of up to 40 qubits. A statistical mechanics model, which the authors ran classically for chains of up to 100 qubits, gave 0.4970 and 1.0000.

At the transition, the measured magnetization grew over time with a dynamical exponent of 2.00, the value expected for an unbiased random walk.

The team also measured the shot-to-shot variance of the magnetization within each fixed circuit and how that variance changed between circuits. The shot-to-shot fluctuations come from the randomness of quantum measurement outcomes and vanish in the model’s classical limit, where a fixed circuit always produces the same outcomes. Near the transition, they broadly followed the predicted scaling, although some fitted exponents differed from the simulations by more than their quoted uncertainties, most clearly the dynamical exponent for the fluctuations, 2.08 against 2.00.

To test whether ordinary noise could produce similar fluctuations, the authors compared the measured distributions with the MPS simulation and with a classical model in which every qubit is fully dephased after each gate, adding a readout-error model to both. For chains of 20 to 40 qubits, in the chaotic phase and at the transition, the authors found the data closer to the quantum simulation than to the dephasing model. Both models still showed a measurable divergence from the data.

The authors reported applying up to nearly 5,000 entangling gates and 5,000 non-unitary mid-circuit operations on systems of up to 100 qubits. The work first appeared as an arXiv preprint in September 2025, and Nature Physics accepted it on August 26, 2026.

Rutgers announced the paper on the day of publication under the headline “Scientists Move a Step Toward Quantum Computers That Can Correct Their Own Mistakes.” The university said the experiment was one of the biggest demonstrations so far of quantum feedback control in real time, a capability it described as necessary for fault-tolerant machines.

Pixley, a physics professor and director of the Center for Materials Theory at Rutgers, said in the release that a fault-tolerant machine will have to repeat such checks and corrections many times during a single calculation. He described the experiment as a landmark toward that goal.

Takita, IBM’s principal research scientist for quantum error correction and dynamic circuits experiments, said the experiment showed that current hardware can interleave gates, measurements and resets thousands of times on a large processor. The release also said the IBM processor let the team test its theory on up to 100 qubits, well past the roughly two dozen qubits the researchers could simulate on conventional computers.

IBM physicist Barbara Jones, a co-author who, according to the release, helped turn the theory into an experiment, died while the manuscript was in preparation.

My Analysis

Pokharel, Pan and their colleagues asked whether a classical recipe for taming chaos still produces a sharp transition when the system is a quantum processor and the control comes from mid-circuit measurements. On 100 qubits they got a clear answer, with a critical point and correlation-length exponent that agree with the noiseless simulations within their quoted uncertainties, and a measurement of quantum fluctuations that needs no post-selection.

IBM’s Heron hardware ran circuits containing thousands of conditional resets across 100 qubits, with no error mitigation, and the averages it produced matched the theory. The team encoded no logical qubits in an error-correcting code and decoded no error syndromes, so the result gives no quantitative reason to revise any fault-tolerance or CRQC timeline.

Thousands of Mid-Circuit Resets Without Error Mitigation

According to the published supplement, a 100-qubit circuit at the critical point interleaved close to 5,000 measurement-conditioned resets with about 5,000 two-qubit gates. At p = 0.2, the two-qubit gate count peaked near 8,000. The supplement states L² time steps per circuit, which gives those counts, and the L²/2 in the main text would halve them. The paper does not reconcile the two. The team ran 50 random circuits for each value of p, with 1,000 shots each, and reported the steady-state results without error mitigation. That is a heavy workload for a dynamic-circuit stack, and IBM’s control system returned data clean enough for finite-size scaling across all ten system sizes.

Measurement-induced phase transitions have been hard to scale. The 2023 experiment on IBM hardware by Koh, Sun, Motta and Minnich used up to 14 qubits. Google Quantum AI reached 70 qubits only by mapping the problem onto circuits with no mid-circuit measurements at all. The obstacle is post-selection. Detecting an entanglement transition directly requires quantities that need the same random sequence of measurement outcomes to recur. The chance of any particular sequence shrinks exponentially as measurements are added.

Pixley’s group designed around that obstacle. The control transition at p = 0.5 changes the average magnetization, which every shot contributes to, so the team needed no post-selection to measure it. The trade-off is that the experiment never measures the model’s entanglement transition, which the theory places near p = 0.3.

According to the release, the researchers called this the largest successful demonstration of the approach to date, and the closest comparison I found supports them. A group at the National University of Singapore reported feedback circuits on 100 IBM qubits in a preprint posted in April 2026. But those circuits were about ten layers deep and used sparse measurements, far from 5,000 resets per circuit.

The Preprint, the Journal Version and the Press Release

The September 2025 preprint ended its abstract by calling the experiment “a signpost on the route towards fault tolerance.” In the journal version, the authors kept the sentence about 5,000 gates and 5,000 mid-circuit operations but ended it on “accurate estimates of the universal critical properties.” The introduction still calls the development of hardware that performs mid-circuit measurement and feedback quickly and accurately “the crucial next step needed to move to the fault-tolerant era of quantum computing.” The conclusion calls the coordination of gates, measurement and feedback “a crucial ingredient to reach fault tolerance.” Both statements are true of the measurement-based error correction that IBM and Google are building.

Where the preprint described the comparison with the dephasing model as certifying quantum behavior, the published text presents it as evidence, and the preprint’s two claims of first-time observations are gone. The authors also added a readout-error model to the comparison and moved its system-size scaling from the supplement into the main figure.

Rutgers went the other way. The university’s press office put error correction in the headline and called the result critical progress toward bigger and more reliable quantum computers, and Pixley’s quotes framed the work as a landmark on the road to fault tolerance. Takita’s quote is the most precise sentence in the release. Its description of thousands of coordinated operations, measurements and resets matches the gate and reset counts in the published supplement.

The release does not claim that the experiment corrected errors. Its own explanation presents error correction as a capability that fault-tolerant machines will need and says that nobody has built such a machine yet. My objection is to emphasis. Rutgers headlined a capability the experiment did not test, and the paper makes no claim to it.

What Quantum Error Correction Requires From Mid-Circuit Measurement

Quantum error correction spreads each logical qubit across many physical qubits and checks it continuously. In a surface code, or in the qLDPC codes IBM plans to use, ancilla qubits are entangled with groups of data qubits and then measured, on a fixed schedule that repeats every cycle and covers the whole code. That repeated measurement is syndrome extraction.

A decoder then turns the stream of syndrome bits into an estimate of which errors occurred, and it has to keep pace with the hardware. Many corrections can be tracked in software, but operations that depend on a decoded result, such as the corrections that follow a magic-state injection for a non-Clifford gate, have to wait for the decoder’s answer. Gidney’s 2025 RSA-2048 estimate, for example, assumes a control-system reaction time of 10 microseconds for that wait, from measurement through decoding to the dependent operation, a figure I discussed in my September piece on the classical loop.

The adaptive Bernoulli circuit exercises one piece of that loop. Each reset measured one qubit and flipped it back to zero if the result was one, and the circuit applied a single operation per time step. No logical qubits were encoded, and no decoder ran.

On measurement volume alone, error-correction experiments are already far ahead. Google Quantum AI decoded a distance-5 surface code in real time for up to a million cycles, with an average decoder latency of 63 microseconds and a cycle time of 1.1 microseconds. In September I reviewed six teams’ work on the classical side of that loop, and none of them had closed the full loop, decoder output included, on a running processor at code distance 9 or above; the one distance-9 loop ran on emulated measurements. The Rutgers and IBM team was not trying to close that gap.

So the experiment is a systems test of IBM’s dynamic-circuit stack under a heavy load of measurement-conditioned operations, and on that test IBM’s hardware did well. Syndrome extraction is a different test, with many measurements in parallel every cycle and the results routed to a decoder.

Classical Simulations of the Same Circuits

Rutgers said the processor took the theory well past the roughly two dozen qubits the team could simulate on conventional computers. The figure matches the team’s 2024 theory paper, which, according to the Nature Physics paper, established the transitions for chains of up to 24 qubits. Simulating every amplitude of a 100-qubit state is out of reach for any classical computer. The quantities this experiment measured need far less. The authors’ own MPS simulations cover chains of 10 to 40 qubits, and their statistical mechanics model reproduces the average magnetization and the critical exponents for chains of up to 100.

At the transition, the authors report 103 seconds on an AMD EPYC 9654 to evolve one 40-qubit trajectory through 2L² steps, and a run time that grew roughly as $$L^{3.15}$$ over the sizes they simulated. MPS simulation becomes expensive in the volume-law phase below p ≈ 0.3, where entanglement grows with system size. At p = 0.4, inside the area-law phase, a significant number of the authors’ trajectories already hit their bond-dimension cap. The MPS comparisons with the experimental data start at p = 0.4.

The volume-law phase has the least to check the device against. The average magnetization there stays near zero, as the classical statistical model predicts, and the authors note that the same model does not capture the full distribution of fluctuations inside the phases. The authors write that classical simulations have difficulty producing accurate results for large chains in that phase and that the quantum computer explores the full phase diagram efficiently. They draw that conclusion from their MPS bond dimensions, and they make no claim that the quantities they measured are beyond classical computation.

The Evidence for Quantum Behavior

The fluctuation distributions are the paper’s evidence that coherent quantum dynamics contributed to the data. In the model’s classical limit, where a fixed circuit always produces the same measurement outcomes, the shot-to-shot fluctuations vanish, so the authors treat their presence as a sign of quantum dynamics, unless ordinary noise can produce similar fluctuations. They tested that possibility against a model in which every qubit is fully dephased after every gate, which turns the circuit into a classical random process. To compare the distributions, they used the Kullback–Leibler divergence, a standard measure of how far one probability distribution is from another.

The quantum simulation fit better, by a modest margin. Both simulations included the readout-error model, and for chains of 20 to 40 qubits, in the chaotic phase and at the transition, the divergence between data and MPS simulation stayed below the divergence between data and the dephasing model. The authors call the difference subtle, and every model kept a measurable divergence from the data. In the controlled phase, none of the models matched. The authors attribute that chiefly to device noise. Readout error in particular erases the zero-variance peak that a noiseless chain would show.

That comparison ends at 40 qubits because the MPS simulations end there. The 100-qubit observables follow the predicted scaling, but the classical statistical model reproduces those exponents as well, so at 100 qubits nothing in the paper separates coherent dynamics from incoherent noise. My reading is that the comparison supports coherent dynamics in the tested range and certifies neither the full 100-qubit state nor any computational advantage.

The published supplement plots the estimated duration of every circuit. For 100 qubits near the transition, the authors estimate that the longest circuits ran for about 2 milliseconds, some 20 times the mean T2 coherence time of about 100 microseconds that the supplement lists for the qubits used. The circuit keeps resetting and re-scrambling qubits, so each superposition only has to last part of the run. But the paper does not quantify how idle time, scheduling and feedback latency shape the fluctuation signature.

I found no mention of dynamical decoupling on idle qubits in the paper or its supplement, which leaves open whether it was used. How those hardware details affect the quantum signal is the first thing I’d ask the authors.

Critical Exponents and Hardware Error Rates

Critical exponents are universal. They depend on the symmetries and dimensionality of a problem, and the same values describe very different physical systems, which is why the random-walk exponents of this transition, 1 for the correlation length and 2 for the dynamics, could be predicted before the experiment ran.

The authors checked how much noise their averaged results tolerate. In the published supplement they added depolarizing errors to their statistical model, up to 3% per two-qubit gate and 30% on each measured qubit, and the transition stayed sharp, at p = 0.504 with an exponent of 1.000. On the device itself, they report average errors on the chosen qubits of 1.8% per two-qubit gate and 1.3% per readout. Their plot of the distribution puts the typical two-qubit error near 0.3%, with a small tail between 2% and about 20% that pulls the average up.

From a matching exponent we learn that the averaged observables fall in the intended universality class. A matching exponent does not show that every gate implemented the intended dynamics, and it does not measure gate quality. The team also did not test whether this chip can run an error-correcting code below threshold. Hardware noise dominates the measured fluctuations in the controlled phase, according to the authors. Outside that phase, the fluctuation evidence for quantum dynamics covers only chains of 20 to 40 qubits, and the authors call the quantum simulation’s margin over the dephasing model subtle.

IBM’s Fault-Tolerance Timeline

IBM lists two error-correction targets for 2026 on its roadmap: Kookaburra, a module that pairs a logical processing unit with quantum memory, and a prototype of the real-time decoder its error correction will need. IBM’s fault-tolerant target is still Starling, planned for 2029 in Poughkeepsie, New York, with 200 logical qubits capable of 100 million quantum operations.

The adaptive Bernoulli experiment is not one of those targets, and the paper never presents it as one. The team did show that IBM’s control stack sustains thousands of measurement-conditioned operations in a single circuit with the physics intact, and both 2026 items depend on that capability.

For a headline about quantum computers correcting their own mistakes, I’d want evidence that encoded information was protected, such as logical error rates that fall as codes grow, with syndromes decoded in real time on a running processor. IBM’s decoder prototype and Kookaburra are both steps toward showing that on IBM’s own hardware.

From the author