acceptodds
Under review as a conference paper at ICLR 2027

Learning When and How to Use Measurement Feedback in Dynamic Quantum Circuits

Abstract

Mid-circuit measurements let a circuit steer later operations with information it has just created. Yet the feedforward rule that maps outcomes to gates is still designed by hand, one bespoke rule per protocol, while measured qubits create outcome contingencies. External controllers move it outside the circuit, leaving open when conditioning helps. We make the feedforward rule a trainable part of the circuit. Exact branch decomposition enumerates the measurement outcomes inside the optimization loop, so the rule angles and every circuit parameter descend one balanced branch objective together, with no sampling estimator. It also yields analytic floors that bind any parameter-matched circuit ignoring the outcome, a spectral energy gap and a chance-level classification bound, which certify before training whether the outcome can change the attainable objective. Across energy and classification tasks the learned rule crosses those floors, clears every unconditional variant, and at matched budgets and restarts ties or exceeds the best baseline on of grid cells. Under a hardware-informed deployment model its net value is task dependent, repaying the feedback latency on classification while energy recovery crosses to a static preference. A mid-circuit measurement is thus a computational resource whose value can be certified and whose use can be learned.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
dynamic quantum circuits, mid-circuit measurement, measurement feedforward, variational quantum algorithms, differentiable quantum programming, quantum machine learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_vbDzOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes an end-to-end learning framework for measurement feedback in dynamic parameterized quantum circuits (DPQCs). The circuit consists of a pre-measurement block U1(θ1)U_1(\theta_1), a mid-circuit measurement with outcome mm, outcome-dependent gates parameterized by gα(m)g_\alpha(m), and a post-measurement block U2(θ2)U_2(\theta_2). Instead of manually specifying the feedback rule, the authors jointly optimize the feedback parameters and the circuit parameters. To handle the discrete measurement outcome, the paper uses exact branch decomposition, explicitly evaluating all measurement branches and optimizing a balanced branch objective. The paper also derives analytic lower bounds for outcome-independent circuits and evaluates the approach on VQE ground-state preparation and classification tasks, together with studies of different architectures, measurement configurations, readout noise, and feedback latency.

Critical strengths

I do not identify a major strength that compensates for the fundamental problem-formulation issue discussed below. The paper contains substantial theoretical and experimental analysis within its assumed DPQC model, but in my view the significance of these results depends on first establishing that the assumed mid-circuit measurement addresses a meaningful need in the target application.

Critical weaknesses
  1. The main problem formulation appears artificial for the targeted VQE application. The paper assumes a VQE architecture containing a mid-circuit measurement and then proposes to learn how to respond to its outcome. However, unlike quantum error correction, fault-tolerant protocols, or measurement-based quantum computation, the VQE problem considered here does not intrinsically generate or require such an intermediate measurement. The paper does not establish that introducing this measurement is necessary or advantageous for solving the underlying ground-state-preparation problem. Consequently, the measurement creates the branch-dependent feedback problem that the proposed method subsequently solves.

  2. The theoretical advantage does not establish the usefulness of mid-circuit measurement for VQE. The analytic floor compares outcome-dependent processing against outcome-independent processing after the measurement has already occurred. Once a measurement has generated orthogonal branches, it is natural that allowing different operations on the branches can outperform forcing the same operation on both. However, an ordinary VQE circuit that does not perform the intermediate measurement does not suffer this measurement-induced restriction. Therefore, crossing the proposed floor establishes the usefulness of the measurement record given that the measurement was performed, not the usefulness of performing the measurement itself.

  3. The paper lacks the baseline needed to support the VQE motivation. The relevant comparison should include a resource-matched VQE circuit without mid-circuit measurement and, ideally, a coherent conditional circuit that preserves the quantum information rather than measuring it and converting it into a classical branch variable. Without these comparisons, it is difficult to determine whether the proposed architecture provides any benefit over solving the original VQE problem directly.

  4. The training objective is not generally the physical expected loss. The paper optimizes \frac{\ell_0+\ell_1}{2}, whereas actual executions encounter the branches with probabilities p0p_0 and p1p_1, giving p_0\ell_0+p_1\ell_1. The paper explicitly distinguishes these quantities. When branch probabilities are unequal, the two objectives may favor different solutions. This further weakens the connection between the proposed optimization problem and the physical VQE objective.

Questions for authors

My main concern is with the premise of the work rather than the optimization technique itself. Before asking how to learn a feedback rule following a mid-circuit measurement, the paper needs to establish why such a measurement should be present in the target VQE problem. In applications such as quantum error correction, syndrome measurements arise naturally and independently of the feedback algorithm; therefore, learning a response to those outcomes constitutes a meaningful problem. I do not see an analogous motivation established for the VQE task studied here.

The current results show that once a mid-circuit measurement is imposed, using its outcome can outperform ignoring its outcome. This is not sufficient to demonstrate that mid-circuit measurement is beneficial for VQE. Indeed, the measurement itself creates the orthogonal branches responsible for the outcome-independent floor. I therefore encourage the authors to compare against a resource-matched unitary VQE circuit with no intermediate measurement and against an appropriate coherent-control baseline. The central question should be whether introducing measurement and feedback improves the solution of the original ground-state-preparation problem, rather than whether feedback helps after measurement has already been imposed.

Alternatively, I believe the proposed learning framework would be substantially better motivated in an application where intermediate measurement is intrinsically required, such as quantum error correction or a fault-tolerant measurement-based protocol. In such a setting, the problem “given an unavoidable measurement outcome mm, learn the optimal response g(m)g(m)” would have a clear application-level meaning.

In its current form, I view the problem-formulation issue as fundamental. The technical results may be valid within the assumed DPQC model, but the paper has not established that solving this assumed feedback problem advances the target VQE application. For this reason, I would not consider additional optimization results within the same assumed architecture sufficient to address my primary concern; the authors would first need to justify the architecture itself or move to an application in which the measurement has an independent purpose.

2 comments
AuthorsOpenReview ↗

Part 1 of 2

We thank the reviewer for the report. The standard it sets is clear: justify the architecture, or move to a setting where the record has a purpose. We answer on both sides. We start with the contribution, because the disagreement is about whether it survives the formulation.

The paper contributes three things. A certificate computed from the task before training, which says whether the record can change the attainable objective at all, with no analogue in a hand rule or a trained decoder. A rule that lives inside the circuit as four angles, trains without a sampling estimator, and beats a tuned 12-rule hand family by 1.72x at h=2. A measured boundary, where the channel pays when the measured bit carries label information, 0.175 bits on MNIST, and stops paying when it does not, 0.006 bits on CIFAR. The scope is explicit: one to three measured qubits, state-vector simulation, no hardware claim. The abstract will say this in its first sentence, and that framing was ours to make explicit.

W1. The object is the circuit family, not VQE alone. The premise is correct for VQE taken alone, and the paper does not rest on it. The object is the dynamic parameterized circuit family, whose definition includes one mid-circuit measurement and which is provably at least as expressive as unitary circuits of equal depth. VQE is the benchmark we instantiate it on, not an application claim, chosen because its floor has a closed form in the Hamiltonian spectrum. The question is conditional: given a record, when does using it pay, how much, and can the rule be learned.

W2. The floor prices the record, and a unitary circuit has no such price. It separates using the record from ignoring it once the measurement exists, which is the certified quantity. It is not a restatement of the design: it holds over the whole m-independent family, is attained exactly, and moves with the payload spectrum, 0.184, 0.732, 1.507 at one, two and three measured qubits. The measurement does not create a problem for the method to solve; it defines a resource class with a computable ceiling, and the method reaches that ceiling. A coherent circuit has no record, hence no branch class to bound, so the certificate is not defined for it rather than contradicted by it.

W3. The baselines are added, and they do not subsume the reset heads. The stronger claim, that the dynamic circuit solves the ground-state problem better than a unitary VQE at the same budget, is a different claim and we will measure it: a no-measurement arm at matched depth, gate count and restarts on all nine energy cells. The existing outcome-independent arms hold depth and gate count and drop only the dependence on the record, 0.620 and 0.552 against 0.200, but they still project at the mid-point, so they bound the value of the record rather than the cost of the projection. The coherent arm is added on all 27 cells and reported beside it. Neither subsumes the two reset heads, where measure-reset-condition has no unitary counterpart and needs an ancilla.

W4. Two quantities, two jobs, and they coincide on 18 of 27 cells. In the phase head p_m=1/2 exactly for both classes, and in XOR branch m is the sample, so L_bal is the empirical risk rather than a surrogate. The nine energy cells are where they differ, and there the physical cost is in the deployment table.

AuthorsOpenReview ↗

Part 2 of 2

Part 2 answers the four questions and reports the syndrome experiment.

Q1. Why a measurement at all. The paper does not answer this by asserting that VQE needs one. Where the measurement is not required, it does not claim it should be added. It answers a question well posed in any architecture: given a record, what is it worth, and can the response be learned. The measurement is the family's primitive and the floor is what makes its value decidable before training.

Q2. Is the unitary comparison the central question? Yes, and it belongs in the headline table rather than an appendix. W3 adds both arms, and the revision reports them in the first table of the energy section. On the two reset heads the same question has no coherent counterpart.

Q3. An application where the measurement is intrinsic. The setting you recommend is already two of our three task families: in the phase and XOR heads the measured bit is the only carrier of the label. We also ran the syndrome task, a three-qubit repetition code with one round of bit-flip noise at rate p, two ancillas measured mid-circuit, and a 12-angle recovery table indexed by the observed syndrome. Training uses the exact logical error, which is computable here because every error pattern is a basis state, so there is no surrogate mismatch either. The learned rule matches the Bayes decoder in every setting we ran, largest gap 1e-4, and it finds that decoder from the logical error alone, without being told the code or the noise model.

preadout qno recoverytextbookBayes handlearned
0.0500.1360.0070.0070.007
0.1000.2440.0280.0280.028
0.2000.3920.1040.1040.104
0.050.100.1360.1780.1020.102

At q=0 the learned table is majority vote. With readout error it does something a textbook rule does not: at p=0.05, q=0.10 it corrects only the syndrome that noise cannot easily fake, so it beats the textbook rule, 0.102 against 0.178, and beats doing nothing, 0.136. The same ordering holds at q=0.20 and p=0.10, where it reaches 0.204 against 0.316 textbook and 0.244 no recovery.

Q4. Would further results within the same head help? No, and the revision is built accordingly. W3 changes the comparison class, Q3 changes the application. Neither is another optimization result inside the assumed head.

The certificate is the piece that outlives any single arm. It is computed from the task rather than from the trained circuit, and it transfers unchanged to the syndrome setting, where the record exists by necessity.

Reviewer_nBA3OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper makes the feedforward rule of a dynamic circuit a small trainable lookup table of four angles for a single measured qubit. The rule is trained jointly with all circuit parameters by exactly enumerating both measurement branches under a branch-averaged loss. The paper proves "floors" Ky Fan for energy, ln⁡2\ln 2 for classification that bound any circuit ignoring the outcome. It evaluates on a 3×3×33\times3\times3 grid task ×\times architecture ×\times depth at n=8n=8, with scaling to n=20n=20, readout noise, and a simulated "deployment" model. I have listed many weaknesses, but they are not equally important. If the authors address the most important ones i.e. the coherent control baseline, the mismatch between the balanced and the deployed cost, a complete and protocol-consistent 27 cell table, and the scope of the floor claims; and I find the response satisfactory, I am willing to raise my score.

Critical strengths
  • The Ky Fan argument for the energy floor is correct and neat.
  • The experients of the paper are quite broad covering grid, scaling, multiple measurement forms, readout noise and ablations such as estimator, placement and shared budget.
  • The authors honestly discuss limitations. The scope of the paper stops at three measured qubits as shown in Appendix A, and family parity on the two-angle head is called "structural" in Appendix F.
  • The branch-collapse analysis is a useful observation.
Critical weaknesses
  1. The floors follow from task design rather than serving as an independent certificate. (1a) In both classification tasks the measured bit is reset and is the only carrier of label information. The ln⁡2\ln 2 floor therefore says only that a model without label information cannot predict the label. (1b) The paper does not show a nontrivial floor on a task that is not engineered this way, which "certify before training" would require. (1c) The energy floor is loose. At n=8n=8, unconditional circuits sit 1.21.2–4.2×4.2\times above it. At n=20n=20 the floor falls to 0.0770.077 while unconditional circuits sit at 0.810.81–1.111.11. (1d) The abstract says the learned rule "crosses those floors". In the main h=1h=1 grid with one measured qubit, the learned gap 0.2000.200 is above the floor 12(E1−E0)=0.184\tfrac12(E_1-E_0)=0.184. The paper concedes this for two-branch forms. Crossing is shown only at h=2h=2 or with 22–33 measured qubits.

  2. I believe a coherent-control baseline is missing. (2a) A mid-circuit measurement followed by classically controlled gates is a dephased version of a controlled gate [Nielsen, Michael A., and Isaac L. Chuang. Quantum computation and quantum information. Vol. 1. Cambridge: Cambridge university press, 2000.]. A unitary circuit with controlled rotations from QQ to the readout can match or beat it without any measurement. (2b) Every "unconditional" control either removes the channel or erases the information with a reset. (2c) I believe that the claim that a mid-circuit measurement is a computational resource needs a parameter-matched coherent-control comparison.

  3. The trained objective is not the deployed cost. (3a) Training uses the unweighted branch average Lbal=12∑mℓm\mathcal{L}_{\text{bal}}=\tfrac12\sum_m \ell_m. The deployed cost is ∑mpmℓm\sum_m p_m \ell_m, which the paper acknowledges is a different quantity. (3b) The floors bound the balanced objective only. A weighted floor would depend on min⁡mpm\min_m p_m, so the certificate does not transfer. (3c) Equal weighting of rare branches produces gradients scaling as 1/p11/\sqrt{p_1}. In the scan, one collapsed initialization (p1=1.5×10−5p_1=1.5\times10^{-5}) moves a single parameter's gradient by about 10310^{3} and raises the averaged variance statistic to 104.610^{4.6}–10510^{5}. The "self-limiting" claim rests only on the absence of spikes in the reported runs. (3d) It is unclear whether the tables report the balanced or the weighted quantity, and the weighted-cost penalty of balanced training is never shown.

  4. The comparisons with hand rules are weaker than claimed in the paper. (4a) A hand rule is a point of the learned family (e.g. αm=πm, βm=0\alpha_m=\pi m,\ \beta_m=0), so it cannot beat the learned rule at the global optimum. On the main energy task the gain is negligible (0.2000.200 vs 0.2040.204). (4b) The strict win in Table 2 compares a continuous eight-parameter head against hand angles restricted to a discrete set ({0,±π/2,π}\{0,\pm\pi/2,\pi\}). A fairer baseline would fix part of the rule by hand and train the rest, or give the hand family continuous angles. (4c) In the R∈{1,2,3}R\in\{1,2,3\} sweep (Table 8), the learned rule is strictly best in only 33 cells, co-optimal in 1818, and behind in 66. It is clearly worse than manual in places, e.g. ring R3R3 (0.3010.301 vs 0.1460.146) and brick R2R2 (0.4540.454 vs 0.2280.228).

  5. The methodological novelty is not shown to matter, and the scope of the paper is small. (6a) In Figure 9a exact evaluation and a straight-through surrogate give identical final costs. The authors say exactness wins "on variance and shot requirements, not on the optimum". (6b) With a binary outcome any estimator is cheap. The regime where estimators fail (large kk) is the one the method cannot reach, since the cost is 2k2^k state vectors and the study stops at k=3k=3. (6c) The motivation is 2k2^k outcome contingencies, yet the method scales exactly like that. The paper itself says external decoders "remain the right tool" at large registers. (6d) There is no comparison with the external learned approaches it positions against such as RL policies, neural decoders or with Gumbel-softmax.

  6. Real-data and deployment evidence is quite weak. (7a) The learned rule reaches 0.945±0.0000.945\pm0.000, exactly the classical all-features ceiling. The zero variance suggests saturation. The CIFAR variant shows no gain. (7b) The deployment model has three anchor points, one 88-qubit ring, and 33 seeds ×\times 3232 trajectories. The feedback delay of 0.10.1–0.2 T20.2\,T_2 is not tied to any device. Everything is simulated, so "hardware-informed" is overstated. (7c) The task-dependent crossover amounts to "the benefit disappears as noise grows". It is not a mechanism or a map. (7d) "CPU-sized results transfer to hardware-roadmap widths" is not supported by a one-measured-qubit state-vector study up to n=20n=20.

  7. There are several technical and presentation issues. (8a) Proposition 3 assumes E[g0g1]≥0\mathbb{E}[g_0 g_1]\ge 0, which the authors show fails in a natural flip-symmetric case. It handles the preparation-layer dependence of the normalized payload only by restricting to shallow depth (R≤6R\le 6). (8b) Gradient scan covers n≤16n\le 16 only. The rule-channel variance is 0.80.8–1.01.0 orders of magnitude lower than the baselines, with no discussion of whether this is harmful. (8c) Proposition 1 proof in the appendix; the γ\gamma-overlap argument is looser than the Ky Fan trace argument and should be replaced by it. (8d) In XOR data encoding the statement "a 5-bit, 32-point dataset stored in the register's amplitudes" is unclear. It does not say how single samples are classified or what pmp_m means there.

  8. The paper needs an improvement in clarity in the abstract and introduction. In Abstract: (a1) Abstract: Many sentences stack three or four ideas (e.g. "exact branch decomposition ... no sampling estimator"). Phrases such as "crosses those floors," "clears every unconditional variant," and "crosses to a static preference" are metaphors that say little. (a2) "Certify before training" does not say how tight the certificate is. "Ties or exceeds on 2626 of 2727 grid cells" does not say what the grid is. (a3) There are no qubit counts, named baselines, or effect sizes. The mixed deployment result is stated without a reason, so it is unclear whether it is a finding or a limitation. In Introduction: (I1) The first paragraph opens with a broad claim about "hybrid learning systems", then spends a paragraph on a citation list that does not lead to the gap. Key terms such as "floor," "balanced branch objective," "unconditional," "measurement-independent," "grid cells" are used before they are defined, or interchangeably. (I2) The claim that existing approaches "neither keep pace with the hardware" has no support. The introduction does not say why the unweighted branch average is the right target instead of the expected loss. Exact enumeration works only for "small registers," yet the motivation is that the rule space grows as 2k2^k. The paper does not say how this limit squares with that motivation. (I3) In contribution bullet 1 contains a broken sentence "turning whether the channel can pay into a quantity ...". Bullets 1 and 2 are not parallel. Bullet 1 states a theoretical result, while bullet 2 only re-describes the mechanism already introduced in the paragraph before the list. Its only new content, the final clause which also duplicates bullet 3.

Questions for authors

Here I am going to ask questions regarding the most important points that needs to be clarified from the weaknesses:

  1. Can you give a task where the floor is nontrivial and the measured bit is not the only carrier of label information i.e. without reset? Without one, the certificate only restates the information removed by design.

  2. The energy floor is loose and at h=1h=1 with one measured qubit the learned gap stays above . Should the abstract's claim that the learned rule "crosses those floors" be restricted to h=2h=2 and to two or three measured qubits?

  3. By the principle of deferred measurement, measure-then-condition can be replaced by a controlled gate with identical readout statistics. How does the learned rule compare with a parameter-matched and qubit-matched coherent-control circuit on all three tasks?

  4. If a coherent controlled circuit matches or beats the learned rule, what remains of the claim that a mid-circuit measurement is a computational resource?

  5. Do the tables report Lbal=12∑mℓm\mathcal{L}_{\text{bal}}=\tfrac12\sum_m \ell_m or ∑mpmℓm\sum_m p_m \ell_m? What is the penalty in ∑mpmℓm\sum_m p_m \ell_m from training on Lbal\mathcal{L}_{\text{bal}}, and does a floor hold for the weighted cost?

  6. One initialization with p1=1.5×10−5p_1=1.5\times10^{-5} gives 1/p11/\sqrt{p_1} gradients and variance statistics up to 104.610^{4.6}–10510^{5}. What supports the "self-limiting" claim beyond its absence in the reported runs?

  7. hand rule is a point of the learned family, and the main energy gain is only 0.2000.200 vs 0.2040.204. Can you add a baseline that fixes some angles by hand and trains the rest, or a hand family with continuous angles, to make the Table 2 win meaningful?

  8. Can you report all 2727 cells under one fixed protocol same restarts and steps for every cell, with per-seed values and confidence intervals or paired tests? Nine cells are currently re-run with 66–1010 restarts and 2×2\times steps.

  9. Figure 9a shows that exact evaluation and a straight-through surrogate reach identical final costs, and the method costs 2k2^k state vectors with the study stopping at k=3k=3. In what regime does exact enumeration change the optimum, and how does the method address the large-kk setting that motivates it?

  10. Why are RL policies, neural decoders, and Gumbel-softmax relaxations not compared, given that the paper positions itself against them?

  11. The learned rule reaches 0.945±0.0000.945\pm0.000, exactly the classical all-features ceiling, on a constructed MNIST task. How were the delays and noise parameters in the deployment model chosen, and what mechanism explains the task-dependent crossover?

  12. The assumption E[g0g1]≥0\mathbb{E}[g_0 g_1]\ge 0 fails in a flip-symmetric case, and the XOR encoding ("a 5-bit, 32-point dataset stored in the register's amplitudes") does not say how single samples are classified or what pmp_m means. Can you clarify both, and define "grid cells" and "balanced" in the abstract?

5 comments
AuthorsOpenReview ↗

Part 1 of 2

We thank the reviewer for the report. The four items you name as score-bearing fall in Q1 to Q3 and Q8. We answer the 12 questions in order, with numbers where we have them and with the protocol where we do not. Q1 to Q6 are here, Q7 to Q12 in the next comment.

Q1. A task where the floor is nontrivial and the measured bit is not the only carrier. The energy family is that task. No reset, no label, and the floor comes from the Hamiltonian spectrum rather than from removed information: 0.184 at one measured qubit, 0.732 at two, 1.507 at three. At three measured qubits the learned gap is 0.24 to 0.44 while the outcome-ignoring arms sit above 1.507, so the conditional family passes under a bound those arms cannot. At one measured qubit the learned gap 0.200 does sit above 0.184, and there the certificate is silent rather than false. That spread is what makes the floor a property of the task rather than a restatement of the design.

Q2. The loose floor and the abstract claim. Tightness and validity are different properties. At the median the outcome-ignoring arms sit 3.0 to 3.4 times above the floor, and at h=2 the learned gap is below 1.095 while the control is 1.8 times above it. A bound the learned rule passes under where it is large is decisive whether or not it is tight, and we will report the ratio per family. The abstract will say "crosses" only where it is shown: the multi-branch forms (learned 0.21 to 0.44 against 0.732 and 1.507) and the stretched field (43 of 45 cells under 1.095).

Q3. The coherent comparison. By deferred measurement the measure-and-condition circuit is a dephased controlled circuit, so the coherent version is a strict superset: it keeps the branch coherence that the measurement destroys. A superset cannot certify the subset, it can only price the dephasing. Two facts. In the two reset heads no coherent circuit on the same register reproduces measure-reset-condition, since the reset is non-unitary, discards the state and needs a fresh |0>, hence an ancilla and a different gate count. And the conditional class is native where the record must be classical, syndrome extraction being the canonical case. We add the arm on all 27 cells, that is all three task families, parameter-matched and two-qubit-gate-matched, since the coherent rotation costs one extra entangling gate. One correction: on the energy task the parameter-matched control keeps the measurement and the branch split and differs only in not reading m, while the reset is the task's head in the two classification families, shared by every mode including ours.

Q4. What remains if the coherent arm matches. The certificate, which reads the task rather than the trained circuit, and the two reset heads, where there is no coherent counterpart at all. The record is also the only classical interface in the settings that motivate the paper, so a tie on the energy task would bound where the channel pays rather than remove the claim.

Q5. Which quantity the tables report. The main tables report L_bal. In 18 of 27 cells the two quantities coincide by construction: the phase head has p_m=1/2 exactly for both classes, and in XOR branch m occurs exactly on the samples with x0=m, scored on those samples only. The nine energy cells differ, and there the deployment table already reports the trajectory-averaged cost, which still favors the learned rule at the low-noise anchor. We add the weighted column and a weighted-trained arm, both for the same trained rules, so the penalty of balanced training is measured rather than argued. Sec 3.3 will state that the floor bounds L_bal and that a weighted floor depends on p_m and is not claimed.

Q6. What supports self-limiting. The 1/p_m enters the payload normalizer, not the value of the objective, so it rescales the θ1 gradient and leaves the α direction intact. The sum is exact, so there is no sampling variance to amplify, and Adam is invariant to a per-coordinate rescaling. The scan localizes the effect to one initialization in 200, and the manual, unconditional and no-gate modes share it, so it belongs to the task family at branch collapse rather than to this objective. The spike measures the statistic at initialization, not the trajectory. We add a recovery run from that exact initialization, reporting p1 and the cost over steps.

AuthorsOpenReview ↗

Part 2 of 2

Q7. A fairer hand baseline. A hand rule is a point of the learned family, which is why the h=1 tie is expected: 0.200 against 0.204. The eight-parameter optimum sits off the discrete grid, α=(0,-π/3) in the energy cells and antiparallel β in the phase cells, so no member of a {0,π}x{0,±π/2,π} family reaches it however long it trains. That is why the tuned family loses by 1.1x at h=0.5 and 1.72x at h=2.0. A continuous hand family with all eight angles free is the learned family, so the informative baseline is the partial one: we add both directions, fix α and train β, and fix β and train α.

Q8. One protocol for all 27 cells. Yes. We retire the two-protocol split: one restart budget, the same steps, all 27 cells, per-seed values and paired tests. The R in {1,2,3} table is the appendix's completeness table at a smaller budget and will be relabelled as such. One correction to the two cells you cite: ring 0.198 against 0.195 is manual ahead by 0.003, while brick 0.204 against 0.228 is learned ahead by 0.024.

Q9. When exact enumeration changes the optimum, and the large-k regime. It does not change the optimum where the surrogate is unbiased, and we say so: its value is that it removes the sample count S, the estimator variance and the shot cost. At one measured qubit the straight-through arm ties at every budget. At three measured qubits, where the 2^k cost first bites, we will add the same arm. Large k is the boundary we state plainly: the branch count of the objective and the size of the rule table are the same 2^k, so the method stops where the table stops being small, and the certificate does not, because it reads the task. Beyond that boundary external decoders remain the right tool.

Q10. Why RL, decoders and Gumbel are not compared. Gumbel-softmax is added on the same cells. RL policies and neural decoders are trained under a different sample budget and a different objective, and a 400-step comparison would misrepresent them; the honest comparison is a separate study, and we flag it as future work rather than a baseline.

Q11. The MNIST ceiling and the deployment parameters. The ceiling is the design: the label is a threshold on f1 whose threshold moves with m, so the bit's information is the only thing the channel can add, and reaching the classical all-feature ceiling is the calibration that it adds exactly that. On the parameters, the present three anchors were chosen to span the range where the ordering flips rather than to represent a device, which is why they become a 5x5 grid over (p_ro, τfb/T2) plus one device-anchored point from published readout and T2 numbers. The mechanism behind the crossover is headroom: a control pinned at ln2 has none, so the channel keeps paying on classification at every anchor, while the energy control has headroom above 0.184 and loses it as waiting grows.

Q12. The assumption, the XOR encoding, and the definitions. The assumption is stated in the proposition. Its failure is an exact cancellation, hence a saddle that a symmetry-broken initialization or the weighted term leaves, and the shallow-depth condition on the normalizer channel becomes part of the proposition's scope. The XOR encoding is one basis state per sample: 32 samples, qubit 0 carries x0, R_Y(πxi) prepares each bit, the measurement is deterministic given x0, and each branch is scored on the samples that take it. Grid cells and balanced will be defined at first use in the abstract, which also carries the qubit count and the two floors, and the contribution bullets are rewritten in parallel with bullet 1 repaired.

Reviewer_nBA3OpenReview ↗

Brief comment to the Authors

Thank you for the detailed response particularly highlighting answers for each and every questions. You indeed provide useful clarification specifically regarding the spectral origin of the energy task floor, the settings in which the learned rule crosses the floor, the distinction between balanced and probability weighted objectives, and the XOR sample encoding.

But my principle concern remain mostly unresolved based on the evidence that is currently available. Specifically, the resource matched coherent control comparison in Question 3 and the uniform protocol evaluation in Question 8 are indeed proposed but not reported in the response. I believe these are important in assessing the interpretation of the advantage of the measurement condition and the robustness of the comparative results. The stronger partial training baselines and the weighted-objective comparisons similarly remain speculative.

No revised manuscript have been provided for these additions, so I cannot assess whether they support the existing conclusions or require further changes. My whole concern is bsed not the absence of a revision by itself, but the absence of the evidence needed to resolve the central issues.

Accordingly, I retain my current score based on the available submission and rebuttal.

Reviewer_ZmrxOpenReview ↗
Rating
3
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on understanding when and why measurement-based feedback is useful in dynamic parameterized quantum circuits (DPQCs), and how such feedback can be learned automatically rather than manually designed. The authors propose an approach in which the outcomes of intermediate quantum measurements determine the rotation angles of subsequent gates. Instead of sampling measurement outcomes during training, they evaluate all possible measurement branches and minimize a balanced loss. The authors provide theoretical justification for why feedback is required.

The authors evaluate their approach on ground-state energy estimation, phase classification, and XOR classification using different circuit architectures, depths, and qubit sizes. They compare the learned feedback rule with manually designed rules and unconditional circuits. The results show that learned feedback generally improves performance over unconditional circuits and can outperform manual feedback in some specific tasks. Overall, the paper provides an interesting theoretical and experimental justification for learning measurement- feedback quantum circuits. Note that the theoretical results mainly establish the importance of conditioning, not the necessity of learning the feedback rule.

Critical strengths
  1. The authors provide useful mathematical motivation for feedback-based dynamic quantum circuit architectures. Rather than showing only empirical improvements. This gives a clearer explanation of why feedback/conditioning can matter.

  2. I appreciate that the comparison includes multiple baseline types. The authors compare unconditional circuits, manual feedback, random rules, etc. Their additional experiment with a tuned family of 12 hand rules is particularly important because a single weak manual baseline would not be convincing. The measurement-form and noise experiments also strengthen the paper.

Critical weaknesses
  1. The theoretical propositions in Section 3.3 establish why mid-circuit measurement feedback can be necessary to overcome certain limits. They do not prove that the feedback angles (alpha_m, beta_m) must be learned.

  2. The experiments strongly support the usefulness of conditional gates in the tested circuit families, but the evidence that learned feedback outperforms simple hand- designed feedback is more task-dependent.

  3. I believe the experiment involving MNIST data classification demonstrates that measurement feedback is important for classification, but it does not demonstrate a quantum advantage over a classical classifier.

Questions for authors
  1. I noticed that in eqn. (2), the feedback rule is defined as g_alpha(m) = (alpha_m, beta_m)), but the subscript only refers to alpha. Since both alpha_m and beta_m are independently trainable. Also, here g_alpha looks like a pair of scalars for each m, but in eqn. (4), it looks like gate operations. Please consider addressing these notation inconsistencies.

  2. Balanced branch objective (eqn. (4)). I understand the reasoning the authors provided for considering L_bal, but I am not fully convinced that giving every branch equal importance is necessarily appropriate for every physical application.

  3. I noticed that for the physical energy objective, as in VQE, substituting the normalized post-measurement state cancels the branch probability p_m, and this allows the loss to be expressed directly in terms of unnormalized branch states. However, in the balanced objective, the inverse probability remains. I wonder whether this creates additional numerical conditioning issues during optimization, since the gradient contains p_m in the denominator.

Also, one practical concern I have is on actual quantum hardware, a branch with probability p_m must be observed through repeated measurements. So, if a branch occurs with probability 0.001, obtaining 100 observations of it would require about 100,000 circuit executions on average. I would therefore distinguish between differentiability in an exact classical simulator and efficient trainability on physical quantum hardware.

  1. Proposition 1: I understand that it certifies the necessity of feedback, not the necessity of learning the feedback rule. Therefore, it remains unclear when learning outperforms manual feedback or when it is necessary compared with it.

  2. I looked at the XOR classification results (Table 1) and noticed that the manual feedback rule achieves zero cross-entropy. But in Proposition 2, the authors mentioned that the remaining payload is independent of the measured bit after reset. Since the XOR label depends on that bit, I don’t understand how the manual feedback rule preserves this information for classification. Is there another mechanism that allows the model to achieve zero cross-entropy?

2 comments

We thank the reviewer for the positive assessment, and for pressing on the tuned 12-rule family, the measurement forms and the noise study, which sharpened how we state the boundary below.

The paper makes two claims and keeps them apart. C1: given a mid-circuit measurement, the value of its record is priced by an analytic floor binding every parameter-matched circuit unable to read the outcome, computable before training. C2: the rule is four angles inside the circuit, trained end-to-end by exact branch enumeration under one objective, with no sampling estimator. We do not claim that VQE needs a measurement or classical advantage on MNIST.

W1. Two claims, two kinds of evidence. The paper separates them itself (Sec 3.3, Sec 4.1). What the floor adds is a number readable before training: it binds a family, is attained exactly, and is computed from the task, so it tells a design at the floor from one that is badly optimized, which a training curve cannot. What learning adds is a rule found without per-task design, measured by the tuned 12-rule family (0.020/0.373/0.484 against 0.023/0.428/0.834 at h=0.5/1/2), by MNIST at 0.945 against 0.920, and by 17 strict wins in 45 form cells.

W2. Learning buys coverage, and coverage is measured. A hand rule is a point of the learned family, so where the prior is right a tie is the only possible outcome, which is the h=1 energy tie. The informative cases are where the prior stops covering the task: the gap opens to 1.1x at h=0.5 and 1.72x at h=2.0, phase classification goes from 0.035 to 0.008 on the same task, and XOR ring R1 and R2 go from 0.107 and 0.462 to 0.000. The learned angles sit off the hand grid, α=(0,-π/3) in 7 of 9 energy cells, which no discrete family reaches.

W3. MNIST measures the channel, not quantum advantage. The experiment is a controlled measurement of the channel, not a benchmark. The bit carries 0.175 bits, and the classical all-feature models give the ceiling the channel should reach, not a competitor. Channel on reaches 0.945, channel off 0.810, CIFAR at 0.006 bits gains nothing. That is the calibration showing that the channel takes exactly the available information, which is the mechanism at issue.

Q1. Notation. Fixed as φ=(α0,α1,β0,β1), g_φ(m)=(αm,βm), G_φ(m)=R_Y(βm)_r R_X(αm)_Q, with a notation table.

Q2. Equal branch weights. In the two reset heads this is not a modeling choice. The phase head has p_m=1/2 exactly for both classes, and in XOR the branch is the sample, 16 and 16, scored on its own samples only, so equal weight is the uniform data measure. The nine energy cells are where a real choice exists, and there equal weight keeps the rule trained on the branch where the record is informative. The weighted cost is now reported beside it.

Q3. Inverse probability in the gradient. The 1/p_m sits in the payload normalizer, not in the value of the objective, so it rescales the θ1 gradient and leaves the α direction intact. There is no sampling variance to amplify, since the sum is exact. Adam is invariant to a per-coordinate rescaling. The scan localizes it to one collapsed initialization in 200, and the manual, unconditional and no-gate modes share it, so it belongs to the task family. We add a recovery run from that initialization.

Q4. Simulator differentiability against hardware trainability. Training never samples, so no branch is estimated from repetitions. At deployment one shot realizes one branch and applies that branch's angles, so a rare branch needs no repeated observation. The 10^5 figure prices estimating a 10^-3 branch's loss to 100 observations, and the deployed cost weights that branch by p_m anyway. The rule is a four-entry table, so execution costs what the circuit costs.

Q5. When learning beats manual. The claim is coverage, not necessity, and W2 measures it on the same tasks as well as across them. Where the designer already knows the rule, learning matches it and the paper says so.

Q6. XOR, the manual rule, and Prop. 2. The reset erases m from the payload, and the conditional gate writes it back on the reset qubit itself. The head projects Q, normalizes, resets it to |0>, then applies R_X(αm) to Q, so with α=(0,π) branch 1 gets R_X(π)|0>=|1> and the branch states are orthogonal again. Prop. 2(ii) bounds processors applying the same map to an m-free payload, and this rotation is what makes the state m-dependent, so manual CE=0 is the predicted behavior of the conditional family. Where the routing fails, manual fails with it, and the learned rule does not.

Reviewer_ZmrxOpenReview ↗

Staying with my positive assessment.

I thank the authors for their clarification for many of my concerns. I reaffirm my positive assessment of a weak acceptance.

Comments on the paper

4 comments

Thank you for reading the response closely, and for saying plainly what is still missing. We appreciate that you acknowledged the clarifications on the floor's origin, the crossing settings, the objective distinction and the XOR encoding. Below are the numbers for the four items you find unresolved, followed by the scope change and the corresponding revision.

Q3. On ring R2 under the matched-restart protocol, coherent 0.136, learned 0.191, parameter-matched unconditional control 0.762. The direction is expected: by deferred measurement the coherent circuit is a strict superset, since it keeps the coherence the measurement destroys. We do not contest it. The abstract no longer calls a mid-circuit measurement a computational resource, and states the conditional claim instead. The two reset heads have no coherent counterpart, and the syndrome experiment sits there.

Q8. All 27 cells re-run under the paper's own matched-restart rule; no cell trails its best baseline by more than 0.004 gap.

Q7. The fixed-alpha hybrid is within 0.02 gap of the learned rule in eight of nine energy cells, which matches the paper's prediction that the hand prior is already near-optimal there.

Q5. In seven of the nine energy cells the deployed cost of the same trained rule is within 0.002 gap of its balanced cost. The weighted column enters the same revision, and Sec 3.3 now states the boundary, that the floors bound the balanced objective only.

The manuscript has been revised, and every passage that answers this reply is marked in blue so it can be located without a diff: the abstract (the conditional claim, and the crossing statement restricted to branch counts beyond two), Sec 3.1 (the energy task as a benchmark rather than a claim about the ground-state problem), Sec 3.3 (the scope of the floors), and a new Appendix K carrying the syndrome table above.

We are grateful for the standard you set, which has made the revision sharper. If one of the four still falls short, a pointer to which one would help us a great deal.

Reviewer_nBA3OpenReview ↗

Wrong comment box

Dear Reviewer vbDz and Authors,

In the busyness and approaching deadlines I have mistakenly responded to the wrong reviewer comment box. I am currently going to delete this assessment and put it in the correct (my own evaluation comment box) I apologize for the delay.

Kind regards,

Reviewer_nBA3OpenReview ↗

Asking for clarification to the authors

Dear Authors,

As you have noticed I (nBA3) made a mistake while commenting my response to your review. Could you kindly clarify if this response is meant for my review so that I can finally evaluate it properly?

Kind regards,

Yes. This is for you (nBA3).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.