acceptodds
Under review as a conference paper at ICLR 2027

Geometric Flow Matching for Learning Quantum Many-Body States with Classical Shadows

Abstract

Classical shadows provide an efficient representation of quantum many-body states, enabling machine learning approaches to predict properties of unseen states. The distribution of shadows can be viewed as a tomographic representation of a quantum state, from which many observables can be estimated. Learning its Hamiltonian-conditional distribution therefore provides a generative surrogate for estimating multiple properties of quantum states unseen during training. Existing approaches, however, either generate qubits autoregressively or diffuse continuous token embeddings; neither explicitly exploits the discrete and physically meaningful geometry of shadow outcomes. We ask whether incorporating this structure improves conditional shadow generation. We introduce ShadowFlow, a non-autoregressive flow-matching framework that explicitly incorporates the geometry of shadow outcomes through two complementary constructions. We evaluate ShadowFlow on one-dimensional quantum systems up to 100 qubits and two-dimensional systems up to 6x6, comparing against direct regression, autoregressive generation, diffusion, and continuous and discrete flow baselines. Across these benchmarks, ShadowFlow improves observable estimation over existing generative approaches in several settings. We also further demonstrate its extension beyond Pauli measurements to other IC-POVMs.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Flow Matching, Geometric Deep Learning, Riemannian Geometry, Quantum Many-body States, AI for Science

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_5sqCOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a non-autoregressive method to generate classical shadow samples and predict properties of many-body quantum states. The authors theoretically analyze and find that the spin error is more severe than the basis change error during property prediction, and develop two geometric approaches to design flow matching for learning classical shadows. The first one is Spherical Flow, which embeds local measurement outcomes on the Bloch sphere and uses geodesic transport. The second one is Anisotropic Dirichlet Flow, which modifies the probability path to attract samples toward a target outcome while repelling its opposite outcome. Evaluated on 1D and 2D quantum systems of up to 100 qubits, the proposed method exhibits some improvement on predicting two-site correlations and two-site R'enyi-2 entropy compared to baselines.

Critical strengths

The paper introduces a novel method to integrate flow matching and classical shadow for quantum property prediction. The theoretical analysis is interesting. It connects the geometry of measurement outcomes to their consequences for observable estimation. The experiments are extensive and thorough across multiple 1D and 2D quantum systems and different qubit numbers.

Critical weaknesses
  1. Clarification on the spin flip error and comparison with error mitigation for shadow is needed. To my understanding, the motivation for the two proposed methods is to reduce the impact of the spin-flipped error for conditional shadow generation. However, the paper does not clearly state where the error comes from. Is is from the generative model, the hardware noise or the readout of classical shadow? The spin-error analysis considers independent outcome flips, whose bias can be corrected by established robust shadow [1] or readout error mitigation methods when the noise rate is known (which can be estimated by calibration). If the modeling errors can be approximated by a calibratable channel, a comparison with appropriate postprocessing would help establish the benefit of the geometric construction.

  2. The empirical evidence for the benefit the proposed method is limited. The paper considers two properties, i.e., the two-site correlation function and the two-site entropy. The results do not establish whether the improvements extend to higher-order correlations or larger-subsystem properties. Moreover, the gains over the DirichletFM baseline are unclear, especially for large system sizes on more complex Hamiltonians like the Heisenberg.

[1] https://link.aps.org/doi/10.1103/PRXQuantum.2.030348

Questions for authors
  1. How to interpret the results in Table 2? The marking for the best and second best results are not complete.

  2. Consistency of the reported results. Please reconcile the ShadowGPT results in Tables 2 and 7. For TFIM L=10 at B=10k, Table 2 reports correlation/entropy RMSE of (0.021/0.038), whereas Table 7 reports (0.023/0.046). In addition, Section 5.2 states that ADFM performs best at (L=50), but Table 2 shows lower errors for Spherical Flow on both properties and a lower correlation error for ShadowGPT.

  3. The gains over DirichletFM are small in several Heisenberg settings. Can the authors clarify when the geometric modification is expected to help?

3 comments

Thank you for your careful review and for pointing out these points.

Your questions helped us clarify where errors enter a generated shadow record and what evidence would demonstrate a benefit beyond two-site observables. We address these points with a clearer account of the error model, a three-site evaluation, and corrected benchmark entries. Please see the passages marked in orange in our revised manuscript for the changes we made in response to the reviews.


W1. Is spin error from generation or hardware, and can robust-shadow calibration remove it?

Thanks for pointing this out. The shadow error is coming from generation not from a hardware. Since we are using machine learning to optimize generative models, they are inherently imperfect, and some bias and variance resides in their prediction of generative trajectory. Now our idea was to incorporate geometry to more faithfully follow the inrinsic geometry of shadows and relate this to a better observable estimation of quantum states. We agree that a calibrated readout error can be corrected but the spin errors we study arise at a different stage. Even with an ideal measurement apparatus, a generator can assign the wrong probability to a sign given its basis and the outcomes at nearby sites. Our idea is to use geometry to guide the model when learning those probabilities.

Chen et al. address the measurement stage in Robust Shadow Estimation (PRX Quantum 2, 030348). They prepare a known calibration state, apply randomized measurements on the device, and use the calibration data to estimate the effective noisy measurement map. Its calibrated inverse is then used to estimate observables of the unknown state. Their analysis assumes a sufficiently well-prepared calibration state and gate-independent, stationary, Markovian circuit noise.

That correction works because the device channel acts on the measurement and can be characterized separately. Our generator has no device-induced attenuation factor to calibrate in these experiments and we are only dealing with an error coming from our trained generative model, which inherently has some scale of noise due to imperfect optimization of a machine learning model. Our method tried to levarage the geometry of shadows in order to mitigate this scale of error in generation of shadows. See Figure 2 and Appendix E in our revised manuscript.


W2. Only two-site properties; unclear benefit for higher-order or larger subsystems and Heisenberg

We agree that predicting observables on a larger subsystem is a more demanding and interesting test. We therefore evaluated three-site Pauli observables and Rényi-2 entropy against exact diagonalization. The generated shadows also predict these more complex properties accurately in the tested setting. Spherical Flow improves on DirichletFM for both three-site metrics, and ADFM gives the lowest entropy error among the learned methods.

Three-site Rényi-2 entropy probes how well the generated records reproduce the purity of the larger subsystem.

On the 10-site chain, three-site entropy RMSE falls from 0.079 for DirichletFM to 0.077 for Spherical Flow and 0.074 for ADFM. On larger lattices the gains are more modest. At 6×6, ADFM improves correlation RMSE and ties DirichletFM on entropy at the reported precision. See Tables 1 and 3 and Appendix E in our revised manuscript for the main benchmarks and estimators. The additional test below uses eight adjacent triplets per Hamiltonian and a common budget of 10,000 shadows.

For each of 100 test Hamiltonians, we evaluate XXXXXX, YYYYYY, ZZZZZZ, and XYZXYZ on eight consecutive triplets Ai=(i,i+1,i+2)A_i=(i,i+1,i+2). A snapshot contributes P^=3∣A∣1[bA=P]∏i∈Aσi\widehat P=3^{|A|}\mathbf1[b_A=P]\prod_{i\in A}\sigma_i. For S2(A)=−log⁡Tr⁡(ρA2)S_2(A)=-\log\operatorname{Tr}(\rho_A^2) we estimate purity by averaging Tr⁡(ρ^A(a)ρ^A(b))\operatorname{Tr}(\hat\rho_A^{(a)}\hat\rho_A^{(b)}) over distinct snapshot pairs. The RMSEs average squared deviations from exact diagonalization over 100×8×4100\times8\times4 Pauli targets and 100×8100\times8 entropy targets, respectively, before taking the square root.

About the benefit of ADFM over DirichletFM, we noticed that specific fix or ingredients further improves our proposal from the initial version of our implementation. For example, we refined one of inference formula for ADFM (Appendix. D) for more accurate roll out of generation trajectory and also added basis alignment technique for the spherical flow's prior. We also added more rigorous validation protocol for choosing best model during training. After that, we were able to see larger gap between our method and some of the strongest existing baselines (refer to tab 1 and 3 in the revised manuscript). Now, we believe the benefit of these geometric flow models are more pronounced based on these recipes. Please refer to the following tables and the ones in our revised manuscript.

Additional rebuttal experiment — adjacent bonds and three-site subsystems.

L=10, B=10kTwo-site correlationTwo-site entropyThree-site PauliThree-site entropy
LinearFM0.0980.1480.0650.167
Antipodal-weighted LinearFM0.0910.1330.0640.145
ShadowGPT0.0730.1140.0790.099
DirichletFM0.0580.0900.0680.079
ADFM0.0560.0860.0690.074
Spherical0.0630.0970.0620.077
Measured shadows0.0160.0230.0520.025
Wang et al.0.0900.1430.0630.112

About benefit of our method over DirichletFM, please refer to Tab 1, Tab 3 in the revised manuscript. Moreover, we want to emphasize that what we call DirichletFM in our work is our own implementation on this particular shadow generation task. We belive meaningful contribution lies in finding the optimal Hamiltonian conditioning layer, testing DirichletFM framework in quantum property estimation task (original work only shows its usage in language modeling or DNA sequence generation), and showing discrete state modeling of shadows give actually gives benefit. We further push these frameworks by incorporating inherent geometry of shadows amd explore whether that gives benefit in property estimation (prediction) of unseen quantum states.


Q1. Incomplete best/second-best table markings

Thank you for catching the missing markings. We corrected them and applied the same rule throughout: bold and underlining identify the two lowest distinct rounded values within the CFM/Diffusion and CS-DFM rows, including ties. See Table 2 in our revised manuscript.


Q2. Inconsistent TFIM baseline values and L=50L=50 rankings

Thank you for pointing out the inconsistency. We reconciled the main and appendix ShadowGPT entries against the same three-run record and corrected the L=50L=50 discussion to match the per-property rankings. See Tables 2 and 11 and Section 5.2 in our revised manuscript.


Q3. When should geometric modifications improve over DirichletFM?

We observed a tendency across the tested Heisenberg systems. The smaller systems often favored Spherical Flow, while several of the larger systems favored ADFM.

We understand this pattern in terms of the models’ inductive biases. Spherical Flow gives the learned field freedom to move probability on the sphere. ADFM prescribes how concentration changes at the target and its antipode. That extra structure may help when learning a larger joint distribution. The coverage sweep at fixed size shows why we treat this as an interpretation of the observed trend.

For example, at 10,000 shadows, Spherical Flow leads on both metrics on the 4×4 lattice, and ADFM leads on the 5×5 lattice. At 6×6, ADFM retains a small correlation advantage over DirichletFM and the two tie on entropy at the reported precision. Training coverage also affects the ranking. See Tables 1 and 3 in our revised manuscript for the size sweeps and Appendix H for the Haar coupling comparison.

We also figured out that basis aligning the Haar prior for spherical flow was important recipe to achieve improvement in performance: for the spherical coupling, draw U=g/∥g∥2U=g/\|g\|_2 with g∼N(0,I3)g\sim\mathcal N(0,I_3) and permute its largest absolute coordinate onto the sampled basis axis. This uses no target sign and preserves the Haar marginal when averaged over uniformly sampled bases. Holding the Haar prior fixed on 4×4, the independent and aligned recipes give 0.07438/0.11652 and 0.06737/0.10137. Their validation-selected training durations differ, so the comparison measures the two recipes as run. See Appendix H in our revised manuscript.

Table 3 in our revised manuscript — 4×4 lattice, selected rows.

4×4Corr. 1kCorr. 10kCorr. 20kEntropy 1kEntropy 10kEntropy 20k
ShadowGPT0.1250.0880.0840.1500.1250.124
LinearFM0.1430.1100.1110.2180.1970.198
QuaDiM0.1180.0760.0710.1570.1390.138
StatisticalFM0.1130.0730.0740.1600.1360.137
DirichletFM0.1140.0760.0730.1380.1160.116
ADFM0.1130.0720.0690.1320.1020.102
Spherical Flow0.1090.0670.0680.1300.1010.102

Table 3 in our revised manuscript — 5×5 lattice, selected rows.

5×5Corr. 1kCorr. 10kCorr. 20kEntropy 1kEntropy 10kEntropy 20k
ShadowGPT0.1350.1000.0980.1740.1540.152
LinearFM0.1270.0860.0830.1800.1610.160
QuaDiM0.1240.0850.0830.1700.1490.146
StatisticalFM0.1210.0900.0870.1620.1460.144
DirichletFM0.1220.0800.0760.1370.1100.106
ADFM0.1180.0770.0730.1290.1050.103
Spherical Flow0.1170.0820.0780.1460.1210.119

Table 3 in our revised manuscript — 6×6 lattice, selected rows.

6×6Corr. 1kCorr. 10kCorr. 20kEntropy 1kEntropy 10kEntropy 20k
ShadowGPT0.1290.0920.0900.1580.1320.131
LinearFM0.1210.0820.0790.1570.1360.135
QuaDiM0.1240.0860.0820.1650.1450.142
StatisticalFM0.1250.0850.0810.1780.1560.154
DirichletFM0.1160.0790.0770.1400.1090.107
ADFM0.1150.0770.0740.1370.1090.106
Spherical Flow0.1190.0780.0750.1390.1130.110

We hope these changes address your concerns about our manuscript. Thank you for taking your interest and consideration into reviewing our work.

Reviewer_93q7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ShadowFlow, a geometry-aware flow-matching framework for Hamiltonian-conditioned generation of classical shadows of quantum many-body states. The authors argue that existing autoregressive and diffusion-based approaches do not explicitly exploit the physical geometry of quantum measurement outcomes, particularly the distinction between opposite spin outcomes and different measurement bases. To address this limitation, the paper introduces two complementary approaches: (1) Spherical Flow, which embeds Pauli measurement outcomes on the Bloch sphere and performs Riemannian flow matching along spherical geodesics, and (2) Anisotropic Dirichlet Flow (ADFM), which extends standard Dirichlet Flow by introducing target–antitarget asymmetry into the probability path. Both approaches employ Hamiltonian-conditioned Transformers to generate complete shadow sequences non-autoregressively, enabling the estimation of multiple physical observables.

The authors provide theoretical and empirical analyses of different measurement errors and evaluate their methods on 1D and 2D quantum systems, including Heisenberg and transverse-field Ising models. Comparisons with direct regression, autoregressive generation, diffusion models, and existing flow-matching approaches demonstrate improvements in observable estimation across several experimental settings. The paper also investigates training-data requirements, sampling efficiency, and extensions to non-Pauli POVMs.

Critical strengths

Physically motivated geometric modeling. The paper identifies a meaningful distinction between opposite spin outcomes and changes in measurement basis, supported by theoretical analysis and controlled corruption experiments. This provides a concrete motivation for incorporating measurement geometry into classical-shadow generation, rather than treating all discrete outcomes as unrelated categories.

Comprehensive evaluation across quantum systems and generative paradigms. The experiments cover multiple Hamiltonian families, 1D systems with up to 100 qubits, and 2D systems with up to 6 × 6 qubits. The comparisons include recent autoregressive, diffusion-based, and flow-matching approaches, making the evaluation relatively broad and relevant.

A mathematically structured extension of Dirichlet Flow. ADFM introduces an explicit target–antitarget relationship into the probability path and derives a corresponding velocity field through the continuity equation. This provides a principled formulation beyond simply adding an auxiliary geometric regularization term.

Reusable generation and extension to different measurement schemes. By learning the conditional shadow distribution, the framework supports the estimation of multiple observables without training separate observable-specific predictors. The experiments on Cube-8 and tetrahedral POVMs also provide preliminary evidence that the proposed ideas extend beyond standard Pauli-6 measurements.

Critical weaknesses

Insufficient evidence supporting the necessity and effectiveness of the proposed geometric mechanisms. Although the paper demonstrates that opposite spin errors can be more detrimental than basis errors, this observation does not directly establish that modifying the flow geometry or probability path is necessary for improving shadow generation. Simpler alternatives, such as geometry-aware classification losses or cost-sensitive objectives, are not investigated. Furthermore, the paper does not directly demonstrate that the proposed approaches reduce the specific types of generation errors identified in the theoretical motivation. This weakens the connection between the theoretical analysis, methodological design, and empirical improvements.

Limited methodological novelty and unclear relationships between the two proposed approaches. Spherical Flow largely applies existing Riemannian Flow Matching to a physically motivated embedding of Pauli measurement outcomes, while ADFM modifies the probability path of existing Dirichlet Flow. Although both approaches are reasonable, they appear to be largely independent solutions to the same problem rather than components of a unified algorithm. Their relationship, complementary roles, and individual methodological contributions need to be better justified.

Insufficient consideration of computational efficiency and comparison fairness. Despite its non-autoregressive formulation, ShadowFlow does not demonstrate a practical inference-speed advantage. The reported sampling time for flow-based models is substantially higher than that of ShadowGPT and QuaDiM. Moreover, the proposed flow models use a larger Transformer architecture than the reported ShadowGPT baseline, making it difficult to attribute all performance improvements to the geometric modeling strategy. Comparisons under matched computational and model-capacity budgets would be valuable.

Inconsistent improvements and limited component-level ablations. The performance gains over standard Dirichlet Flow are relatively small or absent in several settings, such as the 2D 6 × 6 experiments. The paper lacks a systematic analysis of the anisotropy parameter, alternative geometric representations, and the contribution of each design choice. Although three-seed statistics are reported, the significance of some marginal improvements remains unclear. Additionally, certain baseline results appear inconsistent between the main tables and supplementary material, which should be clarified.

Questions for authors

Motivation and validity of the corruption experiments. The paper compares spin errors and basis errors using the same per-token corruption probability (p). While this ensures matched corruption rates, the basis-error channel randomly resamples the measurement sign rather than using the state-dependent measurement probabilities. What motivates this particular corruption model, and how representative is it of physically realistic measurement errors? Would the conclusion remain valid under alternative or physically motivated corruption channels?

Necessity of the proposed geometric mechanisms. The theoretical analysis motivates treating opposite spin outcomes differently, but why is modifying the flow geometry or probability path preferable to simpler alternatives, such as geometry-aware classification losses or cost-sensitive learning? Could the authors provide controlled comparisons demonstrating that the proposed flow constructions offer benefits beyond incorporating the same geometric information into the training objective?

Relationship between Spherical Flow and ADFM. The two proposed approaches appear largely independent, with one modifying the continuous geometric space and the other modifying the categorical probability path. What is the intended conceptual or methodological relationship between them? Are there specific theoretical or empirical reasons for preferring one approach under certain conditions, beyond the observed differences across system sizes?

Ablation studies and computational fairness. Could the authors provide a more systematic sensitivity analysis of the anisotropy parameter (\gamma), including how performance changes relative to standard Dirichlet Flow at (\gamma=0)? Additionally, given the differences in backbone capacity and reported sampling times, how do the proposed methods compare with existing approaches under matched model sizes, inference latency, or computational budgets?

Experimental reliability and interpretation of improvements. Some improvements over the strongest flow-matching baselines are relatively small. Are these differences statistically significant across training runs? Can the authors provide evidence that the proposed methods specifically reduce antipodal-outcome errors or improve the fidelity of the generated measurement distribution? Furthermore, the ShadowGPT results for certain TFIM settings appear inconsistent between Table 2 and Appendix Table 7 (e.g., the TFIM-10 entropy RMSE at (B=10k)). Could the authors clarify which results are correct and whether the discrepancy affects the reported comparisons?

5 comments

Thank you for taking the time to examine the motivation and evaluation of our methods.

You asked whether a simpler loss could capture the benefit of the geometry and whether the gains justify the sampling cost. We focus on those questions by comparing the loss and flow constructions directly and reporting the accuracy and cost of the evaluated models. Please see the passages marked in orange in our revised manuscript for the changes we made in response to the reviews.


W1. Necessity of geometric paths versus simpler geometry-aware losses; no direct error evidence

We tried geometry aware loss in linearFM and observed this improves the shadow generation. We are calling that as antipodal-weighted loss and we believe this further supports our claim about the benefit of levaraging geometry of shadows. In our controlled comparison, the geometric flows improve both observable estimates further.

The motivation comes from both a proof and a controlled experiment. Theorem 4.2 in our revised manuscript shows that, for a weight-ww Pauli string with expectation μ\mu, independent sign flips at rate pp attenuate the expected estimate to (1−2p)wμ(1-2p)^w\mu, while basis corruption at the same rate gives (1−p)wμ(1-p)^w\mu. For a nonzero signal and sufficiently many shots, the sign-flip channel gives larger mean squared error (0<p≤1/20<p\leq 1/2). Figure 2 tests this prediction on adjacent XX, YY, and ZZ correlations: the paired 95% confidence interval for the excess RMSE from sign flips lies above zero at every nonzero rate tested. This is why the loss control specifically penalizes the opposite sign.

What we can measure directly is the accuracy of the estimated observables. A generated shot has no paired measurement with a known correct sign, so we cannot count how often the generator flipped that sign. See Figure 2 in our revised manuscript for the controlled corruption experiment and Q2 below for the loss definition.

Here is how the antipodal weighting works. For a target y=(b,σ)y=(b,\sigma), its antipode is the outcome yˉ=(b,−σ)\bar y=(b,-\sigma) on the same axis. At each site we use the endpoint probabilities from the existing LinearFM predictor and minimize L=−log⁡qθ(y∣xt,t,h)+λ qθ(yˉ∣xt,t,h),λ=1.\mathcal L=-\log q_\theta(y\mid x_t,t,h)+\lambda\,q_\theta(\bar y\mid x_t,t,h),\qquad \lambda=1. The first term rewards probability on the correct outcome. The second charges the model in proportion to the probability it assigns to the opposite sign. For a +Z+Z target, two predictions with the same q(+Z)q(+Z) have identical cross-entropy. If one assigns 0.20 to −Z-Z and the other assigns 0.05, the first incurs 0.15 more loss. This penalty changes the training signal while keeping the original path and sampler. At the same evaluation budget, it lowers correlation RMSE by 7.5% and entropy RMSE by 9.5% relative to plain LinearFM.

** antipodal weighted loss test on adjacent bonds, 100 test Hamiltonians.**

L=10, B=10kAdjacent-bond correlationAdjacent-pair entropy
LinearFM0.0980.148
Antipodal-weighted LinearFM0.0910.133
ShadowGPT0.0730.114
DirichletFM0.0580.090
ADFM0.0560.086
Spherical0.0630.097
Measured shadows0.0160.023
Wang et al.0.0900.143

W2. Limited novelty and unclear relationship between Spherical and ADFM

As the reviewer noted in the strengths, shadow outcomes have a geometry that matters for estimating quantum observables. Reversing a sign on the measured axis can affect an estimate differently from changing the axis. Our contribution is to identify this distinction, build it into conditional flow models, and show that it helps predict properties of quantum many-body systems. Spherical Flow represents opposite outcomes as antipodes on a sphere. ADFM encodes the same relationship in its probability path.

Moreover, we want to emphasize that what we call DirichletFM in our work is our own implementation on this particular shadow generation task. We belive some contribution lies in finding the Hamiltonian conditioning recipe for applying and testing DirichletFM framework in quantum property estimation task and showing discrete state modeling of shadows give benefit

The empirical finding is that a non-autoregressive model can generate all sites jointly and match or exceed the observable accuracy of the autoregressive baselines on the tested systems. The comparison with DirichletFM and the loss-only control in Q2 show why the geometric modification is useful, beyond the choice of a non-autoregressive sampler.

We evaluated the approach on chains of up to 100 sites and lattices of up to 6×6. We also tested adaptation to a new Hamiltonian family, prediction at fields withheld around a phase transition, and forecasting at times beyond the training window. These experiments establish where the learned generator generalizes and where accuracy depends on target data. See Sections 4.2.1–4.2.2 and Tables 1–3, 6, and 7 in our revised manuscript.


W4. Small/inconsistent gains, missing gamma and component ablations, statistics and table discrepancies

The updated comparison shows a clear size-dependent pattern at the reported shot budget. Spherical Flow leads on the smallest lattice, and ADFM gives the strongest overall results on the larger lattices. The size of the gain depends on the observable.

We revisited both geometric flows. For ADFM, we corrected how the evaluator computes the velocity of the target and its antipode. For Spherical Flow, we paired a Haar-uniform starting point with the measured basis. We then updated the 2D comparison.

At 10,000 shadows on the 5×5 lattice, ADFM lowers correlation and entropy RMSE from 0.080 and 0.110 for DirichletFM to 0.077 and 0.105. The gap narrows at 6×6, where ADFM improves correlation and ties entropy to the reported precision. We interpret the two RMSEs together at the same checkpoint and shot budget: the lower correlation error and equal rounded entropy error at 6×6 indicate a gain in correlation accuracy while preserving entropy accuracy to the reported precision. These lattice comparisons are point estimates. In the separate three-seed time forecast, the paired tests do not establish a significant gain. See Tables 3 and 7 and Appendix H in our revised manuscript. Q4 below gives the sensitivity to the antipodal weight, and R3–Q2 explains the TFIM baseline correction.

The clearest revision is at 5×5. In the earlier table ADFM had slightly higher entropy RMSE than DirichletFM (0.112 versus 0.111). In the updated table it is lower (0.105 versus 0.110), alongside lower correlation RMSE. In ADFM the corrected evaluator uses the appropriate Beta marginal for each of the target and antipode velocity coefficients. The Haar coupling change applies to Spherical Flow. These changes do not explain revisions to other baselines, and the gains differ by lattice and observable.


Q1. What does the corruption experiment tell us about errors in shadow generation?

Our task is to learn the distribution of shadow records conditioned on the Hamiltonian. The corruption experiment probes the consequences of errors in those records for observable estimation.

Starting from clean records, we perturb either the sign within a fixed Pauli basis or the basis–outcome token at the same per-token rate. When we change the basis, we choose uniformly between the other two Pauli axes and draw a fresh sign uniformly so the perturbation does not carry over the original sign or introduce a state-dependent sign bias. This gives us a controlled comparison of how sensitive the estimators are to these two changes.

Physical circuit and readout noise are outside the scope of this experiment. We deliberately perturb records to measure how two kinds of errors affect observable estimates. That comparison does not depend on whether either perturbation occurs in a quantum device. A physically realistic channel would address device noise and is unnecessary for this diagnostic. We assess the effect of geometry on the generator separately using observable RMSE and the cost-sensitive LinearFM control in Q2.


Q2. Controlled comparison with cost-sensitive or geometry-aware classification losses?

Yes. A simpler way to use the geometry is to teach LinearFM that some wrong outcomes are more costly for the observable we want to estimate. We tested that idea by changing only the endpoint loss. The penalty improves both observable errors, and both geometric flows improve further in the same evaluation.

For an observed outcome y=(b,σ)y=(b,\sigma), let yˉ=(b,−σ)\bar y=(b,-\sigma) be the opposite sign on the same axis. If qθq_\theta is the model’s endpoint distribution at an intermediate state xtx_t, our control uses L=−log⁡qθ(y∣xt,t,h)+λqθ(yˉ∣xt,t,h),λ=1.\mathcal L=-\log q_\theta(y\mid x_t,t,h)+\lambda q_\theta(\bar y\mid x_t,t,h),\qquad \lambda=1. The first term rewards the correct outcome. The second makes the model pay for the probability it puts specifically on its antipode. Suppose the target is +Z+Z. Two predictions assigning the same probability to +Z+Z receive the same cross-entropy. If one places 0.20 on −Z-Z and the other places 0.05 there, the former incurs 0.15 more loss. This distinction has a physical motivation. A −Z-Z record reverses that shot’s contribution to a ZZ estimate, whereas a measurement on another axis contributes zero to that estimate.

Both LinearFM runs start from fresh weights and use 200 epochs, the same training path, and the same sampler. On 100 held-out 10-site Heisenberg Hamiltonians, the penalty lowers adjacent-pair correlation RMSE by 7.5% and entropy RMSE by 9.5%. The table reports both test errors from one checkpoint per model at 10,000 shadows per Hamiltonian (one training seed). The comparison between the two LinearFM rows isolates the effect of the loss. The geometric-flow rows show the accuracy of the complete methods under the same evaluation.

L=10, B=10kAdjacent-bond correlationAdjacent-pair entropy
LinearFM0.0980.148
Antipodal-weighted LinearFM0.0910.133
ShadowGPT0.0730.114
DirichletFM0.0580.090
ADFM0.0560.086
Spherical0.0630.097
Measured shadows0.0160.023
Wang et al.0.0900.143

Q3. How are the methods related and when should one be preferred?

Both methods give special treatment to the opposite sign because it reverses what a measurement tells us about that observable. Spherical Flow encodes this through distance, while ADFM builds it into the probability path. We choose between them on validation at the intended size and training budget.

Suppose a shot records +Z+Z. Replacing it with −Z-Z changes a positive contribution to the ZZ estimate into a negative one. Replacing it with +X+X gives zero contribution to that estimate. This difference is why we single out the antipode.

Spherical Flow expresses the idea through distance. Starting at +Z+Z, the opposite outcome −Z-Z is half a turn away, while each outcome on another axis is only a quarter turn away. ADFM expresses the same idea directly in the probability path. As it strengthens +Z+Z, it also weakens −Z-Z. The other four outcomes receive no such extra suppression.

In our benchmarks, Spherical Flow tends to do better on smaller systems and ADFM on larger ones. Training coverage can change the ranking even at fixed size, so we use validation at the intended data budget to choose between them. See Figure 2 and Tables 1 and 3 in our revised manuscript.

The two data sweeps help explain why we use validation to choose the method. The cleanest comparison varies the number of training Hamiltonians at fixed chain length, shots per Hamiltonian, and optimizer update cap. At 50 Hamiltonians, Spherical Flow leads on both metrics. At 400, ADFM leads on both. The middle rows show how the ranking changes along the way.

Training HamiltoniansDirichletFMADFMSpherical
500.069 / 0.1080.069 / 0.1100.063 / 0.093
1000.057 / 0.0870.055 / 0.0840.057 / 0.086
2000.045 / 0.0660.046 / 0.0670.046 / 0.065
4000.046 / 0.0680.042/ 0.0590.046 / 0.066

The separate three-seed sweep varies measurements per Hamiltonian on the 30-site chain. At 1,000 shots Spherical Flow has the lowest correlation/entropy RMSE of the three methods (0.085/0.135). At 2,000 shots DirichletFM has lower errors than ADFM on both metrics. At 4,000 shots ADFM improves on DirichletFM (0.07863/0.12181 versus 0.08169/0.12788), while Spherical Flow gives 0.07718/0.12196. Reading the pair together, Spherical Flow leads on correlation at this budget and ADFM narrowly leads on entropy. See Figure 4 in our revised manuscript for the full measurement sweep and across-run variation.


Q4. Gamma sensitivity including gamma0 and matched size/latency budgets?

A small positive anisotropy gave the best validation accuracy in our sweep. Here γ\gamma controls ADFM’s probability path. The loss-only LinearFM experiment in Q2 instead changes the training objective through λ\lambda.

For target ii, the path sets αi(s)=1+s\alpha_i(s)=1+s, αiˉ(s)=1−γs\alpha_{\bar i}(s)=1-\gamma s, and the other concentrations to one. In our implementation, DirichletFM is exactly the γ=0\gamma=0 setting of ADFM. Setting γ=0\gamma=0 keeps the antipodal concentration at one throughout the path, so only the target concentration increases. We use this setting to run the DirichletFM baseline. Table 1 in our revised manuscript therefore gives the zero-anisotropy test comparison. The table below reports validation errors for the four positive values we swept on the 10-site chain. Of these, γ=0.01\gamma=0.01 gave the lowest error on both observables.

ADFM γ (L10 validation)Correlation RMSERényi-2 RMSE
0.010.0540.076
0.030.0570.079
0.10.0580.086
0.20.0550.079

For computational fairness, we timed comparable-capacity samplers on the same device and batch. ADFM and DirichletFM use the same number of network calls and nearly identical matrix FLOPs. Accuracy is reported at a common generated-shot budget. See Tables 1 and 9 in our revised manuscript for the test comparison and measured costs.

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

This paper introduces a framework for generative modeling of classical shadow measurement results for predicting properties of ground states of unseen Hamiltonian. The main contribution is incorporating measurement-outcome geometry to flow matching. It particularly investigates two constructions: spherical flow on the Bloch sphere and anisotropic Dirichlet flow. The authors evaluate correlation and Renyi-2 entropy estimation on 1D systems unto 100 qubits and 2D lattices up to 6x6, comparing against direct regression, autoregressive, diffusion, and other flow-matching baselines. They report improvements in several settings and additionally investigate training-measurement budgets and extensions to non-Pauli informationally complete measurements.

Critical strengths
  1. The explicit incorporation of the geometry of shadow outcomes into flow matching is an interesting methodological contribution.
  2. Improvements over standard flow matching baselines in several settings support the usefulness of this physically motivated inductive bias.
Critical weaknesses
  1. Wang et al. proposed learning Hamiltonian-conditioned measurement distributions and using generated classical shadows to predict correlations and Renyi-2 entropies of unseen ground states (arXiv:2211.16943). That work also includes experimental Rydberg atom data and extrapolation tests, but is neither cited nor directly compared with here. The authors should clarify the advance over this work and provide a comparison under matched data and computational budgets to establish the significance of the proposed geometric flow constructions.
  2. Both Spherical Flow and Anisotropic Dirichlet Flow incorporate the geometry of local single-qubit measurement outcomes, while correlations across sites must still be learned by the Transformer. The experiments do not establish whether the benefit of the proposed geometry persists as the correlation and entanglement structure becomes more complex. The 1D benchmarks are classically tractable, and the 2D experiments reach only 6 × 6 lattices. Thus, it does not establish scalability to more challenging many-body states.
Questions for authors
  1. Have you investigated generalization to a Hamiltonian family not represented in training? If so, can the trained model be applied directly, or does it require fine-tuning with measurements from the new family?
  2. Can ShadowFlow generalize to a phase not represented in the training data? How accurately does it predict observables near a phase transition when no training data from that region are provided?
  3. How do the required numbers of training Hamiltonians and measurements per Hamiltonian scale with system size at fixed prediction accuracy?
  4. The sample complexity bound in Appendix A appears to omit dependence on the shadow norm of the observable.
3 comments

Thank you for taking the time to review our work and for drawing our attention to Wang et al.

Your review raises two main questions. The first is how ShadowFlow compares with the conditional generator of Wang et al. The second is how far its performance carries beyond the systems and observables used in training. We address both through a direct baseline comparison and additional tests of transfer, extrapolation, and training data requirements. Please see the passages marked in orange in our revised manuscript for the changes we made in response to the reviews.


W1. Missing Wang et al. citation and matched-data/computation comparison, including extrapolation

We agree that a direct comparison with Wang et al. is important. On the Heisenberg benchmarks, both geometric flows yield lower reported correlation and entropy errors than the official Wang baseline at comparable capacity and shot budgets.

We trained their model implemented with causal attention based Transformer, and joint Pauli-6 sampler on our random 2D Heisenberg data. This was based on their official codebase. Wang et al. and the main comparing flow models have about 31M parameters. We used 1,000 training shadows per Hamiltonian and compared at 10,000 generated shadows per test Hamiltonian. At 4×4, Wang et al. gives correlation and Rényi-2 entropy RMSEs of 0.113 and 0.177, compared with 0.072 and 0.102 for ADFM and 0.067 and 0.101 for Spherical Flow.

Our leading hypothesis is that ShadowGPT benefits from bidirectional attention within the observed prefix. Both models generate autoregressively, and both can use all previously generated outcomes to predict the next one. The difference is how those observed outcomes are represented. For example, suppose two neighboring sites have already been measured. In ShadowGPT, the representation of the first site can be updated in light of the second site's outcome before predicting a third. In Wang et al., the later site can attend to the earlier one, but the causal mask prevents the earlier site from updating its representation in return. Since the order in which we list the sites is a modeling choice, allowing this exchange in both directions may make spatial correlations easier to learn.

The reported 4×4 comparison is consistent with this hypothesis. At 10,000 shadows, ShadowGPT gives correlation and Rényi-2 entropy RMSEs of 0.088 and 0.125, compared with 0.113 and 0.177 for Wang et al. The comparison also includes differences in conditioning and decoding. Wang conditions on a graph embedding and predicts basis and sign jointly. ShadowGPT uses sitewise couplings, draws the basis from its known distribution, and predicts the sign.

While Wang et al. performs some experiments on real quantum measurement data, we want to emphasize that our paper focuses on developing new ML framework for this particular task, focusing on geometric property of shadows providing thorough comparison with various baselines (autorgressive, non-AR). We also want to note that our framework can be easily extended to these real quantum measurement data. Since Wang et al. also performs experiments on random 2D Heisenberg, we provide comparisons under 4x4 2D Heisenberg and some other 1D systems. See Section 3 and Tables 1 and 3 in our revised manuscript for the comparison.

Table 1 in our revised manuscript — 10-site Heisenberg chain, selected rows.

Heisenberg chain, L=10, B=10kCorrelation RMSERényi-2 RMSE
Direct shadows0.0150.020
RBFK0.0870.151
NTK0.0850.142
Lasso0.1030.168
Ridge0.1030.168
ShadowGPT0.0580.089
Wang et al.0.0900.143
LinearFM0.0940.141
QuaDiM0.0830.114
StatisticalFM0.0960.143
DirichletFM0.0660.101
ADFM0.0550.083
Spherical Flow0.0520.078

Table 3 in our revised manuscript — 4×4 Heisenberg lattice, selected rows.

4×4 test, B=10kCorrelation RMSERényi-2 RMSE
Wang et al.0.1130.177
LinearFM0.1100.197
ShadowGPT0.0880.125
QuaDiM0.0760.139
StatisticalFM0.0730.136
DirichletFM0.0760.116
ADFM0.0720.102
Spherical0.0670.101

Table 1 in our revised manuscript — 30-site Heisenberg chain, selected rows.

L=30 test, B=10kCorrelation RMSERényi-2 RMSE
Wang et al.0.1660.246
ShadowGPT0.1000.161
DirichletFM0.0950.150
ADFM0.0950.150
Spherical Flow0.0850.135

Table 3 in our revised manuscript — 5×5 Heisenberg lattice, selected rows.

5×5 test, B=10kCorrelation RMSERényi-2 RMSE
Wang et al.0.1530.227
ShadowGPT0.1000.154
DirichletFM0.0800.110
ADFM0.0770.105
Spherical Flow0.0820.121

W2. Local geometry does not establish scalability to complex correlation/entanglement structures

We agree that the geometry of individual measurement outcomes cannot by itself capture correlations across sites. The Transformer learns those dependencies from full measurement records. Our claim is that the geometric modification helps the model to learn this correlation better and induces more accurate generation by incorporating the geometry.

To test whether the local geometry still helps when the prediction depends on a larger part of the joint distribution, we evaluated Pauli observables and Rényi-2 entropy on eight consecutive triplets of a 10-site Heisenberg chain. We used 10,000 shadows per Hamiltonian and compared the resulting estimates with exact diagonalization across 100 held-out Hamiltonians.

Three-site observable RMSE on the 10-site Heisenberg chain (10,000 shadows per Hamiltonian).

MethodThree-site Pauli RMSEThree-site Rényi-2 entropy RMSE
LinearFM0.0650.167
Antipodal-weighted LinearFM0.0640.145
ShadowGPT0.0790.099
DirichletFM0.0680.079
ADFM0.0690.074
Spherical Flow0.0620.077
Measured shadows0.0520.025
Wang et al.0.0630.112

Relative to DirichletFM, Spherical Flow lowers Pauli RMSE from 0.068 to 0.062 and Rényi-2 entropy RMSE from 0.079 to 0.077. ADFM gives the lowest entropy RMSE among the learned methods at 0.074, with a Pauli RMSE close to that of DirichletFM. The entropy result is particularly relevant here because it tests how well the generated records recover the purity of a three-site reduced state. Together, these results show that the benefit of measurement geometry extends beyond two-site prediction in this benchmark. The ground-state benchmarks in Tables 1 and 3 in our revised manuscript separately test larger systems, up to 100 sites in 1D and 6×6 in 2D. The three-site experiment extends the observable support at fixed system size. See Appendix E in our revised manuscript and R3–W2 for the estimators and evaluation details.


Q1. Generalization to Unseen Hamiltonian family: few-shot transfer learning vs. from scratch training

This is an interesting question to answer. We compared whether pre-training on different Hamiltonian class would be helpful in transfer learning (or few-shot training) on different target Hamiltonian class. We tested transfer by taking a checkpoint trained on random Heisenberg chains and fine-tuning its weights on TFIM measurements. Fine-tuning a Heisenberg checkpoint on a small set of TFIM measurements improved both metrics for both geometric flows relative to training from scratch. This is evidence for transfer with target-family measurements.

We compared the two initializations under the same conditions. Both chains have 10 sites. The fine-tuned and randomly initialized versions of each architecture use the same 20 training Hamiltonians and records, a fresh optimizer, and 200 epochs. The comparison uses a fixed training budget and validation-based checkpoint selection. In this single seed, ADFM gives 0.094/0.216 from random weights and 0.084/0.181 after transfer. The comparison tests initialization at that budget and does not isolate a stopping-time effect. We select checkpoints on 20 separate TFIM Hamiltonians and evaluate adjacent two-site observables on 100 held-out Hamiltonians with 1,000 generated shadows each.

Both geometric flows improve on their own scratch runs in correlation and entropy. For Spherical Flow the two RMSEs fall from 0.073 and 0.164 to 0.065 and 0.156. DirichletFM worsens on both metrics. The table reports one training seed per setting.

Additional rebuttal experiment — few-shot transfer to TFIM.

TFIM target N_H=20Random init. corr.Random init. entropyTransfer corr.Transfer entropy
ShadowGPT0.0900.2630.0830.239
LinearFM0.0770.2530.0860.227
DirichletFM0.0710.1780.0910.234
Ours (AD)0.0940.2160.0840.181
Ours (Spherical)0.0730.1640.0650.156

Q2. Generalization to unseen phase and transition region excluded from training?

This is another interesting questions to ask. We tried some experiments and observed that the model can interpolate across a gap around the transition when training includes fields on both sides. Predicting a phase that was completely absent from training was considerably harder.

On the 10-site TFIM chain, ADFM achieved correlation and entropy RMSEs of 0.048 and 0.046 in the withheld transition interval. When we trained and validated only below the transition and tested on the other side, the errors rose to 0.138 and 0.263. Neither test used fine-tuning. See Table 6 in our revised manuscript for the transition-gap comparison.

We used the open chain H(h)=−(1−h)∑i=1L−1ZiZi+1−h∑i=1LXiH(h)=-(1-h)\sum_{i=1}^{L-1} Z_i Z_{i+1}-h\sum_{i=1}^L X_i at L=10L=10, with exact diagonalization in the even global-XX sector. The thermodynamic critical point is h=1/2h=1/2. The 200 training and 40 validation fields span both sides but exclude [0.45,0.55][0.45,0.55]. We used 1,000 training shots per Hamiltonian and tested on 40 fields in the gap at B=10,000B=10{,}000. The separate one-sided result is for ADFM alone, trained and validated below h=0.5h=0.5 and tested at h≈0.515h\approx0.515–0.6310.631.

Table 6 in our revised manuscript — withheld TFIM transition gap.

TFIM L10 interpolationCorrelation / Rényi-2 RMSE
LinearFM0.049 / 0.176
DirichletFM0.052 / 0.052
ADFM0.048 / 0.046
Spherical0.065 / 0.074

Q3. How do M and N_H scale with system size at fixed accuracy?

Our sweeps show how accuracy responds to additional training data at each tested size. They do not yield a fixed-accuracy scaling law across sizes. On the shorter chain we varied Hamiltonian coverage, and on the longer chain we varied the number of measurements at each Hamiltonian.

For example, increasing coverage from 50 to 400 Hamiltonians on the 10-site chain lowered ADFM’s correlation and entropy RMSEs from 0.069 and 0.108 to 0.042 and 0.059 . See Figure 4 in our revised manuscript for the measurement sweep and the table below for the coverage sweep.

The coverage sweep uses one seed, 1,000 training shots per Hamiltonian, a cap of 20,000 optimizer updates, and 10,000 test shadows. At 50 Hamiltonians Spherical Flow has the lowest pair of errors (0.06296, 0.09350). At 400 ADFM leads (0.04175, 0.05904). The separate 30-site measurement sweep uses three seeds. DirichletFM leads ADFM on both metrics at 2,000 training shots, while at 4,000 shots ADFM gives 0.07863 and 0.12181 versus 0.08169 and 0.12788 for DirichletFM. The ordering depends on the budget.

Additional rebuttal experiment — training Hamiltonian coverage at L=10L=10.

Training HamiltoniansDirichletFMADFMSpherical
500.069 / 0.1080.069 / 0.1100.063 / 0.093
1000.057 / 0.0870.055 / 0.0840.057 / 0.086
2000.045 / 0.0660.046 / 0.0670.046 / 0.065
4000.046 / 0.0680.042 / 0.0590.046 / 0.066

Q4. Appendix A sample-complexity bound omits the observable shadow norm

You are right that the bound must depend on the observables being estimated. We corrected it to include the largest squared shadow norm of the centered observables. Thank you for pointing this out.

This gives the simultaneous error guarantee for independent measured shadows. For generated shadows, approximation of the measurement distribution can introduce additional bias. See Appendix A in our revised manuscript for the corrected bound and its assumptions.

For KK observables on nn qubits, median-of-means estimation gives simultaneous additive error ϵ\epsilon with failure probability at most δ\delta using M=O(log⁡(K/δ)ϵ2max⁡j≤K∥Oj−Tr⁡Oj2nI∥shadow2)M=O\left(\frac{\log(K/\delta)}{\epsilon^2}\max_{j\le K}\left\|O_j-\frac{\operatorname{Tr}O_j}{2^n}I\right\|_{\rm shadow}^{2}\right) independent measured shadows.

The comparison with Wang et al. places our results alongside an established baseline. The transfer and held-out-field experiments clarify where target measurements help, while the three-site and time-forecast tests extend the evaluation beyond the original two-site ground-state setting.


We hope these additions, together with the corrected shadow-norm bound, address your concerns and clarify the scope of our generalization claims. We hope these changes address your concerns about our manuscript. Thank you for taking your interest and consideration into reviewing our work.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of the rebuttal

Dear AC and reviewers,

Thank you for taking your time and consideration into review our manuscript. Here, we provide a summary of the rebuttal for easier consolidation. Please see the passages marked in orange in our revised manuscript for the changes we made in response to the reviews.

What the reviewers found valuable

  • KFvK, 93q7, and 5sqC recognized the sound connection between measurement geometry and observable estimation.

  • Their comments highlight the physical motivation for distinguishing opposite spin outcomes from changes of measurement basis.

  • 93q7 and 5sqC highlighted the thorough benchmark coverage, and 93q7 noted that one set of generated shadows can estimate multiple observables and that the approach extends beyond Pauli-6.

What the reviewers asked us to address

KFvK (R1)

  • Prior work. We trained the official Wang baseline at comparable capacity and shot budgets. Both geometric flows lowered correlation and entropy RMSE on the 10-site chain and 4×4 lattice (Tables 1 and 3).

  • Generalization to unseen phase. Both flows benefited from fine-tuning on 20 TFIM Hamiltonians. Predicting withheld fields around the transition was easier than predicting a phase absent from training (R1–Q1/Q2 and Table 6).

  • Scalability and Classical tractability. While Wang et al. performs some experiments on real quantum measurement data, we want to emphasize that our paper focuses on developing new ML framework for this particular task, focusing on geometric property of shadows providing thorough comparison with various baselines (autorgressive, non-AR). We also want to note that our framework can be easily extended to these real quantum measurement data. But in this paper, we focus on method development and do not show experiments with real quantum data which would be a path for classically intractable regime. Also, we tested three-site observables on a 10-site chain and time extrapolation on a 30-site chain. Separate sweeps measured sensitivity to Hamiltonian coverage and shots per Hamiltonian at fixed sizes (R1–Q3, Figure 4, and Table 7).

  • Sample complexity. We restored the observable-dependent shadow norm in the bound for independent measured shadows (Appendix A).

93q7 (R2)

  • Why geometry? We clarified how the two flows encode the antipode and tested a simpler loss-only alternative. We want to emphasize that what we call DirichletFM in our work is our own implementation on this particular shadow generation task. We belive some contribution lies in finding the Hamiltonian conditioning recipe for applying and testing DirichletFM framework in quantum property estimation task and showing discrete state modeling of shadows give benefit (R2–W1/Q2).

  • We explained why sign resampling defines a controlled perturbation. Fresh generated shots have no paired reference sign, so we assess their distribution through observable RMSE (Figure 2 and R2–Q5).

  • When are the gains reliable? The anisotropy sweep and coupling comparison probe the design choices. Spherical Flow leads at 4×4 and ADFM at 5×5 and 6x6. (Table 3, Appendix H, and R2–Q4).

  • Sampling cost. We timed complete samplers at comparable capacity on one A100. ADFM and DirichletFM use 20 calls and nearly equal matrix FLOPs, taking 0.929 and 0.824 ms per shadow, respectively (Table 9). Since our method is non-autorgressive, the benefit over autorgressive models gets pronounced for larger systems.

5sqC (R3)

  • Hardware versus generator error. We explained which measurement errors robust-shadow calibration can correct and why our benchmark has different scope compared with the prior work's. Our work is about decreasing error in generative model, not in the hardware (R3–W1).

  • Higher-order observables. We extend metrics to 3-sites observables. Spherical Flow improved both Pauli and entropy RMSE over DirichletFM. ADFM gave the lowest three-site entropy RMSE among learned methods (R3–W2). The size sweeps address when each method helps (R3–Q3).

  • Reporting consistency. In response to both 5sqC and 93q7, we corrected the rank markings, reconciled the TFIM ShadowGPT entries against the same three runs, and corrected the 50-site discussion (Tables 2 and 11).

What the rebuttal establishes

  • We want to emphasize that our contribution is to identify an observable-relevant distinction between opposite signs and changes of measurement axis, and to use that geometry in non-autoregressive flows for quantum many-body prediction. Across benchmarks reaching 100 sites in 1D and 6×6 in 2D, joint generation can match or exceed the autoregressive baselines on the observables we evaluate. On 4×4, for example, Spherical Flow lowers both errors relative to ShadowGPT as well as DirichletFM. Our transfer, withheld-field, time-forecast experiments, and evaluating on 3-sites observables also test how the learned distributions generalize beyond training conditions in harder settings or in harder evaluation settings.

  • We made several improvement in our method's implementation and now 2D comparison favors Spherical Flow on 4×4 and ADFM on 5×5 and 6x6. We report inference costs separately and clarify the scope of the corruption diagnostic. We also reconciled the TFIM baseline values and corrected the observable dependence of the shadow-norm bound. See Tables 1–3, 6, 9, and 11 and Appendix A in our revised manuscript.

  • A cost-sensitive (antipodal-weighted) loss improves LinearFM on both observables, and both geometric flows improve further in the adjacent-bond control. This supports our claim: geometric modification gives improvement in generative performance for predicting shadows of unseen quantum states.

  • Transfer initialization (first pre-trained on Heisenberg) helps all the shadow generative models, including both geometric flows, adapt to TFIM (unseen target Hamiltonian class). We also provide time extrapolation and unseen phase generalization experiment.


We hope the new baseline and loss comparisons, the transfer and three-site experiments, and the improved benchmark table with refined implementation help you assess the contribution of this work. We hope our responses resolve reviewers' concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.