acceptodds
Under review as a conference paper at ICLR 2027

QuMUSE: Generative Pretraining from Local Measurements for Characterizing Quantum Many-Body Ground States

Abstract

Characterizing quantum many-body ground states is fundamental to understanding quantum matter and advancing quantum technologies. However, most existing learning methods remain specialized to a particular Hamiltonian family, system size, or tasks, limiting knowledge reuse across physical systems. We introduce QuMUSE, a local measurement-native generative foundation model that learns unified, reusable representations of many-body ground states from local observables, targeting Hamiltonian families of variable scales and lattice geometries. QuMUSE consists of a variational autoencoder that compresses variable-dimensional measurement data into a fixed-size, physical system-aware latent state representation that decodes local observables, and a conditional latent flow model that generates state representations from physical specifications alone. Therefore, one unified representation supports both characterization from measurements and prediction without target-state measurements. Jointly pretrained on five 1D and 2D Hamiltonian families with up to 100 qubits, QuMUSE achieves the best overall observable-reconstruction and conditional-generation performance among baselines. Despite training only on local observables, the representations support predicting nonlinear and global properties, data-efficient adaptation to unseen Hamiltonian families, and unsupervised phase boundary witnessing. These results demonstrate the potential of measurement-native generative pretraining to reuse physical information across quantum systems and learning tasks.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Quantum state learning, Quantum property estimation, Variational autoencoder, Latent diffusion model

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

The paper introduces QuMUSE, a generative pretraining framework for learning shared representations of quantum many-body ground states from local observables. A variational autoencoder compresses variable-sized measurement data into a fixed-size latent representation, while a conditional flow model generates latent representations from Hamiltonian and lattice specifications. The framework is pretrained on five Hamiltonian families with up to 100 qubits and evaluated on observable reconstruction, conditional generation, quantum property prediction, phase-boundary identification, and adaptation to unseen Hamiltonian families.

Critical strengths
  1. The paper addresses the important problem of learning reusable representations of quantum many-body states from local observables. The unified architecture supports both measurement-based reconstruction and conditional generation across different Hamiltonian families, system sizes, and lattice geometries, providing a meaningful connection to generative representation learning.
  2. The authors evaluate five Hamiltonian families with up to 100 qubits and demonstrate improvements in observable reconstruction and conditional generation over the evaluated baselines. The additional experiments on few-shot adaptation, nonlinear and nonlocal property prediction, phase-boundary identification, and unseen system sizes provide evidence that the learned representations capture physically relevant information.
Critical weaknesses
  1. The proposed framework combines established techniques, including query-based attention, variational autoencoders, conditional flow matching, and representation alignment. Although their integration for quantum many-body systems is useful, the methodological contribution beyond adapting these existing techniques appears relatively limited. In particular, it remains unclear whether the conditional flow model offers substantial advantages over simpler supervised models that directly predict observables from Hamiltonian parameters.
  2. Although a single model is pretrained across five Hamiltonian families, zero-shot performance deteriorates substantially on the three unseen families. Accurate predictions require fine-tuning using six target-system ground states and 2,500 optimization steps. Furthermore, the pretrained model does not consistently outperform training from scratch on every conditional-generation benchmark. These results demonstrate the benefits of pretraining for adaptation but provide weaker evidence for the broadly transferable representations suggested by the foundation-model framing.
  3. The training data are generated using classical DMRG simulations with a maximum MPS bond dimension of 64, and the main experiments use local expectation values computed from these approximate states. While the authors state that the DMRG calculations are sufficiently converged, quantitative convergence evidence for local observables and nonlocal properties is limited, particularly for larger 2D and frustrated systems. Although finite-shot measurement noise is investigated, the framework is not evaluated using experimental quantum measurement data. It remains unclear whether representations pretrained on classically tractable quantum states can generalize to more strongly entangled states that are difficult to simulate using tensor network methods.
Questions for authors
  1. What is the main methodological advance beyond combining existing VAE, query-based representation learning, and conditional flow-matching techniques? Could the authors compare against a simpler model that directly predicts local observables from Hamiltonian parameters and lattice specifications, without latent generative modeling?
  2. The zero-shot errors on unseen Hamiltonian families are much higher than those after fine-tuning. What limits the model's generalization to genuinely unseen Hamiltonians? In particular, how are previously unseen Hamiltonian-family identifiers and their coefficients represented by the pretrained physical-condition encoder?
  3. How sensitive are the local observables and nonlocal properties to increasing the MPS bond dimension beyond 64?
  4. How does the number of ground states and local measurement observables required for pretraining scale with system size, the diversity of Hamiltonian families, and the desired prediction accuracy? Can the authors quantify the computational or quantum measurement resources required to obtain these training data?
  5. The learned representations support linear probing of entanglement entropy and other nonlocal properties despite being trained only on local observables. To what extent does this capability depend on correlations specific to the training Hamiltonian families? How would the framework distinguish states with similar local observables but substantially different global properties?
4 comments

Weakness 1 and Question 1. What is the contribution of QuMUSE, and how does it compare with direct prediction?

For the discussion on the novelty, please refer to the reply Weakness 1: What is the methodological contribution beyond combining existing components? to reviewer oS7N.

For the comparison with direct prediction, please refer to the reply Weakness 2: Stronger conditional-prediction baselines. to reviewer oS7N.

Weakness 2 and Question 2. How are unseen Hamiltonians represented, and what limits the zero-shot transfer?

In our implementation, we use a simple design. The condition encoder receives a sixteen-value vector: six family indicators, three coefficient slots, three masks and four lattice descriptors. The vector is normalized and passed through an MLP. This encoding is enough for representing the specified families and coefficients studied in the manuscript, but cannot express an arbitrary Hamiltonian term list. Extending this to arbitrary Hamiltonian is left for future work, which could be handled by a graph neural network.

In the adaptation experiments, triangular Ising reuses the 2D-Ising indicator with coefficients (1,gg,0), long-range Ising reuses the 1D-Ising indicator with (1,α\alpha,0), and Kitaev reuses XXZ with (1,1,JzJ_z). The same numerical slots therefore acquire different physical meanings in the target systems. This gives a plausible reason why fine-tuning is needed, together with the change in interaction and query geometry.

In addition, in the reply 2. Does pretraining help on unseen families? to reviewer 7G3C, we have shown that the clearest benefit is in reconstruction after adaptation. Since the ground states of different Hamiltonian families differ substantially, as illustrated in the visulization in Appendix E of the manuscript, building a model that has zero-shot transferability to any other system would require substantial amount of training data. We leave this for future work, because the goal of our paper is a first step towards a reusable representation for quantum states, which has never been discussed in previous quantum state characterization literatures.

Weakness 3. Can representations pretrained on DMRG data remain useful for more strongly entangled states, experimental quantum measurement data, and classically intractable states?

For reference accuracy, please see our reply 1. Are the 10×10 J1–J2 reference states converged? to reviewer 7G3C, which includes exact-diagonalization checks, larger-bond calculations and reevaluation of the frozen model against the refined references. For realistic measurement noise, please see Question 2. Does the reconstruction benefit survive device-specific measurement noise? in our reply to reviewer oS7N. Under the IBM Torino and Brisbane hardware with 1024 shots per observable, QuMUSE reduces raw measurement RMSE by 56.1–56.5% without retraining.

The additional question is whether using DMRG for pretraining imposes its entanglement restriction on QuMUSE. The encoder takes observable values and their measurement supports as inputs, rather than MPS tensors. Neither its inputs nor its learned representation carries an MPS bond-dimension constraint. Consequently, observables measured from a more strongly entangled state can be supplied through the same interface. Whether the pretrained model predicts them accurately is an empirical question about transfer to that state distribution.

Our higher-bond study provides a first test of this issue. For the 10×10 J1–J2 system at J2/J1=0.4J_2/J_1=0.4, increasing the reference bond dimension from 64 to 384 raises the central-cut entropy from 1.336596 to 2.221383, while the unchanged model's generation RMSE improves from 0.095016 to 0.074503. Thus, predictions learned from the lower-bond data can remain useful against a substantially more entangled reference.

A complementary result is that pretraining improves reconstruction after adaptation in all nine target–seed comparisons across three unseen families, as reported in 2. Does pretraining help on unseen families? in our reply to reviewer 7G3C. Together, these experiments support using classically generated data to learn a reusable starting representation, with target measurements providing a route to adaptation.

Question 3. How do larger bond dimensions affect local and nonlocal quantities?

For the bond-dimension study, please refer to our reply 1. Are the 10×10 J1–J2 reference states converged? to reviewer 7G3C.

For the nonlocal quantities specifically, we computed the central-cut von Neumann entropy for all twelve saved J1–J2 MPS states and checked the spectra using independent canonicalizations. At J2/J1=0.5J_2/J_1=0.5, the entropy increases from 1.381108 to 1.720615, 2.128044 and 2.410560 as DD increases from 64 to 128, 256 and 384. The corresponding results at the other two couplings are included in the same response to reviewer 7G3C. These measurements show why local-observable accuracy and nonlocal-label convergence should be assessed separately: reliable local predictions do not by themselves establish convergence of entanglement entropy. We discuss how the learned entropy predictor depends on the training distribution in our reply to Question 5 below.

For the measurement-noise concern in Weakness 3, please refer to our reply Question 2. Does the reconstruction benefit survive device-specific measurement noise? to reviewer oS7N. It reports the frozen model's performance under IBM Torino and Brisbane calibration-based noise simulations, including comparisons with raw measurements and known-channel inversion.

Question 4. What data and measurement resources does pretraining require?

The current training set contains fixed eighty states for each family–size combination, giving 1200 states across five families and three sizes. We audited the number of distinct nonidentity Pauli observables per state:

FamilyN=9N=36N=100
1D Ising / XXZ994231191
2D Ising1356481920
Cluster-Ising35116474719
J1–J21809903078

For these measurement layouts, the number of distinct observables grows as O(N)O(N). For classical acquisition cost, the overall completed DMRG stages took 2.8 hours, with each stage taking approximately 0.6 hours on a 128-core CPU.

Question 5. Does entropy prediction rely on correlations within the training families? How would the framework distinguish states with similar local observables but different global properties?

The frozen QuMUSE representation supports accurate entropy prediction with only 24 target-family labels and no encoder retraining, outperforming the target-mean control in all five families. To examine how this depends on family-specific correlations, we first excluded each target family's entropy labels from both probe training and validation. We then fitted a fresh linear readout using 24 target labels, repeating this experiment with three subsets. Normalization and regularization were chosen using other-family data only. The encoder was pretrained on all five families without entropy labels and remained fixed throughout, so this experiment tests transfer of the entropy readout. The table reports entropy RMSE, with mean and standard deviation across the three subsets for the 24-label readout:

Target familyNo target-family probe labels24 target labelsTarget-mean control, 24 labelsRandom encoder, 24 labelsPauli sketch, 24 labels
1D Ising0.5583610.162837 ±\pm 0.0210570.2372610.221658 ±\pm 0.0089450.234257 ±\pm 0.004354
2D Ising0.7528030.201524 ±\pm 0.0162290.2770500.254544 ±\pm 0.0081610.274190 ±\pm 0.002924
XXZ1.1937690.111915 ±\pm 0.0193450.4408730.141948 ±\pm 0.0079040.191556 ±\pm 0.009275
Cluster-Ising1.2731070.317961 ±\pm 0.0650910.5022250.352886 ±\pm 0.0416490.332537 ±\pm 0.071244
J1–J22.8634990.093691 ±\pm 0.0161650.2937450.136695 ±\pm 0.0042870.284839 ±\pm 0.008810

The errors without target-family labels show that the mapping from the representation to entropy depends on the Hamiltonian family. However, only 24 labels are sufficient to obtain a useful readout. The frozen representation outperforms the target-mean, random-encoder and Pauli-sketch controls in all five families. This supports the value of pretraining: it makes information relevant to entropy accessible through a linear readout, even though entropy was not a pretraining target.

For states with similar local observables, the encoder can exploit informative differences in the measured correlations. Alignment with physical-system representations encourages it to learn such structure, but cannot supply information absent from the inputs. If two states have identical values for every supplied observable and identical measurement metadata, the encoder cannot distinguish them. This is an information constraint shared by any method using those inputs.

We verified this distinction with an exact 12-qubit example. Two states have identical reduced density matrices on every contiguous region of up to three sites, but half-system entropies of 0 and 6ln⁡26\ln 2. An additional Z0Z3Z_0Z_3 measurement distinguishes them, with expectation values of 1 and 0. Thus, ambiguous cases require additional informative measurements. Within the tested families, the probing results demonstrate that QuMUSE already captures correlations useful for entropy prediction and makes them accessible with limited property supervision.

I thank authors for additional experiments and responses. The new DMRG results show some convergence issues. Generalization to unseen Hamiltonian families remains limited. The methodological novelty is relatively modest. Thus I keep my score.

Reviewer_oS7NOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces QuMUSE, a generative framework that learns a unified representation of quantum many-body ground states from local measurements. It pairs a query-based VAE, which reconstructs observables, with a conditional latent flow model, which generates state representations from physical conditions. Experiments on five families of Hamiltonians show promising results in observable reconstruction, conditional generation, few-shot transfer, quantum property estimation, and phase boundary detection.

Critical strengths
  • Well-motivated problem. Learning a shared representation from local measurements is an interesting direction, since it avoids having to reconstruct the full quantum wavefunction.

  • Promising experimental results. The proposed method shows strong reconstruction and generation performance across five Hamiltonian families and system sizes of up to 100 qubits, and it generally outperforms the baselines it is compared against.

Critical weaknesses
  • Limited methodological novelty. The framework mainly combines existing techniques, including a query-based VAE, physical-condition alignment, and latent flow matching. The decoder-aware reconstruction loss is reasonable, but the main contribution seems to be integrating existing methods rather than introducing a fundamentally new learning approach.

  • Relatively weak baselines. Conditional generation is compared mainly against QuaDiM. Stronger comparisons with direct supervised regression or simpler conditional latent prediction methods are missing. Since the Hamiltonians studied have relatively few varying parameters, these baselines are essential to show that a generative flow model is actually needed.

Questions for authors
  • Limited generalization to unseen systems. Table 2 shows large zero-shot errors on unseen Hamiltonian families, especially the Kitaev honeycomb model. In addition, for Kitaev conditional generation, training from scratch slightly outperforms transfer learning. Table 10 also shows noticeably higher errors for several families at unseen system sizes. These results weaken the claim of a broadly generalizable foundation model.

  • Limited validation on real measurements. Most experiments use exact expectation values obtained from numerical simulations. Although Appendix C.1 studies a finite number of measurement shots, the paper does not evaluate real hardware noise or experimental data. It remains unclear whether the reported improvements carry over to real quantum systems.

  • Missing analysis of physical consistency and efficiency. The evaluation focuses mainly on prediction RMSE and does not check whether the generated observables are physically consistent or correspond to valid local quantum states. Moreover, conditional generation uses 50 Euler steps and averages over 8 samples, yet its inference cost is not compared with simpler alternatives.

3 comments

Weakness 1: What is the methodological contribution beyond combining existing components?

Although VAEs and latent generative models are well established in machine learning, adapting them to this setting requires a representation that accommodates these differences while retaining information useful for quantum state characterization. QuMUSE demonstrates that a shared latent space can support both quantum state reconstruction and property estimation, as well as Hamiltonian-conditioned prediction and adaptation to additional systems. We view this unified formulation, together with the demonstrated reuse of the learned representation, as the central contribution of the work.

This formulation also suggests a broader direction: representing quantum systems as an additional modality in multimodal generative models. The fixed-size continuous latent representation could provide an interface to architectures such as Transfusion, which combines language modeling with diffusion over continuous image representations. Extending that approach to QuMUSE's quantum-state representations could allow descriptions of quantum systems to be processed alongside text and, eventually, other modalities such as images, video and audio, as a native representation to the large language models.

Weakness 2: Stronger conditional-prediction baselines.

In the ablation study in Appendix C.5 of the manuscript, we have already discussed the simpler latent prediction model, showing a latent flow model is necessary for better generation performance. We now further train a direct observable regressor using the same training data. The direct regressor has the same number of trainable parameters as the latent flow model. We report the test mean RMSE and standard deviation in the following table:

HamiltonianDirect regressorLatent interpolationLatent flow
1D TFIM0.0094140.0083750.006235
2D TFIM0.0148970.0088780.008727
XXZ0.0400300.0305540.025159
Cluster-Ising0.0947250.1252690.023229
J1J_1–J2J_20.0509740.0254030.024767
Overall0.0420080.0396960.017623

The latent flow outperforms other two methods on all settings, showing that it is preferable for the quantum representation generation task.

Question 1. What generalizes to unseen systems?

Please refer to the reply 2. Does pretraining help on unseen families? to reviewer 7G3C. For generalization across system sizes, with training at 9, 36 and 100 qubits, the frozen model predicts observables at unseen sizes of 25, 49 and 81 qubits without target-size fine-tuning. Mean generation RMSE is 0.006658 for 1D Ising and 0.013579 for 2D Ising, demonstrating useful zero-shot size transfer for these families. Errors are higher for XXZ, Cluster-Ising and J1–J2. Thus, a shared representation can support prediction at unseen sizes, but the accuracy depends on the Hamiltonian family. Besides, note that in Figure 2 (a) of the manuscript, the plot of number of qubits N=200 is obtained by zero-shot prediction. This demonstrates that our model supports reproducing the physical behavior of the studied finite Hamiltonian with zero-shot transfer.

Question 2. Does the reconstruction benefit survive device-specific measurement noise?

To test this, we evaluated the frozen manuscript model using real experimental data from IBM Torino and Brisbane devices. We retain all 120 held-out states with complete observable blocks at 9, 36 and 100 qubits, and performed no retraining on IBM noise. Besides the raw data, we also use the calibrated noise model to reduce the noise. At 1024 shots per distinct observable, we obtain:

DeviceRaw RMSEQuMUSE on raw dataCalibration inversionInversion + QuMUSE
Torino / 5 µs0.0836730.0363610.0983080.025449
Brisbane / 5 µs0.0696150.0305930.0374490.013951

QuMUSE reduces the raw measurement error by 56.1–56.5%, with lower mean error in every family and noise setting. The reduction remains 56.3–56.8% after identity observables are excluded, so it is not explained by trivially recovering identity entries. It also complements calibration-based correction: applying QuMUSE after exact-channel inversion reduces the inversion error by 62.7–74.1%. We also tested training from noisy data, with noisy inputs and targets in both training and validation. At 1024 shots, reconstruction RMSE is 0.014043 compared with raw measurement error 0.028993. These results extend the validation beyond ideal finite-shot sampling to noisy measurement channels, showing that our framework can also act as an implicit error mitigation model to hardware noise.

Question 3. Are the outputs physically consistent, and what does inference cost?

For physical consistency, please see the reply 3. Are the predicted local density matrices physically consistent? to reviewer 7G3C.

For inference cost, we highlight that 50 Euler steps are not necessary if inference speed is prioritized, and averaging over 8 samples is only used to evaluate the variance of the prediction, which is never required in actural inference stage. We now benchmark the original frozen checkpoint and the simpler alternatives using one RTX 4090 GPU:

PredictorRMSELatency
Observable interpolation0.0338410.050 ms
Latent interpolation0.0370211.142 ms
Flow: 10 steps0.0178574.459 ms
Flow: 50 steps0.01762321.946 ms

The inference time is negligible even for the 50 Euler steps configuration. Thus, this is not a bottleneck for our method. Further, Reducing flow inference from 50 to 10 Euler steps gives a 4.92x speedup with a 1.33% increase in RMSE, without retraining. If speed is a real need, established distillation methods like DMD for diffusion models can be easily applied to our framework. We leave this for future work.

Dear Authors.

Thank you for the rebuttal. The new regression baseline and IBM hardware results address some of my main concerns. But the physical consistency results also need to be checked. Overall the rebuttal improves the paper, but some key concerns remain. So I decided to keep my score.

Best regards,

Reviewer

Reviewer_7G3COpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors present a unified generative modelling framework for learning ground states of a family of Hamiltonians, given limited local measurements from the ground state. This is done using a variational autoencoder that compresses local expectation values into a fixed-size latent representation, together with a decoder that predicts local observables. The authors then show numerical results on a variety of interesting lattice models including one- and two-dimensional transverse-field Ising models, the one-dimensional XXZ and cluster-Ising models, and the two-dimensional J1-J2 Heisenberg model, with systems of up to 100 qubits. They also demonstrate adaptation to unseen Hamiltonian families, prediction of nonlinear and nonlocal properties, and identification of phase boundaries without phase labels.

Critical strengths

This paper is well written and uses ideas from generative AI to solve a natural problem in quantum tomography. They introduce an learning framework that can potentially have applicability beyond the ground state learning task being solved in the paper

Critical weaknesses
  1. The J1-J2 experiment for a 10x10 lattice only uses a DMRG run with a max bond dimension of 64. This is not usually not large enough to capture the essential physics of a 2D ground state. Are the authors certain that the DMRG runs have converged in bond-dimension and also in the variational energy? That is, if they do a bond-dimension = 128, 256, 384, 512 runs, will they see that same converged energy as the 64 bond dimension used in the paper? This is especially relevant in the frustrated region where J2/J1≊0.5J_2/J_1 \approxeq 0.5.

For instance in the following paper [1] [[https://link.aps.org/doi/10.1103/PhysRevLett.113.027201]], the J1-J2 model is studied much larger bond-dimensions [1] Gong, Shou-Shu, et al. "Plaquette ordered phase and quantum phase diagram in the spin-1 2 J 1-J 2 square Heisenberg model." Physical review letters 113.2 (2014): 027201.

  1. The evidence for the fact that this model generalizes to unseen model families is rather weak. Accurate predictions generally require fine-tuning on six states from the target family, while zero-shot performance is substantially poorer. Moreover, pretraining does not consistently improve conditional generation: for the Kitaev model, training from scratch achieves a slightly lower mean error.

  2. It also worries me that no type of physical consistency is enforced in the model. For instance, are all the predicted three body observables consistent with a physically correct density matrix (i.e. unit trace and positive?). Furthermore, reduced density matrices predicted on overlapping regions should agree on their shared subsystems. While a full global consistency check may be difficult, the authors should at least report local positivity and overlap-consistency tests, particularly because they interpret the latent variables as representations of quantum states.

  3. More generally, while the results of the paper are interesting. I don't see why a QuMUSE type model is more useful to train than a bespoke model for each Hamiltonian class. While the ML approach used here is interesting, the ground state finding task would be much more interesting than the tomography task studied here.

Questions for authors

Please address the weakness section directly

4 comments

Thank you for the careful assessment. Your questions prompted us to test both the reliability of the reference states and the practical value of reusing QuMUSE. The additional experiments show that local physical constraints can be enforced without retraining and with improved accuracy, and that pretraining consistently improves reconstruction after adaptation. We address the questions as follows.

1. Are the 10×10 J1–J2 reference states converged?

During our dataset construction, we measure the average energy differences across different bond dimensions for different Hamiltonian families to examine convergence, and D=64 is enough for the overall convergence. Raising it to 128 results in only 10−710^{-7} energy difference. We further test the specific case: the open 10×10 J1–J2 model at J2/J1=0.4, 0.5 and 0.6. For these three instances, the results show that D=64 is not bond-converged. We use DMRG and increase the bond dimension through D=64, 128, 256 and 384, and report the energy as follows.

J2/J1E/N, D64 →\rightarrow D384Local-observable change, RMSEHalf-system entropy, D64 →\rightarrow D384
0.4−1.954008 →\rightarrow −2.0163960.0444671.336596 →\rightarrow 2.221383
0.5−1.863103 →\rightarrow −1.9235550.0494461.381108 →\rightarrow 2.410560
0.6−1.814384 →\rightarrow −1.8540260.0494071.111496 →\rightarrow 2.054633

We will therefore revise the convergence statement for these 10×10 J1–J2 references in the manuscript.

To determine how this affects the reported learning results, we evaluated the unchanged QuMUSE checkpoint against each completed reference:

J2/J1Reconstruction RMSE, D64 →\rightarrow D384Fixed-generation RMSE, D64 →\rightarrow D384
0.40.088907 →\rightarrow 0.0708780.095016 →\rightarrow 0.074503
0.50.038943 →\rightarrow 0.0571100.043696 →\rightarrow 0.059757
0.60.035104 →\rightarrow 0.0486010.035570 →\rightarrow 0.058686

The error even decreases at J2/J1=0.4 but slightly increases at 0.5 and 0.6. Thus, the 10×10 J1–J2 experiment demonstrates learning of finite-bond reference observables. However, we highlight that all methods are evaluated on the same data, so the reported values in the manuscript are reliable.

We also checked whether the reference sensitivity found in the 10×10 J1–J2 case extends to the smaller systems across the benchmark. We independently diagonalized the Hamiltonians for all fifty nine-qubit test states, covering all five families, and compared the original local-observable labels with exact ground-space values:

FamilyMean local-observable RMSEMaximum local-observable RMSEMaximum energy error per site
1D Ising2.58e-050.0001384.93e-06
2D Ising3.63e-050.00025.99e-06
XXZ6.41e-050.0004499.8e-07
Cluster-Ising4.2e-050.0002337.57e-06
J1–J20.0001320.00078.26e-06

These reference differences have a negligible effect on the measured nine-qubit prediction errors. With the original QuMUSE predictions held fixed, replacing the labels by the exact references changes mean reconstruction RMSE from 0.00590681 to 0.00590602, and mean generation RMSE from 0.01408223 to 0.01408298. The reconstruction inputs and all model weights are unchanged. This directly supports the numerical reliability of the local-observable results for the nine-qubit test sets across all families. The conclusion is specific to these sizes and observables; larger-system convergence and nonlocal properties require separate checks.

We additionally completed the bond-dimension study for all eighty held-out 36- and 100-qubit states from 1D Ising, 2D Ising, XXZ and Cluster-Ising. Each calculation starts from the original saved MPS, refines it at D=64, and continues to D=128 and D=256. This separates the original solver accuracy from the effect of increasing the bond dimension.

FamilyQubitsD128 →\rightarrow D256 local RMSE, mean / maxMaximum energy change per site
1D Ising362.45e-15 / 1.22e-141.87e-14
1D Ising1004.52e-09 / 4.48e-081.65e-11
2D Ising369.7e-08 / 5.23e-078.16e-08
2D Ising1009.4e-05 / 0.0004830.000155
XXZ361.08e-12 / 4.72e-121.74e-14
XXZ1005.77e-09 / 1.21e-084.16e-11
Cluster-Ising360.00213 / 0.01662.16e-11
Cluster-Ising1001.15e-08 / 6.09e-081.64e-09

This further demonstrates that the current D=64 bound dimension is enough to train a reliable model.

To measure the effect of bond dimension on the evaluation in the main experiments, we also compare the original labels with D=256 and rescore the same frozen predictions without training, as shown in the table below.

FamilyQubitsOriginal →\rightarrow D256 reference RMSE, mean / maxReconstruction RMSE, original →\rightarrow D256 labelsGeneration RMSE, original →\rightarrow D256 labels
1D Ising367.13e-06 / 1.62e-050.004012 →\rightarrow 0.0040120.006009 →\rightarrow 0.006009
1D Ising1002.3e-05 / 8.54e-050.005525 →\rightarrow 0.0055250.007022 →\rightarrow 0.007022
2D Ising364.86e-05 / 0.000150.007002 →\rightarrow 0.0070010.008803 →\rightarrow 0.008802
2D Ising1000.000877 / 0.005010.008155 →\rightarrow 0.0082040.011335 →\rightarrow 0.010904
XXZ366.09e-05 / 0.0001890.010200 →\rightarrow 0.0101960.013667 →\rightarrow 0.013660
XXZ1000.00977 / 0.08270.020564 →\rightarrow 0.0266330.027300 →\rightarrow 0.023822
Cluster-Ising360.0162 / 0.08620.017083 →\rightarrow 0.0186120.025626 →\rightarrow 0.027070
Cluster-Ising1000.00308 / 0.02560.014542 →\rightarrow 0.0143950.028463 →\rightarrow 0.028320

These results quantify reference sensitivity for the full held-out sets of these four families at the two larger sizes.

2. Does pretraining help on unseen families?

The clearest benefit is in reconstruction after adaptation. To test whether it persists across data budgets and random seeds, we train paired pretrained and randomly initialized models on nested subsets of two, four and six target states, using three seeds. Both models had the same two validation states and twenty held-out test states. We evaluated fixed checkpoints through 2,500 updates and selected checkpoints using validation RMSE alone. The table reports mean test RMSE with standard deviation across seeds at the 2,500-update budget:

TargetScratch, 6 statesTransfer, 2 statesTransfer, 4 statesTransfer, 6 states
Triangular Ising0.006150 ±\pm 0.0000440.019655 ±\pm 0.0036050.011953 ±\pm 0.0072050.003562 ±\pm 0.000033
Long-range Ising0.008233 ±\pm 0.0009060.003222 ±\pm 0.0002130.003054 ±\pm 0.0002320.002885 ±\pm 0.000012
Kitaev0.008209 ±\pm 0.0001840.046480 ±\pm 0.0075530.024769 ±\pm 0.0187830.005117 ±\pm 0.000033

With six optimization states, pretraining improves reconstruction in all nine target–seed comparisons. Long-range Ising also shows a clear reduction in the required target data: transfer with only two optimization states outperforms scratch training with six in all three seeds.

3. Are the predicted local density matrices physically consistent?

We now test positivity, unit trace and consistency on overlapping regions, and provide a correction that enforces all three jointly. Starting from the predicted observables, we reconstruct the local density matrices and project them onto these constraints. The procedure needs only the predictions and their measurement supports, without reference labels or retraining.

Across 120 test states with complete local observable blocks, all 240 reconstruction and generation corrections converged. For each task, the audit covers 6,750 local blocks and 12,010 overlapping pairs. After correction, all minimum eigenvalues exceed −1e−8, trace and overlap discrepancies are zero to the reported precision, and no constraint violations remain at tolerance 1e−6. Enforcing the constraints also improves prediction accuracy:

TaskOriginal RMSECorrected RMSE
Reconstruction0.0095370.008639
Conditional generation0.0158380.015083

Before correction, 59.3% of reconstruction blocks and 57.4% of generation blocks violate positivity, so this is a substantive correction to the outputs. Therefore, we will add this correction as an optional post-processing step for better property estimation performance. This does not change the representations of the quantum states or the main goal of our research -- learning neural representations for quantum states.

4. Why train a shared model instead of a separate model for each Hamiltonian family?

The practical motivation is to reuse what has already been learned when the Hamiltonian or downstream task changes. A separate model can be effective when sufficient data are available for one fixed family. QuMUSE addresses the setting in which several related systems must be studied and only a small number of reference states can be obtained for a new system. In that setting, the relevant question is whether shared pretraining makes the available target data more useful.

Our adaptation experiment tests this directly. We compare pretrained and randomly initialized models with the same architecture, target data, validation procedure and 2,500-update adaptation budget. With six target optimization states, pretraining reduces reconstruction RMSE from 0.006150 to 0.003562 for triangular Ising, from 0.008233 to 0.002885 for long-range Ising, and from 0.008209 to 0.005117 for Kitaev. The improvement holds in all nine target–seed comparisons. Because the target architecture and adaptation budget are held fixed, these gains demonstrate the usefulness of the pretrained initialization for learning the new system. The comparison controls for architecture; it does not exhaust all possible bespoke architectures.

Long-range Ising provides a particularly concrete example of the practical benefit: two target optimization states with pretraining outperform six without pretraining in every seed, with mean RMSE 0.003222 versus 0.008233. Both methods additionally use the same two validation states. Thus, in this example, existing pretraining reduces the number of new optimization states needed to attain the scratch model's accuracy. This matters when obtaining each additional reference state is costly. Pretraining itself has an upfront cost, which can be shared across subsequent applications; we have not established the number of applications needed to recover that cost.

We also test reuse within the original families by withholding the 36-qubit Ising and XXZ cells. Joint training achieves generation RMSE 0.014964 and 0.069866, compared with 0.017804 and 0.084080 for family-specific models of the same architecture. This is complementary evidence that a shared model can use data from other families to improve prediction in a missing family–size combination. Together with the controlled adaptation experiment in the manuscript, it identifies a concrete use case for QuMUSE: extending an existing model to a new setting when target-specific data are scarce.

Application to ground state finding. The suggested ground-state task provides another way to use the learned representation. We decode Hamiltonian-conditioned QuMUSE predictions into local observables, fit a normalized wavefunction to those moments, and then minimize its variational energy under the target Hamiltonian. Across twenty held-out nine-qubit XXZ/J1–J2 conditions, all QuMUSE-assisted trials reach an energy error below 10−410^{-4} per site within two seconds, including prediction and moment fitting. The minimum final ground-subspace fidelity is 0.9999997. These results not only shows another way to use the learned representation, but also provides another direction for ground-state optimization: using observables queried from learned quantum-state representations to initialize a physical wavefunction, followed by variational energy minimization.

Thank you for the responses. Compared to existing DMRG studies on the J1-J2 model, I don't think the DMRG here is converged even at D=384. I have decided not to change my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.