acceptodds
Under review as a conference paper at ICLR 2027

Overlapotron: Zero-Shot Molecular Transfer for Quantum State Preparation

Abstract

Fault-tolerant quantum algorithms for chemistry need an initial state with high overlap with the target state, and the Hartree–Fock overlap decays exponentially with molecular size. Current methods build that state for one molecule at a time, from a classical solution of that same molecule or from a per-molecule quantum optimization. We propose Overlapotron, a model that predicts the initial-state preparation circuit of an unseen molecule from the molecule alone, with no target state and no per-molecule optimization at test time. An equivariant molecular-orbital encoder feeds heads whose output size follows the molecule while the learned weights stay fixed. We show transfer in two regimes. On exact state-vector targets, the heads emit unitary coupled cluster with singles and doubles (UCCSD) and coupled exchange operator (CEO) circuits. On density matrix renormalization group (DMRG) matrix product state (MPS) targets computed at a fixed bond dimension, we introduce the brickwall family, whose overlap with an MPS contracts as a tensor network, so training scales past the state-vector limit. One brickwall model trained on 25,877 QM7 states of at most 88 qubits transfers to QM9 targets of 90 to 112 qubits. Its predicted circuits exceed the overlap of a target-aware non-variational benchmark that compiles each target state directly, at of its T count and of its T depth. On held-out hydrogen chains of 68 to 100 qubits, the margin over the benchmark averages and grows to on the longest chain. Two new datasets of 30,000 CASCI state vectors and 7,165 DMRG matrix product states, with a 25,877-state geometry sweep, support the training.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Quantum circuit synthesis, Quantum state preparation, Geometric deep learning, Out-of-distribution generalization, Quantum chemistry

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_K6pkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Overlapotron, a machine learning model that predicts the parameters of quantum state-preparation circuits directly from molecular geometry and orbital information. The goal is to generate initial states with high overlap with the target ground states, thereby improving the efficiency of downstream quantum algorithms without requiring molecule-specific optimization at inference time.

Critical strengths
  1. The paper presents a comprehensive set of numerical experiments across various molecular systems and experimental settings to evaluate the proposed approach.
  2. Using a machine learning model to directly predict the parameters of state-preparation circuits is an interesting idea that could potentially reduce the need for costly molecule-specific optimization.
Critical weaknesses
  1. The paper does not establish a quantitative criterion for the initial-state overlap required by downstream quantum algorithms. Without such a criterion, it is difficult to determine whether the reported overlap values are sufficiently high to provide meaningful computational advantages. Consequently, the practical significance of the improvements over existing approaches remains unclear.
  2. The main comparisons are against Hartree-Fock (HF) and a target-aware baseline that does not perform per-target optimization. However, HF is typically effective in weakly correlated regimes and can become less competitive for strongly correlated systems. Other potentially relevant approaches, such as MP2-based initial states, selected configuration interaction (selected-CI), are not considered. Without these comparisons, it is difficult to assess the advantages of the proposed learning-based approach over relatively inexpensive classical alternatives.
  3. The paper lacks explanations of several important techniques and concepts, such as MoLE, making the methodology difficult to follow for readers who are not already familiar.
  4. Although the proposed approach improves the initial-state overlap, it remains unclear whether these improvements justify the additional quantum resources required for circuit implementation, particularly the T-gate count.
Questions for authors
  1. Do the reported improvements in initial-state overlap translate into meaningful resource savings for downstream quantum algorithms, such as quantum phase estimation? How do the authors define a "high-overlap" initial state, and what overlap threshold is required to achieve a practical computational advantage?
  2. How does the choice of ansatz depth affect the achievable initial-state overlap and the corresponding T-gate count? Is there an optimal trade-off between circuit complexity and state-preparation accuracy, particularly in the context of fault-tolerant quantum computing?
  3. Why were MP2, selected CI with a small number of determinants not included as baselines?Comparing against at least one of these approaches would help clarify the benefits of learning-based state preparation relative to established classical methods.
1 comment

W1 / Q1 — Do the reported improvements in initial-state overlap translate into meaningful resource savings for downstream quantum algorithms, such as quantum phase estimation?

The success probability of quantum phase estimation (QPE) depends on the squared overlap p between the initial state and the target ground state. More precisely, the number of times QPE must be repeated before success is mostly governed by 1/p. For example, for a molecule with an initial state overlap of 0.2, we expect 5 repetitions of QPE. If that overlap can be increased to 0.8, then only 2 repetitions would be needed, which directly translates in a 2.5x speedup of the quantum algorithm. For our 100 QM9 molecules, the average 1/p falls from 2.93 (HF) to 1.63 (model), a 1.8× speedup; on hydrogen chains, where HF overlap reaches 10⁻⁷, our model yields speedups of several orders of magnitude. Although there is no single threshold, "high overlap" is generally taken to mean that only a handful of repetitions are needed. From a resource standpoint, the relevant quantity is the number of repetitions saved against the extra state preparation cost. Since QPE is generally resource-intensive, saving even a single repetition can be worth a complex state-preparation circuit.

W2 / Q3 — Why were MP2, selected CI with a small number of determinants not included as baselines? Comparing against at least one of these approaches would help clarify the benefits of learning-based state preparation relative to established classical methods.

On the 1500 records of the updated Table 1, MP2 doubles loaded into the UCCSD circuit recover about half of the gain from the predicted UCCSD parameters.

MethodQM9, untouched (1,500 records)
HF0.535 ± 0.010
MP2 in UCCSD circuit0.640 ± 0.011
Brickwall (depth 4), learned0.646 ± 0.011
UCCSD, learned0.715 ± 0.009
CEO, learned0.766 ± 0.008

± = one standard error over the records; learned rows are 5-seed means. Paired difference MP2 − learned UCCSD: −0.075 ± 0.004.

In addition, since MP2 is a perturbation theory method based on Hartree-Fock, we expect it to fail in strongly multireference cases, as is observed in bond dissociation in J. Chem. Phys. 152, 014101 (2020).

Regarding selected CI, it can indeed be used to obtain a compact wavefunction and generate a quantum circuit for initial state preparation. In that sense, it is quite similar to the MPS loader method that we chose as our baseline, since it is also a target-aware method. The recent literature has generally favored the MPS loader approach, which is why we considered it. The two methods were compared in PRX Quantum, 5, 040339 (2024) and in that case, selected CI was found to yield shallower circuits while also being less scalable to large systems. We leave a detailed benchmark of state preparation methods to future work, as our main goal here is to introduce a new competitive method that does not rely on prior knowledge of the target.

W3 — MoLe not explained. We added a self-contained description of the MoLe encoder (App. C.2), with a summary in the main text (Sec. 3.2 and the contributions).

W4 / Q2 — are the extra T gates justified, and what is the optimal depth? Each phase-estimation repetition prepares the initial state once, so the expected cost is (1/p) × (C_prep + C_PE), where p is the squared overlap, C_prep the state-preparation cost, and C_PE the cost of one phase-estimation run. On the 100 QM9 test molecules, our prediction lowers the mean 1/p from 2.93 (HF, whose preparation is free of T gates) to 1.63. The extra T gates therefore pay off whenever 1.63 × (C_prep + C_PE) < 2.93 × C_PE, that is, whenever state preparation costs less than 80% of a single phase-estimation run. The same expression sets the optimal depth.

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

The paper introduces Overlapotron, a machine learning framework that predicts quantum state preparation circuits for unseen molecules without requiring target states or molecule-specific optimization at inference time. It combines an equivariant molecular-orbital encoder with circuit-parameter prediction heads, including a brickwall ansatz that enables scalable training using DMRG-generated matrix product states. The authors demonstrate zero-shot transfer from molecules of up to 88 qubits to unseen molecules containing 90-112 qubits, as well as extrapolation to hydrogen chains of up to 100 qubits, achieving improved state overlaps with reduced estimated circuit resources compared with a target-aware non-variational benchmark.

Critical strengths
  1. The paper addresses an important bottleneck in quantum algorithms by learning to predict state preparation circuits directly from molecular representations. Unlike approaches that optimize circuits independently for each molecule, the proposed framework amortizes state preparation cost across molecular systems and enables zero-shot transfer. The use of equivariant molecular representations and size-independent prediction heads provides a meaningful connection to representation learning.
  2. The experimental results demonstrate transfer to substantially larger molecules than those seen during training, including chemically diverse systems containing 90$-$112 qubits and held-out hydrogen chains up to 100 qubits. The tensor-network-compatible brickwall ansatz enables training beyond the exact state-vector simulation limit.
Critical weaknesses
  1. The evaluations on 90$-$112-qubit molecular systems and long hydrogen chains use DMRG-generated MPS targets with fixed bond dimensions. While this enables efficient training and evaluation, the reported overlaps are measured relative to these approximate states rather than exact ground states. Furthermore, it remains unclear whether the learned model can generalize to more strongly entangled states that cannot be accurately represented by MPS with manageable bond dimensions. This limitation is particularly relevant because such classically intractable states are an important motivation for quantum state preparation.
  2. Although the zero-shot transfer results are promising, the experiments are restricted to C/H/N/O molecules using STO-3G and a common active space construction. The authors also report that the brickwall ansatz performs less effectively when the Hartree-Fock overlap is particularly small, suggesting limitations for strongly correlated states where improved initialization may be most important. The molecular QM9 test set includes only one example with a Hartree-Fock squared overlap below 0.2. Thus, generalization to substantially different chemical systems and more challenging correlation regimes remains insufficiently established.
  3. The proposed architecture builds substantially on the existing M$\overline{\text{o}}$Le framework, and transferable prediction of quantum circuit parameters has also been explored in prior work.
Questions for authors
  1. How sensitive are the reported overlaps to the DMRG bond dimension? Could the authors provide convergence tests on representative systems? More importantly, is there any evidence that the learned model can generalize to strongly entangled states that cannot be accurately represented by MPS with manageable bond dimensions?
  2. The brickwall ansatz appears less effective when the Hartree-Fock overlap is particularly small. Could the authors provide additional evaluations on strongly correlated molecular systems with low Hartree-Fock overlap, where improved initial state preparation is arguably most valuable?
  3. How would the proposed approach compare against target-aware methods incorporating variational refinement, or against alternative initialization strategies at comparable circuit depths and resource budgets? Would the claimed advantages persist under a more realistic fault-tolerant resource model?
  4. The framework relies on a specific orbital construction and MPS ordering. How sensitive is its transfer performance to these choices? In particular, would the learned model generalize to different orbital localization schemes, active-space constructions, or orbital orderings?
  5. How does the number of required training states scale with molecular size and the desired overlap accuracy? Is there any evidence that the training data requirements remain manageable when extrapolating to larger or more strongly correlated systems?
4 comments

(1/2)

We thank the reviewer for the thoughtful comments. We respond to each weakness and question below.

W1 / Q1 (1/2) — How sensitive are the reported overlaps to the DMRG bond dimension? See our answer to Reviewer LApJ (LApJ-W1). For convenience, we reproduce the corresponding data table below. The overlaps and rankings are converged in bond dimension (χ = 250 → 500 → 1000).

χHFModelGain over HFBenchmark (8 layers)
2500.3680.6420.274 ± 0.0060.630
5000.3650.6350.271 ± 0.0060.623
10000.3630.6320.269 ± 0.006—

Model: mean of 3 training seeds, averaged molecule by molecule

W1 / Q1 (2/2) — is there any evidence that the learned model can generalize to strongly entangled states that cannot be accurately represented by MPS with manageable bond dimensions?

We cannot give direct evidence in that regime since we cannot verify overlaps with states that no tractable MPS represents. However, what the paper does show bears on how such states could be reached. The cost of DMRG grows with both the system size and the bond dimension. A target that needs a large bond dimension at a large system size is out of classical reach, but the same kind of state at a small system size is not: small molecules can be solved with a bond dimension large enough to capture any qualitative change in the state. Our results show transfer along the system-size axis at a fixed bond dimension: from QM7 to 90–112-qubit QM9 molecules, to hydrogen chains of up to 100 qubits, and from 8/12/16-orbital to 20/26-orbital active spaces of N₂ at bond lengths not seen in training. Training on small, similar molecules solved at a sufficiently large bond dimension and transferring to large ones is therefore a route to strongly entangled targets that the present framework already supports, even where the state changes qualitatively with the bond dimension.

W2 / Q2 — Could the authors provide additional evaluations on strongly correlated molecular systems with low Hartree-Fock overlap, where improved initial state preparation is arguably most valuable?

Yes, we have additional experiment as we wrote in the response to Reviewer LApJ. To avoid unnecessary back-and-forth, we reiterate our response here.

We agree this is an important direction. We built our work on the standard QM7 and QM9 datasets because this is, to our knowledge, the first study of molecular transfer for quantum initial-state preparation at this scale. A comprehensive study of strongly correlated systems is a necessary next step, but it requires constructing a new large dataset.

Within the rebuttal period, we added two experiments that probe this regime:

(i) N2 dissociation (def2-SVP, SU(2) DMRG, χ = 500): trained on 8/12/16-orbital active spaces at r = 1.0–2.0 Å only, scored zero-shot on 20/26 orbitals at r = 2.1–2.6 Å, where neither the active space nor the bond length was seen in training. The depth-6 brickwall reaches about three times the HF overlap and matches the 8-layer target-aware benchmark.

Test space (r = 2.1–2.6 Å)HFModelBenchmark, 2 layersBenchmark, 8 layersModel − benchmark, 8 layers
20 orbitals (40 qubits)0.166 ± 0.0230.487 ± 0.0380.384 ± 0.0380.475 ± 0.034+0.012 ± 0.008
26 orbitals (52 qubits)0.156 ± 0.0210.460 ± 0.0370.363 ± 0.0290.460 ± 0.027+0.000 ± 0.019

Model: mean of 3 seeds per geometry; ± = one standard error over the 6 geometries (the paired difference removes the shared geometry-to-geometry spread).

(ii) Low-HF QM9 molecules. We generated 25 QM9 molecules (90–108 qubits) deformed until the squared HF overlap fell to 0.01–0.2. The model, trained only on QM7 states with HF overlap ≥ 0.2, raises the overlap above HF on all 25 molecules (2.3× on average) and matches the 2-layer target-aware benchmark, which compiles each target directly:

25 QM9 moleculesHFModelBenchmark, 1 layer2 layers8 layers
Overlap0.089 ± 0.0120.205 ± 0.0290.152 ± 0.0190.202 ± 0.0240.238 ± 0.028
Model − benchmark (paired)+0.052 ± 0.014+0.003 ± 0.013−0.034 ± 0.013

The 8-layer benchmark, which compiles every target separately is better than zero-shot model in this regime (−0.034 ± 0.013). Large datasets in the low-overlap regime — many chemically diverse molecules whose Hartree–Fock overlap is small — do not exist at these sizes, and building them is a necessary next step for this direction.

(2/2)

W3 — The proposed architecture builds substantially on the existing MoLe framework, and transferable prediction of quantum circuit parameters has also been explored in prior work.

For the architecture, our contribution lies in new components that enable molecule-transferable generation of quantum circuit parameters for large systems of 90–112 qubits. For this, the new brickwall head was essential: the ablations show that the new head carries the gain, as we wrote in the response to Reviewer LApJ:

HeadOverlap
Full brickwall head0.378 ± 0.003
without adjacent exchange weight0.371 ± 0.002
without in-loop exchange message0.363 ± 0.004
without chain propagation (the exchange message goes with it)0.238 ± 0.002
inner-product readout (all three removed)0.237 ± 0.001

Previous transferable approaches are limited to predicting circuit parameters across geometries of a single molecule [Koziell-Pipe 2026] or across hydrogen systems of up to 24 qubits [Bincoletto 2025]. We extend transfer to chemically different molecules: the model prepares states for molecules of 90–112 qubits that it never saw in training.

Q3 — How would the proposed approach compare against target-aware methods incorporating variational refinement, or against alternative initialization strategies at comparable circuit depths and resource budgets?

As a target-aware method with variational refinement, we take the circuit generated by the MPS-loading method of Ran (2020) and refine it variationally against the DMRG target (an SU(4) correction per two-qubit gate). For a fair comparison, our method starts from the angles predicted by our model, which needs no target, and refines them against the same DMRG target (L-BFGS on the squared overlap). To match T counts we use the two-layer Ran circuit (≈70k T gates at 10⁻⁴ per rotation for both circuits); its T depth is still about eight times that of our depth-6 brickwall. On all 100 QM9 test molecules, our refined circuit reaches 0.706 against 0.647 for the refined Ran circuit (paired difference +0.059 ± 0.002; higher on 99 of 100). Without any refinement, our zero-shot prediction (0.642) already exceeds the compiled Ran circuit (0.609) by 0.033 ± 0.003 (91 of 100) and is on par with the refined one (−0.005 ± 0.003).

In more realistic fault-tolerant resource models, the number of T-gates that can be executed in parallel depends on the precise architecture used. Our T-depth estimates assume optimal parallelization, and the advantage actually realized will depend on the available architectures. Advantages in terms of the number of T-gates are expected to be architecture-independent.

Q4 — How sensitive is its transfer performance to these choices? In particular, would the learned model generalize to different orbital localization schemes, active-space constructions, or orbital orderings?

In our MPS experiments, we use orbitals localized with the IBO method and the Fiedler ordering, which is the optimal ordering for the MPS since it keeps the chain local, and also allows to efficiently simulate the brickwall Ansatz. For the state vector experiments, the ordering of the orbitals is different, however the MoLe orbital encoder is equivariant to orbital ordering by construction, therefore the model naturally transfers for different orderings.

Regarding the localization scheme, it is important to note that we have freedom to localize orbitals within the active space with any method of our choosing, without changing the corresponding Hamiltonian or eigenvalues. This is also true of the inactive occupied and virtual spaces, as long as the localization does not mix different orbital spaces. Therefore, the main transferability challenge lies in the active space construction, and a rigorous study should consider diverse molecular systems as well as diverse methods for active space selection, which is outside of the scope of the current work. However, our dataset already contains a small variety of active spaces that cover different bonding situations, and we observe reasonable transfer in those cases.

Q5 — How does the number of required training states scale with molecular size and the desired overlap accuracy? Is there any evidence that the training data requirements remain manageable when extrapolating to larger or more strongly correlated systems?

Our hydrogen-chain and N₂ experiments show that, once the model captures the structure of a system well, it transfers successfully with only a small amount of data, although these are simple systems. Whether such small-data transfer is possible for the strongly correlated and general problems we are interested in is future work. We would like to emphasize, as a positive result, that accuracy can be expected to improve further as more data is added.

Thank you for the additional experiments and detailed responses. Many of my questions have been addressed. Nevertheless, my most fundamental concerns remain unresolved.

I understand that the proposed model does not require DMRG calculations for each new molecule at inference time. However, training uses classically tractable MPS of the target ground states. While the authors demonstrate transfer to larger systems and more strongly correlated examples, the method still relies on classically tractable MPS targets. It therefore remains unclear whether the method can generalize to states having larger entanglement complexity than the training data. I think this is an important issue, since the motivation for quantum computing is to solve problems that are classically intractable.

Regarding Q5, I agree that prediction accuracy may improve as more training data are added, but this is generally expected in supervised learning. My question was about the sample complexity. How does the sample complexity grow with the problem size?

Thank you for the additional experiments and detailed responses. Many of my questions have been addressed. Nevertheless, my most fundamental concerns remain unresolved.

Thank you for the careful reading and reviews. Please let us address your remaining concerns.

While the authors demonstrate transfer to larger systems and more strongly correlated examples, the method still relies on classically tractable MPS targets. It therefore remains unclear whether the method can generalize to states having larger entanglement complexity than the training data. I think this is an important issue, since the motivation for quantum computing is to solve problems that are classically intractable.

We agree that generalization to states with larger entanglement complexity is important. We believe that our paper already shows our method has potential to reach states with large entanglement, as we explain below.

The cost of DMRG for a quantum-chemistry Hamiltonian scales as O(L⁴D² + L³D³) per sweep, where L is the number of orbitals and D is the bond dimension, which bounds the entanglement the MPS can represent. Therefore, even for states with the same entanglement (fixed D), the cost grows steeply with molecular size, and DMRG becomes prohibitively expensive for large molecules, where a quantum computer can be useful.

Our paper demonstrates transferability from small to large molecules at similar entanglement regimes. Therefore, our model does not need to extrapolate from small to large entanglement, but can instead learn on small molecules with large entanglement and transfer this knowledge to larger systems, that are out of reach for DMRG. We would also like to emphasize that this "small molecule" regime already includes systems with 88 qubits, where there already is potential for large entanglement.

Regarding Q5, I agree that prediction accuracy may improve as more training data are added, but this is generally expected in supervised learning. My question was about the sample complexity. How does the sample complexity grow with the problem size?

Thank you for clarifying. We define the sample complexity as the number of training states needed to reach a given share of the per-molecule ceiling, i.e. the gain over HF of the same circuit optimized directly against each target. Using our nested training subsets (N = 100 to 25,877, depth-6 brickwall) and splitting the 100 QM9 test molecules (90–112 qubits, all larger than any training molecule) by size, the required N (interpolated in log N between the measured points) is shown below. At every level the larger molecules need fewer, not more, training states, so within the range we can test we see no increase of the sample complexity with system size.

QM9 test molecules60% of ceiling70%75%80%
90–96 qubits≈470≈2,600≈8,000>25,877
98–112 qubits≈120≈620≈2,300≈16,000

Values are interpolated in log N between the measured training sizes (N = 100, 500, 1,500, 3,000, 6,449 and 25,877); >25,877 means the level is not reached within the measured range.

This suggests that the model learns patterns of small substructures that recur in larger molecules, so the difficulty of learning does not change merely because the system becomes larger. How the sample complexity behaves for much larger structures or for strongly correlated systems requires further study.

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

Overlapotron predicts quantum state-preparation circuit parameters for unseen molecules without target-state access or per-molecule circuit optimization at inference. It combines an existing equivariant molecular-orbital encoder with circuit-specific prediction heads. Its most significant contribution is a local brickwall circuit that permits differentiable overlap training against matrix product state targets, enabling training and evaluation beyond dense state-vector sizes. A model trained on QM7 systems of up to 88 qubits transfers to 100 QM9 systems of 90–112 qubits.

Critical strengths
  • Amortizing state-preparation optimization across molecules is useful, and the separation between training-time target access and test-time circuit prediction is clearly defined.

  • The QM7-to-QM9 experiment demonstrates transfer across both molecules and active-space sizes. The hydrogen-chain experiment predicted circuits maintain squared overlaps of approximately 0.32–0.51 on held-out chains despite extremely small Hartree–Fock overlaps.

  • Aligning the circuit’s locality with the MPS ordering makes overlap-based training tractable at the reported sizes.

  • The appendices report formula-clustered uncertainty, benchmark depth sweeps, compilation sensitivity, and rare failures.

Critical weaknesses
  • My biggest concern is that the main molecular experiments measure overlap with DMRG targets at fixed bond dimension 250; hydrogen-chain targets use 500. The authors appropriately acknowledge that these targets are not established as converged ground states. Nevertheless, the proposed downstream application depends on overlap with the actual desired eigenstate, rather than an approximate training target.

  • Also, the principal comparison is with one target-aware, non-variational sequential disentangler. It does not answer whether the molecular encoder and new head outperform simpler transferable predictors.

  • The 300 records used for Table 1 select checkpoints. Splitting by parent molecule prevents training-parent leakage, which is good, but checkpoint-selection performance remains validation performance. Repeating training across five seeds does not remove this selection effect. An untouched parent-molecule test set is needed before treating the CEO and UCCSD results as established generalization performance.

  • Only one QM9 test molecule has Hartree–Fock squared overlap below 0.2. Meanwhile, on the exact-state dataset, brickwall gains only 0.014 on the 59 records below that threshold, compared with 0.512 for CEO. This is a consequential limitation because strongly correlated systems motivate the application. The hydrogen-chain results provide a strong counterexample to a blanket claim that brickwall fails under strong correlation. However, those systems use a highly regular geometry and a specialized orbital ordering. A broader difficult-molecule test is needed to establish chemically diverse transfer in this regime.

  • The encoder and dense readouts are inherited. The clearest contribution appears to be the integration of transferable prediction, the specific brickwall head, tractable overlap training, and new datasets. The paper should distinguish this from novelty of the general brickwall layout.

Questions for authors
  1. How stable are the overlaps and rankings when DMRG target bond dimension increases?
  2. Can the exact-state experiments be evaluated on an untouched test set?
  3. How do simple transferable predictors perform, and what are the separate contributions of chain propagation and exchange conditioning?
  4. How close are predicted circuits to per-target optimized circuits within the same ansatz?
  5. Does the strongest 25,877-state result remain consistent across training seeds?
  6. Can an energy-based check reliably identify the reported near-zero-overlap failures?
2 comments

(1/2)

We thank the reviewer for the thoughtful comments. We respond to each weakness and question below.

W1 / Q1 — How stable are the overlaps and rankings when DMRG target bond dimension increases?

Both remain stable as the target bond dimension increases.

We regenerated all 100 QM9 test targets at χ = 500 and 1000 (SU(2) DMRG) and re-scored the same checkpoints. The model overlap changes by −0.0066 from χ = 250 to 500 and by −0.0037 from 500 to 1000, and its gain over HF only from 0.274 to 0.271 and 0.269. The ranking of all methods is identical at χ = 250 and 500; at χ = 1000 we re-scored only our model, because the benchmark could not be compiled there within our resources.

χHFModelGain over HFBenchmark (8 layers)
2500.3680.6420.274 ± 0.0060.630
5000.3650.6350.271 ± 0.0060.623
10000.3630.6320.269 ± 0.006—

Model: mean of 3 training seeds, averaged molecule by molecule; ± = one standard error over the 100 molecules.

W2 / Q3 — How do simple transferable predictors perform, and what are the separate contributions of chain propagation and exchange conditioning?

We answer the two questions separately (depth-6 brickwall trained on the 500 molecules of QM7-500 with 3 seeds; checkpoints are selected and scored on a separate holdout of 716 QM7 molecules, HF overlap 0.223).

How do simple predictor performs? Replacing MoLe by the Hartree–Fock MO coefficient matrix halves the gain, and with the inner-product readout alone (the simplest transferable predictor) no gain is left.

Encoder inputHeadOverlap
MoLe (full model)full brickwall head0.378 ± 0.003
Hartree–Fock MO coefficient matrixfull brickwall head0.300 ± 0.002
Hartree–Fock MO coefficient matrixinner-product readout only0.224 ± <0.001

Which head features matter? The encoder is fixed to MoLe; head features are removed one at a time. Chain propagation carries about 90% of the gain; the two exchange terms add 0.007 and 0.015.

HeadOverlap
Full brickwall head0.378 ± 0.003
without adjacent exchange weight0.371 ± 0.002
without in-loop exchange message0.363 ± 0.004
without chain propagation (the exchange message goes with it)0.238 ± 0.002
inner-product readout (all three removed)0.237 ± 0.001

Both tables: mean over 3 training seeds ± the standard deviation of the seed means.

W3 / Q2 — Can the exact-state experiments be evaluated on an untouched test set? Yes. Similarly to the QM7 data of Table 1, we stretched 300 QM9 molecules by factors 1.2, 1.4, 1.6, 1.8 and 2.0 (1500 records total) and used them as an untouched test set: no QM9 molecule was used for training or model selection. All methods were scored zero-shot with the Table 1 checkpoints (mean of 5 seeds; ± is one standard error over records):

1,500 QM9 recordsHFMP2 → UCCSDBrickwall (depth 4)UCCSDCEOTarget-aware benchmark (4 layers)
Overlap0.535 ± 0.0100.640 ± 0.0110.646 ± 0.0110.715 ± 0.0090.766 ± 0.0080.820 ± 0.006

On this set the target-aware benchmark, which compiles a circuit for each target state, is the strongest on average (0.820 vs 0.766 for CEO; paired difference 0.053 ± 0.005). For these small systems (16–20 qubits), loading a state with knowledge of the target is strong; the learned predictors never see the target at deployment.

(2/2)

W4 — Strongly correlated systems in addition to the hydrogen-chain experiment.

We agree this is an important direction. We built our work on the standard QM7 and QM9 datasets because this is, to our knowledge, the first study of molecular transfer for quantum initial-state preparation at this scale. A comprehensive study of strongly correlated systems is a necessary next step, but it requires constructing a new large dataset.

Within the rebuttal period, we added two experiments that probe this regime:

(i) N2 dissociation (def2-SVP, SU(2) DMRG, χ = 500): trained on 8/12/16-orbital active spaces at r = 1.0–2.0 Å only, scored zero-shot on 20/26 orbitals at r = 2.1–2.6 Å, where neither the active space nor the bond length was seen in training. The depth-6 brickwall reaches about three times the HF overlap and matches the 8-layer target-aware benchmark.

Test space (r = 2.1–2.6 Å)HFModelBenchmark, 2 layersBenchmark, 8 layersModel − benchmark, 8 layers
20 orbitals (40 qubits)0.166 ± 0.0230.487 ± 0.0380.384 ± 0.0380.475 ± 0.034+0.012 ± 0.008
26 orbitals (52 qubits)0.156 ± 0.0210.460 ± 0.0370.363 ± 0.0290.460 ± 0.027+0.000 ± 0.019

Model: mean of 3 seeds per geometry; ± = one standard error over the 6 geometries (the paired difference removes the shared geometry-to-geometry spread).

(ii) Low-HF QM9 molecules. We generated 25 QM9 molecules (90–108 qubits) deformed until the squared HF overlap fell to 0.01–0.2. The model, trained only on QM7 states with HF overlap ≥ 0.2, raises the overlap above HF on all 25 molecules (2.3× on average) and matches the 2-layer target-aware benchmark, which compiles each target directly:

25 QM9 moleculesHFModelBenchmark, 1 layer2 layers8 layers
Overlap0.089 ± 0.0120.205 ± 0.0290.152 ± 0.0190.202 ± 0.0240.238 ± 0.028
Model − benchmark (paired)+0.052 ± 0.014+0.003 ± 0.013−0.034 ± 0.013

The 8-layer benchmark, which compiles every target separately is better than zero-shot model in this regime (−0.034 ± 0.013). Large datasets in the low-overlap regime — many chemically diverse molecules whose Hartree–Fock overlap is small — do not exist at these sizes, and building them is a necessary next step for this direction.

W5 — inherited components vs new contributions. We agree and now state this explicitly. The MoLe encoder and the dense UCCSD/CEO readouts are inherited from MoLe. What is new is (i) zero-shot prediction of whole state-preparation circuits for unseen molecules, (ii) the brickwall head, which predicts the angles along the chain of the target MPS with chain propagation and orbital-interaction (exchange) features, (iii) overlap training against MPS targets that contract as a tensor network, so training scales past the state-vector limit, and (iv) the new datasets. The general brickwall layout of nearest-neighbour two-qubit gates is standard; the contribution is predicting its angles in a transferable way. The ablations (W2/Q3) show that the new head carries most of the gain: removing chain propagation loses about 90% of it, and the inherited inner-product readout on the same encoder stays at 0.237 vs 0.378 for the full head and 0.223 for HF.

Q4 — How close are predicted circuits to per-target optimized circuits within the same ansatz? On all 100 QM9 test molecules, mean ± standard error: HF 0.368 ± 0.010, zero-shot 0.642 ± 0.012, per-molecule optimum 0.709 ± 0.010; per molecule, the prediction reaches 89.9 ± 0.7% of the optimum and 80 ± 1% of its gain over HF, without the target.

Q5 — Does the strongest 25,877-state result remain consistent across training seeds? Yes. We run additional experiments and obtain gains 0.274 / 0.272 / 0.275 for three seeds; the advantage over the 8-layer benchmark is significant with three seeds (p = 3×10⁻⁵).

Q6 — energy check. We read the question as referring to the rare state-vector failure discussed in the paper (one C₅H₈O₂ geometry at overlap ~10⁻⁷), for which we suggested checking the prepared state's energy. However, the energy alone is in general not a good indicator of the quality of the prepared state. For example, two nearly-degenerate singlet states have very close energies, and yet the true ground state will be orthogonal to the other one. The lack of correlation between energy and overlap has also been reported in PRX Quantum, 5, 040339 (2024), for example in Fig. 12 that shows that methods yielding better energies can result in worse overlap. When the spectral gap of the Hamiltonian is known, the energy error can be used to bound the overlap, however this is rarely the case in practice. The cited PRX Quantum paper proposes a criterion based on energy distributions to assess the quality of the initial state, something that could be explored in future work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response: summary of the update

We thank the reviewers for their careful reading and for recognizing the contributions of this work: predicting state-preparation circuits directly from the molecule, which amortizes the per-molecule optimization (LApJ, auzp, K6pk); transfer to 90–112-qubit molecules and to hydrogen chains of up to 100 qubits while training on smaller systems (LApJ, auzp); a brickwall ansatz aligned with the MPS ordering that makes overlap training tractable beyond the state-vector limit (LApJ, auzp); and a comprehensive set of experiments (LApJ, K6pk).

To address the reviewers' comments, we added the following experiments and uploaded a revised PDF in which changes are marked in red. All overlaps are squared overlaps with the DMRG (or exact) target. For our response to each comment, please see the corresponding reply.

  1. Target convergence (LApJ-W1, auzp-W1). QM9 targets regenerated at bond dimension 500 and 1000. Gain over HF: 0.274 / 0.271 / 0.269 at χ = 250 / 500 / 1000; the ranking of all methods is unchanged.
  2. Ablations and a simple predictor (LApJ-W2, Q3). Chain propagation in the new brickwall head carries about 90% of the gain; replacing the MoLe encoder by the Hartree–Fock MO coefficient matrix halves the gain, and the simplest transferable predictor (that matrix with an inner-product readout) gains +0.001 over HF.
  3. Per-molecule ceiling (LApJ-Q4, auzp-Q3). Optimizing the brickwall angles directly for each QM9 test molecule gives 0.709; the zero-shot prediction reaches 0.642, i.e. 90% of it.
  4. Strong correlation (LApJ-W4, auzp-Q2). N2/def2-SVP, trained on 8/12/16-orbital active spaces at r = 1.0–2.0 Å, zero-shot on 20/26 orbitals at r = 2.1–2.6 Å (neither the active space nor the bond length seen in training): the depth-6 brickwall reaches an averaged overlap for the 20/26 orbital active spaces of 0.487 / 0.460 from HF 0.166 / 0.156, matching the 8-layer target-aware benchmark (0.475 / 0.460).
  5. MP2 baseline (K6pk-W2). MP2 amplitudes in the UCCSD circuit for the 1500 QM9 test molecules: 0.640, vs 0.646 (learned brickwall, depth 4), 0.715 (learned UCCSD) and 0.766 (learned CEO), from HF 0.535.
  6. Seed consistency (LApJ-Q5). Three seeds show consistent results; all QM9 numbers are now three-seed means.
  7. Low-HF-overlap QM9 molecules (LApJ-W4). On 25 QM9 molecules at squared HF overlap 0.01–0.2 (mean 0.089), the model reaches 0.205 for brickwall (depth 6), above HF on every molecule and on par with the 2-layer target-aware benchmark (0.202), though below its 8-layer version (0.238); closing that gap needs datasets in the low-overlap regime.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.