Pith. sign in

REVIEW 4 major objections 4 minor 52 references

Neural quantum states can appear converged above the true ground-state energy because sampling silently drops rare but crucial configurations; the paper shows a temperature-annealed gradient keeps those configurations in play and reaches ch

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 14:06 UTC pith:MPUYUVBL

load-bearing objection Good diagnosis and a plausible idea, but the written algorithm contradicts its own stated mechanism, and the resource mismatch confounds the empirical comparisons. the 4 major comments →

arxiv 2607.18865 v2 pith:MPUYUVBL submitted 2026-07-21 quant-ph cs.LGphysics.chem-phphysics.comp-ph

Enhanced Neural Quantum State via Annealed Gradient Descent

Shiwei Zhou , Yiming Huang , Xiao Yuan , Xiaoxia Cai This is my paper
classification quant-ph cs.LGphysics.chem-phphysics.comp-ph
keywords neural quantum statesvariational Monte Carloannealed gradient descentsubspace trappingfinite-sample biasmode collapseJ1-J2 Heisenberg modelmolecular ground-state energy
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that the practical failure of neural quantum states is often not a limit of what the network can represent but a self-reinforcing sampling instability, which it names subspace trapping. When a physically important configuration is assigned a probability below the sampling resolution, it disappears from many consecutive batches, receives no corrective gradient, and the optimization effectively minimizes the energy of a projected Hamiltonian on the remaining subspace. The result is an apparently converged variational state whose energy lies above the true ground state. To break this loop, the paper introduces annealed gradient descent (AGD), which during an early high-temperature phase reweights sampled gradient contributions so that low-probability configurations are logarithmically amplified, and then anneals to the standard gradient. The paper reports that AGD lets compact recurrent-network ansatze reach chemical accuracy on several molecules and small relative errors on frustrated J1-J2 spin models, with the same per-step cost as plain gradient descent.

Core claim

The central claim is that 'subspace trapping' is a distinct failure mode of neural quantum state optimization. Theorem 1 states that if a configuration σ0 is never sampled over the optimization, the gradient dynamics is mathematically equivalent to finding the exact eigenstate of the subspace Hamiltonian P H P with P = I − |σ0⟩⟨σ0|, so the probability of σ0 is pinned near zero and the state appears converged above the ground-state energy. The numerical evidence for the trapped state is that its variational energy matches the ground state of the projected Hamiltonian to 3.69×10^-8 Ha. AGD is proposed as the remedy: it replaces the linear sampling weight p with a temperature-dependent weight w

What carries the argument

Annealed gradient descent (AGD), a gradient-estimator modification that reweights each sampled configuration's contribution by a temperature factor approximating the entropy-like weight −p log p in its first-order expansion. The temperature T(t) decays linearly from T0 to 1 over a fixed number of steps, so early updates explore a flattened distribution and later updates recover the standard gradient. The theoretical load-bearing piece is Theorem 1, which equates the trapped optimization with exact diagonalization of the projected Hamiltonian P H P, and the numerical load-bearing piece is the cRNN ansatz with exact ancestral sampling under particle-number constraints.

Load-bearing premise

The mechanism stands on the claim that the annealed update in Algorithm 1 averages over the flattened distribution qθ ∝ |ψ|^{2/T}; the pseudocode's weights, applied to samples from |ψ|^2, lack the importance factor |ψ|^{2/T−2} and would instead concentrate probability on already-high configurations.

What would settle it

Run AGD exactly as specified on a small system where one configuration is deliberately suppressed, and track the sign of the probability change for that configuration during the first training steps. If the entropy-like reweighting is active, its probability should increase; direct computation of the expectation under the pseudocode's weights shows the effective target is |ψ|^{2+2/T}, which is more peaked for T>1, so the probability of the suppressed configuration should instead decrease. This sign flip settles whether AGD flattens or sharpens the sampling distribution.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Subspace trapping explains why a neural quantum state can look converged (energy gradient near zero, high fidelity on dominant configurations) while missing the true ground state by more than chemical accuracy.
  • AGD lets compact recurrent networks reach chemical accuracy on LiH, HCl, N2, H2O, NH3, CH4, H2S, C2 and LiCl, with orders-of-magnitude fewer parameters than a transformer-based NQS.
  • On the 2D J1-J2 model at J2/J1=0.5, AGD maintains a relative energy error near 10^-5, whereas the baseline deteriorates to about 10^-3; the gap widens with frustration.
  • AGD adds negligible overhead per optimization step, so it can be combined with large-batch autoregressive sampling without changing the architecture or constructing a parameter-space metric.
  • The projected-Hamiltonian equivalence provides a practical diagnostic: compare the variational energy with the ground-state energy of P H P for the suspected suppressed configuration to detect trapping.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The weight w(p) = −p log p is the Shannon information of the sampled configuration; treating AGD as an information-theoretic regularizer suggests a tunable, schedule-free penalty that could be tested against the annealing version.
  • A matched-budget comparison is needed: AGD runs use 1e12 total autoregressive samples while the baseline uses 500 samples per step, so part of the reported advantage may come from sampling effort rather than annealing.
  • If the importance-weight correction is added, AGD should be exactly equivalent to minimizing energy under a Rényi-flattened distribution; this equivalence would connect it to annealed importance sampling and could be tested on a two-configuration toy model.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper identifies a finite-sample instability in neural quantum state optimization, called subspace trapping, in which important configurations become undersampled and receive vanishing gradient feedback, leading to spurious convergence in a projected subspace. It proposes annealed gradient descent (AGD), a temperature-dependent reweighting of gradient contributions intended to flatten the effective sampling distribution during early training and then anneal to the standard variational gradient. The authors state Theorem 1 connecting unsampled configurations to a projected Hamiltonian, provide a Taylor-expansion analysis of the annealed gradient, and report numerical results on molecular systems and 1D/2D J1-J2 Heisenberg models, claiming chemical accuracy and improved accuracy over GRU-NQS baselines.

Significance. The subspace-trapping phenomenon is a plausible and practically relevant failure mode, and a lightweight gradient-level cure would be valuable. However, the paper's central mechanism is not realized by the algorithm it writes down: the reweighting in Algorithm 1 concentrates probability mass rather than flattening it. The main theorem is largely a restatement of its assumption, and the empirical comparison is confounded by a six-order-of-magnitude mismatch in sampling budget. These are not presentation issues; they undermine the central claim that AGD suppresses metastable trapping. The paper also provides no code or data release, so the possibility that a different reweighting was actually used cannot be checked. If the result held, it would be significant, but the manuscript as written does not establish it.

major comments (4)
  1. [Section II B, Eq. (3), Algorithm 1, lines 3–8] The stated mechanism is not implemented by the written algorithm. Estimating a gradient under q_θ ∝ |ψ_θ|^{2/T} from samples drawn from p_θ = |ψ_θ|^2 requires importance weights q_θ/p_θ ∝ |ψ_θ|^{2/T−2}. Algorithm 1 line 6 instead uses w_i ∝ exp((2/T) log|ψ(σ_i)|) = |ψ_θ(σ_i)|^{2/T}. With normalized weights, E_p[w f] ∝ ∑_σ p_θ(σ)^{1+1/T} f(σ), so for T > 1 the effective target is more concentrated, not flattened. Thus the central claim that AGD “increases the relative contribution of sampled low-probability configurations” is contradicted by the update rule. If the implementation used a different weighting, code or data are needed; as written, the derivation and algorithm are internally inconsistent.
  2. [Appendix D, Eqs. (D1)–(D3)] The analytical derivation starts from the exact sum ∑_σ p_θ(σ)^{1/(1+ε)} over all configurations, not from the estimator implemented by Algorithm 1. The estimator with line-6 weights has an additional factor p_θ(σ) per sample, so it estimates ∑_σ p_θ(σ)^{1+1/T} f(σ), not ∑_σ p_θ(σ)^{1/T} f(σ). Consequently the first-order correction w(p) = -p log p in Eq. (D3), and the claimed logarithmic enhancement for small p, do not describe the algorithm's actual update. The Taylor expansion in ε also does not cover the initial temperatures T0 = 4–10 used in the experiments.
  3. [Theorem 1 and Appendix E] The theorem assumes that σ0 is never sampled and p_θ(σ0) ≪ ε, then concludes that Δp_θ(σ0) = 0. This is largely a restatement of the assumption: if σ0 contributes to no sampled term and its Hamiltonian coupling is dropped, no update can change its probability. The proof in Eq. (E1) drops H_{σσ0} ψ_θ(σ0)/ψ_θ(σ) on the grounds that ψ_θ(σ0) ≈ 0, which is an amplitude assumption, not a sampling assumption. The N2 check in Section II A constructs H_subspace from a configuration identified post hoc as unsampled and verifies energy agreement; that is a consistency check, not evidence for the self-reinforcing trapping mechanism. The theoretical support for the paper's central mechanism is therefore circular in an essential step.
  4. [Appendix C and Fig. 6(b)] The empirical comparison is confounded by computational resources. Appendix C states that AGD uses a total of 1×10^12 autoregressive samples, while the GRU-NQS baseline uses 500 samples per step for 3×10^5 steps, i.e., 1.5×10^8 samples. The claimed advantage of AGD over GRU-NQS in Fig. 6(b) could be due to the six-order-of-magnitude larger sampling budget rather than to AGD itself. Similarly, the parameter-count comparison in Fig. 4(c) compares different architectures and training procedures, not only parameter efficiency. A fair baseline with matched sample budget and computational cost is required before the empirical claims can be assessed.
minor comments (4)
  1. [Throughout] There are numerous typos and grammatical errors (“evluate”, “tempreture”, “suffers from”, “difficulties that arise”), and duplicated references (Refs. [10]/[11], Refs. [32]/[34], Refs. [36]/[33]). These do not affect the technical content but should be corrected.
  2. [Table I and Fig. 4(b)] Reported molecular energies are computed with a fixed random seed (seed = 111) and no error bars or multiple-run statistics are shown in Table I, despite the text stating that shaded areas in Fig. 4(b) represent spreads across runs. This makes it hard to assess the reliability of the claimed chemical accuracy.
  3. [Section II B, Eq. (7)] The statement that the n-th order approximation yields an enhancement scaling as (log p)^n relative to the original gradient is not rigorously defined or derived; Fig. 8 shows numerical comparison for ε = 1 but no closed-form statement is given for general n.
  4. [Appendix C] The sentence “the total number of autoregressive samples is fixed at 1×10^12” is ambiguous: it could mean per run, per molecule, or per experiment. This needs clarification for reproducibility.

Circularity Check

2 steps flagged

Theorem 1 restates its own unsampled-support assumption, and the N2 check fits the projected Hamiltonian post hoc; the AGD weighting mismatch is a correctness issue, not a circularity.

specific steps
  1. self definitional [Section II A, Theorem 1; proof in Appendix E]
    "Assuming the state σ0 is never sampled, that is over the ∆t optimization steps and p θ(σ0)≪ϵ, where ϵ= 1/∆tN s so that p θ(σ0) can be effectively treated as zero throughout these steps. The gradient-based optimization process is mathematically equivalent to finding the exact eigenstate of a subspace Hamiltonian Ĥ, which inevitably leads to the change of p θ(σ0) close to 0, that is ∆p θ(σ0) = 0, during the optimization within this subspace."

    The theorem's conclusion — that p(σ0) cannot recover — is contained in its premise that σ0 is never sampled and p(σ0) is treated as zero throughout. In the proof, the sampled local-energy sum drops the H_{σσ0} term and replaces H with P H P; optimizing that restricted objective cannot, by construction, change the excluded component. Thus the 'mechanism' is a restatement of the assumption that σ0 is invisible to the gradient estimator, not an independent derivation of subspace trapping from finite-sample VMC.

  2. fitted input called prediction [Section II A, N2 subspace-trapping validation]
    "Given Hamiltonian H, let σ0 denote the unsampled configuration and define the projected effective Hamiltonian as Ĥ=P HP, where P=I− |σ0⟩⟨σ0| is the projector onto H S. We find that the variational energy evaluated with the H is extremely close to the ground-state energy of the projected Hamiltonian Ĥ, with |⟨H⟩ − ⟨Ĥ⟩|= 3.69×10 −8 Ha. This agreement indicates that the NQS is effectively trapped within H S, and provides direct numerical evidence for subspace trapping..."

    The projected Hamiltonian is constructed post hoc from the very configuration alleged to be missing. For any state with very small |ψ(σ0)|², ⟨H⟩ almost equals ⟨P H P⟩ automatically, so the 3.69×10^-8 Ha agreement is a consequence of the assumed suppression, not independent confirmation. The check is a consistency check: it fits the explanatory Hamiltonian to the observed trapped state and then reads the agreement as evidence for the trapping mechanism.

full rationale

The paper's main performance claims are benchmarked against external references (FCI, CCSD, and the GRU-NQS baseline), and the cRNN ansatz is built on prior external work [43,44]; those parts are not circularly justified. The circularity is concentrated in the explanatory theory: Theorem 1 assumes the configuration is never sampled and then proves it stays suppressed, and the N2 validation constructs the projected Hamiltonian from the allegedly missing configuration and treats the resulting agreement as independent evidence. I did not count the AGD inconsistency as circularity: Section II B claims qθ ∝ |ψ|^{2/T}, but Algorithm 1 samples from pθ and uses weights |ψ|^{2/T}, whose normalized expectation equals an effective target p^{1+1/T}, not a flattened p^{1/T}; that is an internal-consistency / correctness problem, comparable to the resource mismatch in Appendix C, not a reduction of a prediction to an input. Score 6 reflects that one central theoretical 'prediction' and its numerical validation reduce by construction, while the empirical method itself retains independent content.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central mechanism rests on a temperature schedule with per-system fitted hyperparameters, an idealized 'never sampled' assumption, and an importance-weight identity that the paper's own equations violate. No code or data is shipped, and the empirical advantage is confounded with a much larger sampling budget. No new physical entities are introduced.

free parameters (5)
  • T0 (initial annealing temperature) = 4.0 (most runs), 5.0 (LiCl, 1D J2/J1=0.8), 6.0 (2D J1-J2), 10.0 (H10)
    Controls the strength of the annealed reweighting and is tuned per molecule and per frustration ratio (Tables I-II); no selection rule is given.
  • T_d (temperature decay steps) = 4k-30k depending on system
    Annealing schedule length is tuned per system and affects the exploration-exploitation balance.
  • hidden units per layer = 10-250 depending on system
    Network capacity is chosen per benchmark; the parameter-efficiency claim uses these hand-selected sizes.
  • base learning rate and schedule = 1e-3 with cosine decay or 2.5e-4 with inverse-time decay by system
    Two different learning-rate schedules are selected per difficulty regime, adding a second tuning axis.
  • total optimization steps = 1e4 to 3e5
    Run lengths vary per molecule and per J2/J1 value, so the reported final energies are not obtained under a single fixed protocol.
axioms (5)
  • domain assumption If p_θ(σ0) < ϵ = 1/(Δt N_s), then σ0 is never sampled and can be treated as exactly zero in gradients.
    Basis of Theorem 1; idealization acknowledged in Appendix E as 'highly idealized' and false if σ0 is sampled even a few times.
  • ad hoc to paper A stationary point of VMC optimization in a sufficiently expressive variational manifold is an eigenstate of the effective subspace Hamiltonian.
    Used in the proof of Theorem 1 (Appendix E, Eq. (E3)); generally false for nonlinear or incomplete variational families without additional conditions, and no proof is given.
  • domain assumption Gradient descent over sampled configurations is mathematically equivalent to minimizing the projected Hamiltonian P H P.
    Appendix E drops the H_{σ σ0} term because ψ(σ0)≈0 and ignores indirect Jacobian effects through shared network weights; the equivalence is asserted rather than derived.
  • ad hoc to paper Algorithm 1's reweighting correctly estimates the gradient under the tempered distribution q_θ.
    The paper claims gradients are evaluated over q_θ, but Algorithm 1 samples from p_θ and weights by |ψ|^{2/T}; the standard importance weight for E_q from p-samples would be |ψ|^{2/T-2} (up to a constant). The stated procedure is therefore not an unbiased estimator of the claimed q_θ gradient.
  • standard math The VMC energy and gradient estimators in Eq. (2) are unbiased.
    Standard variational Monte Carlo identity; used throughout as the baseline against which AGD is defined.

pith-pipeline@v1.3.0-alltime-deepseek · 20079 in / 23080 out tokens · 208452 ms · 2026-08-01T14:06:32.943502+00:00 · methodology

0 comments
read the original abstract

Neural quantum states offer expressive representations of quantum many-body wave functions, yet their practical accuracy can be limited by stochastic optimization rather than representational capacity. Here we identify a finite-sample instability, termed subspace trapping, in which physically important configurations become strongly underestimated, remain absent from successive sampling batches and receive insufficient gradient feedback. This self-reinforcing loss of sampled support can confine optimization to an effective subspace and produce apparently stationary states above the true ground state energy. To address this problem, we introduce annealed gradient descent (AGD), a sampling-aware update with annealing factor that temporarily increases the relative contribution of sampled low-probability configurations while limiting the dominance of high-probability ones. We establish the connection between finite-sample support loss and effective subspace optimization, and then evaluate the method across molecular systems, one and two-dimensional $J_1$-$J_2$ models. Annealed gradient descent suppresses metastable trapping, preserves physically relevant configurations and enables compact neural quantum states to attain chemical accuracy and competitive state-of-the-art performance. These results establish AGD as a lightweight complement to expressive neural architectures, improved sampling strategies for scalable quantum many-body optimization.

Figures

Figures reproduced from arXiv: 2607.18865 by Shiwei Zhou, Xiaoxia Cai, Xiao Yuan, Yiming Huang.

Figure 1
Figure 1. Figure 1: FIG. 1. The illustration of subspace trapping in NQS optimization. (Left) Loss landscape in the variational parameter space. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. Subspace trapping in NQS optimization. Comparison between the exact ground state (Full Configuration Interaction, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. The illustration of an AGD-enhanced NQS for a molecular Hamiltonian as an example, showing that vanilla gradient [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4. Performance of AGD-enhanced NQS on molecular systems. (a) Potential energy surfaces for LiH, HCl, and N [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5. Annealing from a higher initial temperature enables convergence to the ground state for [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6. Accuracy of the AGD-enhanced NQS for the [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIG. 7. Comparison between the entropy-like weighting function [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: FIG. 8. Comparison of [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

52 extracted references · 3 linked inside Pith

  1. [1]

    , σN ) in the computational basis, whereNis the number of sites (or spin- orbitals in molecular systems) andσ i ∈ {0,1,

    Autoregressive wave function parametrization For a configurationσ= (σ 1, σ2, . . . , σN ) in the computational basis, whereNis the number of sites (or spin- orbitals in molecular systems) andσ i ∈ {0,1, . . . , d−1}denotes the local state at sitei, the cRNN factorizes the joint probability into a product of conditionals: pθ(σ) = NY i=1 pθ(σi |σ 1, . . . ,...

  2. [2]

    (17) is that it admits exact ancestral sampling: configurations are drawn site by site from the conditional distributions of Eq

    Sampling with physical constraints A key benefit of the autoregressive form Eq. (17) is that it admits exact ancestral sampling: configurations are drawn site by site from the conditional distributions of Eq. (18), producing independent and identically distributed (i.i.d.) samples and thereby avoiding the autocorrelation overhead inherent to Markov-chain ...

  3. [3]

    Love, Al´ an Aspuru-Guzik, and Jeremy L

    Alberto Peruzzo, Jarrod McClean, Peter Shadbolt, Man-Hong Yung, Xiao-Qi Zhou, Peter J. Love, Al´ an Aspuru-Guzik, and Jeremy L. O’Brien. A variational eigenvalue solver on a photonic quantum processor.Nature Communications, 5:4213, 2014

    2014

  4. [4]

    Solving the quantum many-body problem with artificial neural networks.Science, 355(6325):602–606, 2017

    Giuseppe Carleo and Matthias Troyer. Solving the quantum many-body problem with artificial neural networks.Science, 355(6325):602–606, 2017

    2017

  5. [5]

    Quantum entanglement in neural network states.Physical Review X, 7(2):021021, 2017

    Dong-Ling Deng, Xiaopeng Li, and Sankar Das Sarma. Quantum entanglement in neural network states.Physical Review X, 7(2):021021, 2017

    2017

  6. [6]

    Efficient representation of quantum many-body states with deep neural networks.Nature Communications, 8:662, 2017

    Xun Gao and Lu-Ming Duan. Efficient representation of quantum many-body states with deep neural networks.Nature Communications, 8:662, 2017

    2017

  7. [7]

    Deep autoregressive models for the efficient variational simulation of many-body quantum systems.Physical Review Letters, 124(2):020503, 2020

    Or Sharir, Yoav Levine, Noam Wies, Giuseppe Carleo, and Amnon Shashua. Deep autoregressive models for the efficient variational simulation of many-body quantum systems.Physical Review Letters, 124(2):020503, 2020

    2020

  8. [8]

    Hayward, Roger G

    Mohamed Hibat-Allah, Martin Ganahl, Lauren E. Hayward, Roger G. Melko, and Juan Carrasquilla. Recurrent neural network wave functions.Physical Review Research, 2(2):023358, 2020. 13

    2020

  9. [9]

    Barrett, Aleksei Malyshev, and Alexander I

    Thomas D. Barrett, Aleksei Malyshev, and Alexander I. Lvovsky. Autoregressive neural-network wavefunctions for ab initio quantum chemistry.Nature Machine Intelligence, 4:351–358, 2022

    2022

  10. [10]

    Nnqs-transformer: an efficient and scalable neural network quantum states approach for ab initio quantum chemistry

    Yangjun Wu, Chu Guo, Yi Fan, Pengyu Zhou, and Honghui Shang. Nnqs-transformer: an efficient and scalable neural network quantum states approach for ab initio quantum chemistry. InProceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis. Association for Computing Machinery, 2023

    2023

  11. [11]

    Transformer quantum state: A multipurpose model for quantum many- body problems.Physical Review B, 107(7):075147, 2023

    Yuan-Hang Zhang and Massimiliano Di Ventra. Transformer quantum state: A multipurpose model for quantum many- body problems.Physical Review B, 107(7):075147, 2023

    2023

  12. [12]

    Solving the many-electron schr¨ odinger equation with a transformer-based framework.Nature Communications, 16:8464, 2025

    Honghui Shang, Chu Guo, Yangjun Wu, Zhenyu Li, and Jinlong Yang. Solving the many-electron schr¨ odinger equation with a transformer-based framework.Nature Communications, 16:8464, 2025

    2025

  13. [13]

    Solving the many-electron schr¨ odinger equation with a transformer-based framework.Nature Communications, 16(1):8464, 2025

    Honghui Shang, Chu Guo, Yangjun Wu, Zhenyu Li, and Jinlong Yang. Solving the many-electron schr¨ odinger equation with a transformer-based framework.Nature Communications, 16(1):8464, 2025

    2025

  14. [14]

    Transformer refined quantum sampling for strongly correlated electronic structure.arXiv preprint arXiv:2605.24617, 2026

    Xiongzhi Zeng, Ming Gong, Bowen Kan, Yi Fan, Huan Ma, Jianbin Cai, Yancheng Liu, Naibin Zhou, Tao Jiang, Shao- jun Guo, et al. Transformer refined quantum sampling for strongly correlated electronic structure.arXiv preprint arXiv:2605.24617, 2026

    Pith/arXiv arXiv 2026

  15. [15]

    A fully gpu-accelerated framework for high-performance configuration interaction selection with neural network quantum states.arXiv preprint arXiv:2604.15768, 2026

    Daran Sun, Bowen Kan, Haoquan Long, Hairui Zhao, Haoxu Li, Yicheng Liu, Pengyu Zhou, Ankang Feng, Wenjing Huang, Yida Gu, et al. A fully gpu-accelerated framework for high-performance configuration interaction selection with neural network quantum states.arXiv preprint arXiv:2604.15768, 2026

    Pith/arXiv arXiv 2026

  16. [16]

    Accelerating many-body quantum chemistry via generative transformer-enhanced con- figuration interaction.Journal of Chemical Theory and Computation, 21(23):11989–12000, 2025

    Bowen Kan and Honghui Shang. Accelerating many-body quantum chemistry via generative transformer-enhanced con- figuration interaction.Journal of Chemical Theory and Computation, 21(23):11989–12000, 2025

    2025

  17. [17]

    Evaluation of phase networks in transformer-based neural network quantum states.Communications in Computational Chemistry, 7(2):120–126, 2025

    Lizhong Fu, Honghui Shang, and Jinlong Yang. Evaluation of phase networks in transformer-based neural network quantum states.Communications in Computational Chemistry, 7(2):120–126, 2025

    2025

  18. [18]

    From architectures to applications: A review of neural quantum states.Quantum Science and Technology, 9(4):040501, 2024

    Hannah Lange, Anka Van de Walle, Atiye Abedinnia, and Annabelle Bohrdt. From architectures to applications: A review of neural quantum states.Quantum Science and Technology, 9(4):040501, 2024

    2024

  19. [19]

    Neural-network quantum states for many-body physics.The European Physical Journal Plus, 139(7):631, 2024

    Matija Medvidovi´ c and Javier Robledo Moreno. Neural-network quantum states for many-body physics.The European Physical Journal Plus, 139(7):631, 2024

    2024

  20. [20]

    Deep-neural-network solution of the electronic schr¨ odinger equation.Nature Chemistry, 12(10):891–897, 2020

    Jan Hermann, Zeno Sch¨ atzle, and Frank No´ e. Deep-neural-network solution of the electronic schr¨ odinger equation.Nature Chemistry, 12(10):891–897, 2020

    2020

  21. [21]

    Spencer, Alexander G

    David Pfau, James S. Spencer, Alexander G. de G. Matthews, and W. M. C. Foulkes. Ab initio solution of the many-electron schr¨ odinger equation with deep neural networks.Physical Review Research, 2(3):033429, 2020

    2020

  22. [22]

    Fermionic neural-network states for ab-initio electronic structure

    Kenny Choo, Antonio Mezzacapo, and Giuseppe Carleo. Fermionic neural-network states for ab-initio electronic structure. Nature Communications, 11:2368, 2020

    2020

  23. [23]

    S. Sorella. Green function monte carlo with stochastic reconfiguration.Physical Review Letters, 80(20):4558–4561, 1998

    1998

  24. [24]

    Generalized lanczos algorithm for variational quantum monte carlo.Physical Review B, 64(2):024512, 2001

    Sandro Sorella. Generalized lanczos algorithm for variational quantum monte carlo.Physical Review B, 64(2):024512, 2001

    2001

  25. [25]

    Empowering deep neural quantum states through efficient optimization.Nature Physics, 20(9):1476–1481, 2024

    Ao Chen and Markus Heyl. Empowering deep neural quantum states through efficient optimization.Nature Physics, 20(9):1476–1481, 2024

    2024

  26. [26]

    A simple linear algebra identity to optimize large-scale neural network quantum states.Communications Physics, 7(1):260, 2024

    Riccardo Rende, Luciano Loris Viteritti, Lorenzo Bardone, Federico Becca, and Sebastian Goldt. A simple linear algebra identity to optimize large-scale neural network quantum states.Communications Physics, 7(1):260, 2024

    2024

  27. [27]

    Optimizing neural networks with kronecker-factored approximate curvature

    James Martens and Roger Grosse. Optimizing neural networks with kronecker-factored approximate curvature. InProceed- ings of the 32nd International Conference on Machine Learning, volume 37 ofProceedings of Machine Learning Research, pages 2408–2417. PMLR, 2015

    2015

  28. [28]

    A kaczmarz-inspired approach to accelerate the optimization of neural network wavefunctions.Journal of Computational Physics, 516:113351, 2024

    Gil Goldshlager, Nilin Abrahamsen, and Lin Lin. A kaczmarz-inspired approach to accelerate the optimization of neural network wavefunctions.Journal of Computational Physics, 516:113351, 2024

    2024

  29. [29]

    Looking elsewhere: Improving variational monte carlo gradients by importance sampling.Machine Learning: Science and Technology, 7(1):015035, 2026

    Antoine Misery, Luca Gravina, Alessandro Santini, and Filippo Vicentini. Looking elsewhere: Improving variational monte carlo gradients by importance sampling.Machine Learning: Science and Technology, 7(1):015035, 2026

    2026

  30. [30]

    Neural importance resampling: A practical sampling strategy for neural quantum states.arXiv preprint arXiv:2507.20510, 2025

    Eimantas Ledinauskas and Egidijus Anisimovas. Neural importance resampling: A practical sampling strategy for neural quantum states.arXiv preprint arXiv:2507.20510, 2025

    arXiv 2025

  31. [31]

    Aleksei Malyshev, Markus Schmitt, and A. I. Lvovsky. Neural quantum states and peaked molecular wave functions: Curse or blessing?arXiv preprint arXiv:2408.07625, 2024

    Pith/arXiv arXiv 2024

  32. [32]

    A nonstochastic optimization algorithm for neural-network quantum states.Journal of Chemical Theory and Computation, 19(22):8156– 8165, 2023

    Xiang Li, Jia-Cheng Huang, Guang-Ze Zhang, Hao-En Li, Chang-su Cao, Dingshun Lv, and Han-Shi Hu. A nonstochastic optimization algorithm for neural-network quantum states.Journal of Chemical Theory and Computation, 19(22):8156– 8165, 2023

    2023

  33. [33]

    Neural network wave functions and the sign problem.Physical Review Research, 2(3):033075, 2020

    Attila Szab´ o and Claudio Castelnovo. Neural network wave functions and the sign problem.Physical Review Research, 2(3):033075, 2020

    2020

  34. [34]

    Solving statistical mechanics using variational autoregressive networks.Physical Review Letters, 122(8):080602, 2019

    Dian Wu, Lei Wang, and Pan Zhang. Solving statistical mechanics using variational autoregressive networks.Physical Review Letters, 122(8):080602, 2019

    2019

  35. [35]

    Unbiased monte carlo cluster updates with autoregressive neural networks

    Dian Wu, Riccardo Rossi, and Giuseppe Carleo. Unbiased monte carlo cluster updates with autoregressive neural networks. Physical Review Research, 3(4):L042024, 2021

    2021

  36. [36]

    Solving statistical mechanics using variational autoregressive networks.Physical review letters, 122(8):080602, 2019

    Dian Wu, Lei Wang, and Pan Zhang. Solving statistical mechanics using variational autoregressive networks.Physical review letters, 122(8):080602, 2019

    2019

  37. [37]

    Variational neural annealing.Nature Machine Intelligence, 3(11):952–961, 2021

    Mohamed Hibat-Allah, Estelle M Inack, Roeland Wiersema, Roger G Melko, and Juan Carrasquilla. Variational neural annealing.Nature Machine Intelligence, 3(11):952–961, 2021

    2021

  38. [38]

    Unbiased monte carlo cluster updates with autoregressive neural networks

    Dian Wu, Riccardo Rossi, and Giuseppe Carleo. Unbiased monte carlo cluster updates with autoregressive neural networks. Physical Review Research, 3(4):L042024, 2021. 14

    2021

  39. [39]

    Understanding and eliminating spurious modes in variational monte carlo using collective variables.Physical Review Research, 5(2):023101, 2023

    Huan Zhang, Robert J Webber, Michael Lindsey, Timothy C Berkelbach, and Jonathan Weare. Understanding and eliminating spurious modes in variational monte carlo using collective variables.Physical Review Research, 5(2):023101, 2023

    2023

  40. [40]

    Neural annealing and visualization of autoregressive neural networks in the newman–moore model.Condensed Matter, 7(2):38, 2022

    Estelle M Inack, Stewart Morawetz, and Roger G Melko. Neural annealing and visualization of autoregressive neural networks in the newman–moore model.Condensed Matter, 7(2):38, 2022

    2022

  41. [41]

    Machine-learning-assisted monte carlo fails at sampling computationally hard problems.Machine Learning: Science and Technology, 4(1):010501, 2023

    Simone Ciarella, Jeanne Trinquier, Martin Weigt, and Francesco Zamponi. Machine-learning-assisted monte carlo fails at sampling computationally hard problems.Machine Learning: Science and Technology, 4(1):010501, 2023

    2023

  42. [42]

    Empowering deep neural quantum states through efficient optimization.Nature Physics, 20(9):1476–1481, September 2024

    Ao Chen and Markus Heyl. Empowering deep neural quantum states through efficient optimization.Nature Physics, 20(9):1476–1481, September 2024

    2024

  43. [43]

    Bartlett and Monika Musia l

    Rodney J. Bartlett and Monika Musia l. Coupled-cluster theory in quantum chemistry.Reviews of Modern Physics, 79(1):291–352, 2007

    2007

  44. [44]

    Dian Wu, Riccardo Rossi, Filippo Vicentini, Nikita Astrakhantsev, Federico Becca, Xiaodong Cao, Juan Carrasquilla, Francesco Ferrari, Antoine Georges, Mohamed Hibat-Allah, Masatoshi Imada, Andreas M. L¨ auchli, Guglielmo Mazzola, Antonio Mezzacapo, Andrew Millis, Javier Robledo Moreno, Titus Neupert, Yusuke Nomura, Jannes Nys, Olivier Parcollet, Rico Pohl...

    2024

  45. [45]

    Hayward, Roger G

    Mohamed Hibat-Allah, Martin Ganahl, Lauren E. Hayward, Roger G. Melko, and Juan Carrasquilla. Recurrent neural network wave functions.Phys. Rev. Research, 2(2):023358, 2020

    2020

  46. [46]

    Barrett, Aleksei Malyshev, and A

    Thomas D. Barrett, Aleksei Malyshev, and A. I. Lvovsky. Autoregressive neural-network wavefunctions for ab initio quantum chemistry.Nature Machine Intelligence, 4(4):351–358, 2022

    2022

  47. [47]

    Wiley, Chichester, England, 2000

    Trygve Helgaker, Poul Jørgensen, and Jeppe Olsen.Molecular Electronic-Structure Theory. Wiley, Chichester, England, 2000

    2000

  48. [48]

    ¨Uber das Paulische ¨ aquivalenzverbot.Zeitschrift f¨ ur Physik, 47(9):631–651, 1928

    Paul Jordan and Eugene Wigner. ¨Uber das Paulische ¨ aquivalenzverbot.Zeitschrift f¨ ur Physik, 47(9):631–651, 1928

    1928

  49. [49]

    Scalable neural quantum states architecture for quantum chem- istry.Machine Learning: Science and Technology, 4(2):025034, 2023

    Tianchen Zhao, James Stokes, and Shravan Veerapaneni. Scalable neural quantum states architecture for quantum chem- istry.Machine Learning: Science and Technology, 4(2):025034, 2023

    2023

  50. [50]

    Thiessen, Evan E

    Sunghwan Kim, Paul A. Thiessen, Evan E. Bolton, Jie Chen, Gang Fu, Asta Gindulyte, Lianyi Han, Jane He, Siqian He, Benjamin A. Shoemaker, Jiyao Wang, Bo Yu, Jian Zhang, and Stephen H. Bryant. Pubchem substance and compound databases.Nucleic Acids Research, 44(D1):D1202–D1213, 2016

    2016

  51. [51]

    Daniel G. A. Smith, Lori A. Burns, Andrew C. Simmonett, Robert M. Parrish, Matthew C. Schieber, Raimondas Galvelis, Peter Kraus, Holger Kruse, Roberto Di Remigio, Asem Alenaizan, et al. Psi4 1.4: Open-source software for high-throughput quantum chemistry.The Journal of Chemical Physics, 152(18):184108, 2020

    2020

  52. [52]

    McClean, Nicholas C

    Jarrod R. McClean, Nicholas C. Rubin, Kevin J. Sung, Ian D. Kivlichan, Xavier Bonet-Monroig, Yudong Cao, Chengyu Dai, E. Schuyler Fried, Craig Gidney, Brendan Gimby, et al. Openfermion: the electronic structure package for quantum computers.Quantum Science and Technology, 5(3):034014, 2020. 15 Appendix A: Related works on solving training issues of NQS Al...

    2020