{"id":"a0843451-463b-4986-8a65-e5f02a05a0a8","arxiv_id":"2507.19276","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"A bounded-degree graph recurrent neural network, augmented with RBM-inspired correlators and a semistochastic energy estimator, reaches chemical accuracy for H50, an iron-sulfur cluster model, and a 3D H18 cluster.","lead":"This paper designs a neural network for molecular quantum states that combines tensor-network and graph neural network ideas, and shows it can compute ground-state energies of several challenging molecules to chemical accuracy. It also introduces a faster way to evaluate the energy that cuts computation by orders of magnitude.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified; the chemical-accuracy claim is plausible and internally consistent, but the lack of reported statistical error bars on final energies leaves the precision of the claim unverified.","rationale":"The reader identified the BDG graph construction as the weakest assumption; I agree that it is heuristic and empirically validated, but I do not see it as a load-bearing correctness risk because the graph defines the variational ansatz and the correlators can compensate for missing graph edges. The more concrete, testable issue is the absence of statistical error bars on the reported energies, which the reader also noted but did not treat as the central concern. My concern is not that the methods are wrong, but that the precision of the chemical-accuracy claim is not fully quantified. A concrete rerun with standard errors would settle this. The reader's conditional verdict is appropriate; no change is needed.","tokens_in":19748,"tokens_out":1726,"duration_ms":15822,"concrete_test":"Rerun the final energy evaluation for the three headline systems (H50 BDG(1)+cos-RBM, [Fe2S2(SCH3)4]2- with Ising-RBM, H18 BDG(3)+Ising-RBM) with multiple independent optimization seeds and 10 or more independent Markov chains, reporting mean and standard error of the mean for the variational energy. If the reported energy minus two standard errors lies above the chemical-accuracy threshold (for claims at the threshold), the claim weakens; if the error bars are below 0.2 mHa, the claim stands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that BDG-RNN with cos/Ising-RBM correlators reaches chemical accuracy on H50, [Fe2S2(SCH3)4]2-, and H18 at smaller bond dimensions than plain MPS. The weakest link is the heuristic BDG construction (Algorithms S1/S2, based on Fiedler ordering of exchange integrals plus greedy edge addition), which is validated only empirically on the three benchmarks. However, this is not an internal inconsistency: the graph is a variational ansatz choice, and the correlators add expressivity beyond the graph topology. The more concrete concern is the absence of reported statistical error bars on the final energies. Figures show energy convergence curves and errors relative to reference energies, but the paper does not report VMC statistical uncertainties for the final variational energies. Without error bars, the claim of achieving chemical accuracy (e.g., 1.1 mHa vs sub-mHa statistical noise) cannot be fully assessed. The semistochastic estimator is unbiased only after averaging; the paper reports bias/error relative to exact local energy for fixed samples, but not the variance of the final optimized energy. The RBM-inspired correlator derivation (Eqs. 17-19) is internally consistent, and the reweighting is a standard importance-sampling identity. The MPS-initialized training protocol addresses initialization sensitivity. The code is open-source, which is independent support. Thus the central claim is plausible and supported, but the missing statistical precision reporting makes the exact margin of chemical accuracy unverifiable from the manuscript alone.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes three methodological developments for neural-network quantum states (NQS) in molecular electronic structure: (1) a bounded-degree graph recurrent neural network (BDG-RNN) ansatz that generalizes MPS-RNN by using a graph constructed from Fiedler ordering of orbital exchange integrals plus greedy edge addition; (2) RBM-inspired cos and Ising correlators applied on top of a normalized NQS through a reweighting scheme; and (3) a semistochastic local-energy estimator that treats large Hamiltonian matrix elements deterministically and samples the small ones. The methods are benchmarked on H50 in STO-6G, the [Fe2S2(SCH3)4]2- cluster in a CAS(30e,20o) active space, and a 3D 3x3x2 H18 cluster in STO-3G, with the authors reporting energy errors below 1.6 mHa at smaller bond dimensions than plain MPS. The code is released as the open-source PyNQS package.","tokens_in":20082,"tokens_out":9554,"duration_ms":90283,"significance":"If the results hold, the paper is a useful contribution to NQS for quantum chemistry: it provides a practical route to inject tensor-network initialization and graph structure into RNN-based NQS, and the semistochastic local-energy scheme offers substantial computational savings. The reweighting derivation in Eqs. (17)-(19) is clean, internally consistent, and correctly reduces to the standard VMC estimator when f_n = 1. The benchmarks are against external published references, and the energies are variational upper bounds, so the central claim is not circular. The open-source implementation is an additional strength that supports reproducibility. The main reservation is that final optimized energies are presented without statistical error bars, which prevents a full assessment of the chemical-accuracy claim; this is a fixable issue rather than a fundamental flaw.","major_comments":[{"comment":"The paper reports final VMC energies without statistical error bars or standard errors. The energies are stochastic estimates obtained with a semistochastic estimator whose variance depends on the hyperparameters ε and N_ε, yet the reported comparisons (e.g., 0.6 mHa for BDG(1)-RNN at χ=30 in H50, 1.1 mHa for Ising-RBM in Fe2S2, and 1.1 mHa for BDG(3)-RNN at χ=250 in H18) are not accompanied by a measure of sampling uncertainty. Because the central claim is chemical accuracy at the 1.6 mHa scale, the absence of error bars makes it impossible to verify that the differences from the reference energies are statistically meaningful. Please add error bars or standard errors for the final energies and describe how they were estimated.","section":"§2.4, Eq. (22); §4, Figs. 2d, 3c-d, 4c-f"},{"comment":"The BDG graph determines which orbital correlations the ansatz can represent, and its construction is a heuristic based on Fiedler ordering of exchange integrals followed by greedy edge addition. The paper validates this construction only on the three benchmark systems and provides no sensitivity analysis, such as comparison with alternative orbital orderings, random edge selection, or different out-degrees k. Without such evidence or an explicit scope limitation, the claimed systematic improvement over MPS at fixed bond dimension cannot be cleanly separated from the particular graph choice. Please add a sensitivity study or clearly state the heuristic nature and scope of the graph-construction step.","section":"§2.2, Algorithms S1/S2"},{"comment":"The text and figure caption refer to the Fe2S2 result as the 'exact energy', but the reference is taken from Ref. 57, which is a spin-projected MPS calculation rather than an exact result. If that reference has an uncertainty comparable to the reported margins of 1.1-1.7 mHa, the chemical-accuracy statement for this system needs qualification. Please state the precision or convergence of the reference energy, or replace it with a fully converged exact result.","section":"§4.2, Fig. 3 caption and text"}],"minor_comments":[{"comment":"There is a typo: 'maximum out-dedgressk' should read 'maximum out-degree'.","section":"§2.2"},{"comment":"The notation '(n_i)∈{0,1}^{2K}' is ambiguous; it should state that each n_i takes values in {0,1} for i=1,...,2K.","section":"§2.1, Eq. (3)"},{"comment":"The phrase 'rank-(d+1) core tensor' should be 'order-(d+1) core tensor', and the complexity expression O(χ'^d) appears to undercount by one mode unless d is defined differently.","section":"§2.2, Eq. (13)"},{"comment":"The caption says 'CAS(30e,2o)' but the text uses CAS(30e,20o); please correct the typo.","section":"Fig. 3 caption"},{"comment":"The text says 'decreasing variance as N_ε increases', but the figure shows errors relative to the exact local energy; please report the variance or standard error explicitly to support this statement.","section":"§2.4, Fig. 2a"},{"comment":"The bar chart in Fig. 2b is not fully described in the caption; please specify what the bars represent, e.g., computation time in seconds.","section":"§4.1, Fig. 2b"},{"comment":"Minor wording issues: 'we introduce an correlator' should be 'we introduce a correlator', and 'THe' in the first paragraph of §2.1 should be 'The'.","section":"§2.3"},{"comment":"The main text refers to Algorithms S1-S3 and the gradient derivation only in the Supporting Information; for a self-contained reading, include at least a brief description of the greedy edge-selection criterion in the main text.","section":"§2.2 and SI"}],"recommendation":"major_revision","confidential_remarks":"This is a solid methods paper that fits the scope of a quantum-chemistry methodology journal. The reweighting and semistochastic derivations are sound, and the open-source code and external benchmarks are strengths. The most important issue is the missing statistical precision on the final energies, which is load-bearing for the chemical-accuracy claim. With error bars and a sensitivity discussion for the BDG construction, the paper should be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is one of the more honest and useful NQS-for-chemistry papers I've read recently. The central claim—that BDG-RNN with cos/Ising-RBM correlators reaches chemical accuracy on H50, [Fe2S2(SCH3)4]2–, and H18 at bond dimensions well below plain MPS—is credible and backed by variational upper bounds against published references. I don't think the stress-test note needs to push harder; the weak spots are real but they are the usual fixable sort, not load-bearing flaws.\n\nWhat's new: the BDG-RNN generalizes MPS-RNN from chain/2D grids to arbitrary bounded-degree molecular graphs, and the MPS initialization gives it a sensible starting point. The RBM-inspired correlators (cos-RBM and Ising-RBM) are simple multiplicative factors that don't break autoregressive sampling, and the reweighted energy/gradient derivation (Eqs. 17-19) is clean. The semistochastic local energy estimator—deterministic sum over large |H_nm| plus importance-sampled small terms—is unbiased and gives a real speedup, with a 2000x reduction on H50. That's a practical contribution on its own. Code is open source, benchmarks are against external references, and the paper is refreshingly direct about what is and isn't done.\n\nSoft spots, in proportion. The absence of statistical error bars on the final energies is the main one. The semistochastic tests report error relative to exact local energy for fixed samples, but the paper never gives VMC statistical uncertainty for the optimized energies. When your claim is '1.1 mHa vs chemical accuracy,' the reader needs to know whether that's 1.1 ± 0.05 or 1.1 ± 0.5. The graph construction (Fiedler ordering on exchange integrals plus greedy edge addition) is heuristic; it works on the three tested systems, but nothing guarantees it transfers to other molecules. That's a structural assumption worth stating more prominently, but it doesn't invalidate the empirical results. Hyperparameters are tuned per system, which is mild selection, not circularity.\n\nWho should read it: anyone working on NQS for molecular systems, and people doing tensor-network/neural hybrids. It deserves a serious referee; the main revision asks are error bars and a clearer discussion of the graph heuristic's limits. I'd take it; it's a solid step forward.","headline":"Solid methods paper: BDG-RNN plus RBM correlators hits chemical accuracy on three benchmarks, but the missing statistical error bars on final energies need attention before I'd trust the exact margin.","tokens_in":20637,"tokens_out":2540,"would_cite":true,"duration_ms":23769,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A bounded-degree graph recurrent neural network, initialized from a matrix product state and boosted by RBM-inspired correlators, reaches chemical accuracy in strongly correlated molecules at bond dimensions far smaller than plain MPS.","keywords":["neural network quantum states","tensor networks","matrix product states","variational Monte Carlo","strongly correlated electrons","ab initio quantum chemistry","recurrent neural networks","restricted Boltzmann machines"],"falsifier":"Compute the BDG(1)-RNN energy for a molecule whose dominant correlations are known to span orbitals far apart in the Fiedler order (for example, a stretched bond with long-range exchange) and compare with a converged DMRG/MPS reference at large bond dimension; if the error stays above the 1.6 mHa target and does not drop as the graph degree k increases toward full connectivity, the graph-construction assumption is falsified.","tokens_in":19550,"feed_emoji":"⚛️","tokens_out":8640,"duration_ms":73189,"temperature":0.7,"pith_summary":"The paper tries to show that a hybrid wavefunction—a recurrent neural network whose memory cells follow the connectivity of a molecule-specific bounded-degree graph, with parameters initialized from a matrix product state—can solve electronic structure problems at bond dimensions far below those needed by plain MPS, and that adding a cheap restricted-Boltzmann-machine-inspired correlator pushes it to chemical accuracy. It also introduces a semistochastic local-energy estimator that keeps the energy unbiased while cutting evaluation cost by orders of magnitude. The claims are tested on a 50-atom hydrogen chain, a CAS(30e,20o) model of the iron-sulfur cluster [Fe2S2(SCH3)4]2-, and a three-dimensional 18-atom hydrogen cluster, with reported errors below 1.6 mHa in all three.","feed_headline":"Graph-recurrent network states hit chemical accuracy for molecules","feed_subtitle":"Seeded by matrix-product-state wavefunctions and boosted by correlators, it beats plain tensor networks at the same bond dimension.","key_machinery":"The load-bearing object is the bounded-degree graph recurrent neural network (BDG-RNN) memory cell, whose recurrence follows the graph topology rather than a sequential sweep, with the tensor term Tucker-compressed to avoid $O(\\chi^d)$ scaling. The graph topology is fixed by a Fiedler ordering of the exchange integrals $K_{ij}=[ij|ji]$ followed by greedy edge addition to a bounded out-degree $k$; this is what adapts the network to the molecule's entanglement structure. The RBM-inspired correlators $f_n$ (cos and Ising variants) multiply a normalized base NQS $\\Phi(n)$, and the reweighting formula $E_\\theta = \\langle \\tilde{E}_{\\mathrm{loc}}(n)\\rangle_{n\\sim|\\Phi(n)|^2}$ with $\\tilde{E}_{\\mathrm{loc}}(n) = \\sum_m \\tilde{f}_n^* H_{nm} \\tilde{f}_m \\Phi(m)/\\Phi(n)$ preserves autoregressive sampling. The semistochastic local energy completes the machinery by splitting the sum over $m$ into deterministic large-$|H_{nm}|$ and sampled small-$|H_{nm}|$ parts, making the estimator unbiased while reducing cost.","core_discovery":"The central claim is that the expressivity bottleneck of tensor-network states can be lifted by letting a recurrent network act along the edges of a graph chosen from the molecule's own exchange integrals, rather than along a fixed one-dimensional sweep. Concretely, the BDG-RNN memory update at orbital i is $\\mathbf{h}_i^{n_i} = \\sum_{\\alpha \\in \\mathcal{N}(i)} \\mathbf{M}_{i\\alpha}^{n_i} \\mathbf{h}_\\alpha^{n_\\alpha} + \\mathbf{T}_{i\\alpha\\beta\\cdots\\gamma}^{n_i} \\mathbf{h}_\\alpha \\mathbf{h}_\\beta \\cdots \\mathbf{h}_\\gamma + \\mathbf{v}_i^{n_i}$, which reduces to MPS-RNN when the graph is a chain, and the graph itself is generated from the Fiedler ordering of $K_{ij}=[ij|ji]$ plus a greedy edge-addition step. On top of this, multiplying by $f_n = \\prod_k \\cos(b_k + \\sum_i W_{ki} n_i)$ (cos-RBM) or its quadratic generalization (Ising-RBM) yields $|\\Psi\\rangle = \\sum_n f_n \\Phi(n)|n\\rangle$, and the paper shows via a reweighting identity that both the energy and its gradient keep the same autoregressive sampling structure. The semistochastic estimator splits the local-energy sum into a deterministic part over $|H_{nm}| \\geq \\epsilon$ and a stochastic part sampled from $|H_{nm}|$ for the remaining terms, giving an unbiased energy. The numerical evidence is that BDG-RNN with cos/Ising-RBM correlators reaches chemical accuracy at bond dimensions substantially smaller than plain MPS.","pith_inferences":["If the graph-construction recipe transfers, then a molecule's dominant entanglement pattern could be summarized by its exchange-integral connectivity; a natural test beyond the paper is to learn the graph from data rather than greedily adding edges, and compare BDG(k)-RNN energies.","The correlator-reweighting trick is modular: because the base Phi(n) stays normalized and sampling is untouched, deeper or nonlocal correlators could be layered on top without changing the sampler, extending beyond the cos/Ising-RBM forms tested here.","The semistochastic estimator's hyperparameters epsilon and N_epsilon are system-dependent; an adaptive schedule that lowers epsilon as optimization progresses might preserve the speedups on larger molecules without the bias seen at N_epsilon=0.","The near-tie between BDG(2)-RNN and BDG(3)-RNN on H18 suggests that graph topology, not just degree, is the bottleneck; combining the BDG construction with entanglement-minimized orbital orderings (as done for Fe2S2) could be the more important ingredient on 3D systems."],"forward_implications":["On the 1D hydrogen chain H50, BDG(1)-RNN at chi=30 reaches 0.6 mHa error while MPS needs chi>50 for the same accuracy, and adding cos-RBM drops the error to 0.2 mHa.","On the iron-sulfur cluster CAS(30e,20o) model, BDG(1)-RNN reaches chemical accuracy at chi=200 where plain MPS needs chi=400, and the Ising-RBM correlator reaches it at chi=100.","On the 3D H18 cluster, BDG(3)-RNN approaches 1.1 mHa error at chi=250; adding cos/Ising-RBM correlators reaches the same accuracy at chi=150, and BDG(2)-TensorRNN does so at chi=100.","The semistochastic local-energy estimator yields up to a 2000-fold speedup (for H50, about 0.05% of the original cost) and reduces the local-energy scaling from O(K^5) to roughly O(K^2).","The methods are released in the open-source PyNQS package, so the ansatze and estimators can be applied and extended by other groups."],"supporting_citations":[{"why":"Introduces the restricted-Boltzmann-machine neural network quantum state and the VMC/NQS paradigm that this paper extends.","marker":"[12]"},{"why":"Provides the MPS-to-RNN reformulation (tensorial recurrent neural networks) whose memory-cell update BDG-RNN generalizes from 1D/2D lattices to arbitrary bounded-degree graphs.","marker":"[15]"},{"why":"Supplies the Fiedler-ordering algorithm used to order spatial orbitals before constructing the bounded-degree graph.","marker":"[43,44]"},{"why":"Provides the reference energies for the 1D hydrogen chain against which the BDG-RNN errors are measured.","marker":"[55]"},{"why":"Gives the spin-projected MPS reference energy for the iron-sulfur cluster CAS(30e,20o) benchmark.","marker":"[57]"},{"why":"Introduces the quadratic (Ising-type) neural-network correlator form that the Ising-RBM correlator is inspired by.","marker":"[47]"},{"why":"Establishes the screening-based approach to orbital-space VMC whose select-large-terms idea the semistochastic local-energy estimator parallels.","marker":"[32]"},{"why":"Provides the reweighting identity that lets the RBM-inspired correlator be inserted without abandoning autoregressive sampling.","marker":"[48]"}],"fun_headline_variants":["Graph neural network states reach chemical accuracy","Hybrid NQS-tensor network ansatz hits molecular accuracy","Recurrent network on molecular graph outpaces tensor networks","Semi-stochastic sampling speeds up neural quantum states"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the bounded-degree graph built from the Fiedler ordering of the orbital exchange integrals plus greedy edge addition captures the dominant entanglement structure of the molecule; if that graph misses important orbital correlations, the BDG-RNN cannot represent them, and the reported accuracy would not transfer to other molecular systems.","fun_headline_variants_meta":{"raw":{"variants":["Graph neural network states reach chemical accuracy","Hybrid NQS-tensor network ansatz hits molecular accuracy","Recurrent network on molecular graph outpaces tensor networks","Semi-stochastic sampling speeds up neural quantum states"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000591,"raw_usage":{"total_tokens":2883,"prompt_tokens":1166,"completion_tokens":1717,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":782,"completion_tokens_details":{"reasoning_tokens":1655}},"tokens_in":782,"tokens_out":1717,"duration_ms":11904,"temperature":1.0,"reasoning_tokens":1655,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:55:36.796596+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the BDG(1)-RNN energy for a molecule whose dominant correlations are known to span orbitals far apart in the Fiedler order (for example, a stretched bond with long-range exchange) and compare with a converged DMRG/MPS reference at large bond dimension; if the error stays above the 1.6 mHa target and does not drop as the graph degree k increases toward full connectivity, the graph-construction assumption is falsified.","supporting_citations":[{"cited_title":"Expressibility of comb tensor network states (CTNS) for the P-cluster and the FeMo-cofactor of nitrogenase","cited_arxiv_id":null,"evidence_quote":"Establishes the screening-based approach to orbital-space VMC whose select-large-terms idea the semistochastic local-energy estimator parallels."}],"review_version":1}