{"id":"0214282b-1f0e-4e86-b8b9-41f1c3bc51e2","arxiv_id":"2502.09488","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A single coupling-conditioned Transformer wavefunction, trained with ensemble Stochastic Reconfiguration, approximates ground states across Hamiltonian families and interpolates to unseen couplings.","lead":"This paper trains one Transformer-based neural network to serve as a variational quantum wavefunction for many different Hamiltonians, feeding both the spin configuration and the Hamiltonian couplings into the network. The same pretrained model can then estimate disorder-averaged properties and detect quantum phase transitions without knowing the order parameters in advance.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The fidelity susceptibility in Eq. (22) is the QGT of the variational state, not the exact ground state; for the J1-J2-J3 model its peak locations are validated only against order parameters from the same FNQS, so the unsupervised phase-transition claim rests on an unverified proxy.","rationale":"The paper's central value proposition is that one pretrained FNQS can replace many system-specific NQS simulations and provide unsupervised phase-transition detection via fidelity susceptibility. The multi-system training itself is well supported: the Ising chain benchmarks are exact, the random-field Ising chain is validated against free-fermion exact results, and the test-realization generalization error decreases with R. These parts of the paper are credible. The weakest link is the use of Eq. (22) for frustrated multi-coupling systems. The QGT of the variational state is a proxy for the exact fidelity susceptibility; no theorem or numerical experiment in the paper shows that variational chi peaks coincide with exact transitions when the ansatz is not exact. The J1-J2-J3 validation is circular: the same wave function is used to compute both chi and the order parameters, so a systematic variational error would shift both consistently. The SI's negative result on cross-phase-boundary extrapolation reinforces that the ansatz's coupling dependence is not universally faithful. A concrete ED-based test on a small frustrated cluster would settle whether the proxy is trustworthy. If the test passes, the conditional verdict can be upgraded; if it fails, the unsupervised-detection claim must be weakened. Thus the reader's CONDITIONAL verdict is appropriate.","tokens_in":24454,"tokens_out":7491,"duration_ms":70333,"concrete_test":"Train the same FNQS architecture on a 4x4 J1-J2-J3 cluster with J2/J1 in [0,1], J3/J1 in [0,0.6] on the same dense grid; compute chi_FNQS via Eq. (22). Independently compute the exact fidelity susceptibility from exact diagonalization of the 4x4 Hamiltonian on the same grid (using Sz conservation). Overlay the two maps and compare peak locations. If the FNQS peak lines deviate from the exact ones by more than the grid spacing, Eq. (22)'s proxy is invalid for frustrated multi-coupling systems; if they match, the central claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. (22) defines chi_ij(gamma) as the quantum geometric tensor of the variational state psi_theta(sigma|gamma): chi_ij = Re{<O_i^dag O_j>_gamma - <O_i^dag>_gamma <O_j>_gamma} with O_i = d log psi_theta(sigma|gamma)/d gamma^i. This equals the exact fidelity susceptibility only if the variational state matches the exact ground state for all gamma in a neighborhood, including its first derivatives. The paper provides no bound or argument that a fixed-capacity Transformer approximates both the ground states and their gamma-derivatives uniformly. The single-coupling Ising chain is validated against the exact solution, but for the J1-J2-J3 model the claimed phase boundaries are checked only against order parameters computed from the same wave function (Fig. 3b-d). If the variational state is inaccurate in a region, e.g., in the valence-bond phase where frustration is strong, both the order parameters and the QGT can be systematically biased, producing peaks that do not correspond to real transitions. The SI's own result that FNQS fails to extrapolate across phase boundaries underscores that the ansatz's gamma-dependence is not a faithful representation of the true ground-state manifold in all regimes. Without an independent check, the central claim of unsupervised detection of quantum phase transitions in multi-coupling systems is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces Foundation Neural-Network Quantum States (FNQS), a Transformer-based variational ansatz that takes both spin configurations and Hamiltonian couplings as input, together with an ensemble formulation of Stochastic Reconfiguration that minimizes the averaged energy L(θ)=∫dγ P(γ)⟨H_γ⟩_γ over a distribution of couplings. The authors demonstrate that a single FNQS can be trained simultaneously on many systems with a fixed total batch size, can be evaluated on couplings not present in the training set, and can be used to compute disorder-averaged observables and a coupling-space fidelity susceptibility. Validation is provided on the one-dimensional transverse-field Ising chain (exact ground-state energies, square magnetization, fidelity susceptibility with finite-size scaling), on the J1-J2-J3 Heisenberg square lattice (order parameters and claimed unsupervised phase boundaries from the fidelity-susceptibility tensor), on the random transverse-field Ising chain (energies, critical correlation functions, and distributions of squared magnetization against numerically exact results), and on an out-of-distribution generalization example for a J2L/J2R diagonal-frustration model. The trained models are publicly released on the Hugging Face Hub with NetKet examples.","tokens_in":24751,"tokens_out":9080,"duration_ms":88187,"significance":"If the results hold, FNQS provide a practical route to amortized quantum many-body ground-state simulation: a single pretrained network could replace many system-specific NQS optimizations, substantially reduce the cost of disorder averaging, and give access to coupling derivatives and hence fidelity susceptibilities. The ensemble-SR construction is a clean and useful contribution with a clear computational scaling (constant cost in the number of systems at fixed total batch), and the exact benchmarks on the Ising and random-Ising chains are convincing. The public release of the trained architectures and NetKet interfaces strengthens reproducibility. The main reservation is that the multi-coupling phase-transition detection on the J1-J2-J3 model is validated only internally against order parameters obtained from the same ansatz, so the paper's most novel application—unsupervised detection of quantum phase transitions—requires an independent check before the claims are fully established. With such a check, the paper would represent a significant step for neural-network quantum states.","major_comments":[{"comment":"The generalized fidelity susceptibility χ_ij(γ) in Eq. (22) is the quantum geometric tensor of the variational state |ψ_θ(γ)⟩, not automatically the exact ground-state fidelity susceptibility. The equality holds only if ψ_θ(σ|γ) reproduces the exact ground state and its first derivatives with respect to the couplings in a neighborhood of γ. For the transverse-field Ising chain this is checked against the exact solution (Fig. 2c), but for the J1-J2-J3 model the boundaries inferred from χ(γ) in Fig. 3a are compared only with order parameters (m_Néel, m_stripe, d^2) computed from the same FNQS wave function (Fig. 3b-d). The observed correspondence is therefore an internal-consistency check, not an independent validation. The Introduction's claim of 'rigorous, unsupervised detection of quantum phase transitions' is not yet established. Please validate at least one cut of the J1-J2-J3 phase diagram against an independent method—for example, exact diagonalization on small clusters as already done at J3=0 in Fig. 4a, quantum Monte Carlo on sign-positive lines, or published tensor-network results (Refs. 47-48)—and explicitly discuss the SI result (Supplementary Fig. 2) that FNQS do not extrapolate across phase boundaries as a limitation of the unsupervised approach.","section":"Methods, Eq. (22); Results, J1-J2-J3 Heisenberg model"},{"comment":"The abstract and introduction state that FNQS 'can generalize to physical Hamiltonians beyond those encountered during training', with the abstract presenting this without qualification. The evidence in the paper is limited to (i) interpolation between training couplings within the same phase (Fig. 2b), (ii) i.i.d. disorder realizations with the same coupling distribution (Fig. 5), and (iii) extrapolation from the axes J2L=0 or J2R=0 to the symmetric point J2L=J2R, where the relative energy error degrades from about 10^-5 at J2/J1=0.1 to 10^-1 at J2/J1=0.6 (Fig. 7c). Supplementary Fig. 2 explicitly shows that a model trained on one side of a quantum phase transition fails on the other side. The generalization claims in the abstract, the Introduction, and the caption of Fig. 1 should be qualified to state that generalization is demonstrated within the training distribution and for limited extrapolation near the training support, and the phase-boundary limitation should appear in the main text where generalization is advertised.","section":"Results, Out-of-distribution generalization; Supplementary Information"},{"comment":"The claim that accuracy 'remains constant with no observable degradation as the number of systems increases' is supported by aggregate relative energy errors without statistical error bars. Because the total batch size M is fixed while per-system sample counts are M/R, the stochastic noise per system grows with R; the reported flat behavior is plausible but needs a more precise substantiation. Please report per-system worst-case errors with error bars, state whether the regularization λ in Eq. (18) and the learning rate were kept fixed across R, and confirm that the flat total-energy error does not hide a spread in individual-system errors. This is needed to fully support the central scalability claim.","section":"Results, Transverse-field Ising chain, Fig. 2a inset; Fig. 5a"}],"minor_comments":[{"comment":"There is a typographical error in the fidelity expression: the ket in the numerator is written as |ψθ(γ+ε⟩| instead of |ψθ(γ+ε)⟩|.","section":"Methods, Eq. (19)"},{"comment":"The sentence 'In the absence of this weighting, S would reduce to an unweighted integral, leading to large statistical fluctuations as the number of systems increases' is unclear: Eq. (17) already defines S as the P(γ)-weighted ensemble average of S(γ). Please clarify what 'unweighted' means and how a different weighting would alter the statistical fluctuations.","section":"Methods, Eqs. (15)-(17)"},{"comment":"The color bar in panel (a) is not labeled and the clipping interval [0.0, 0.5] is mentioned only in the caption text. Please label the color bar as the leading eigenvalue χ_max of the quantum geometric tensor and state more explicitly that values above 0.5 are clipped for visualization.","section":"Fig. 3 caption"},{"comment":"The horizontal and vertical axes appear to be logarithmic, but the axis labels do not indicate this. Please label the axes as log-log and specify the fitted range used to compare Cav(r) with the power law η≈0.382.","section":"Fig. 5b"},{"comment":"The data availability statement says that data are available 'upon request'; since the trained architectures are openly released, please also deposit the numerical data needed to reproduce the histograms in Fig. 6 and the order-parameter maps in Fig. 3, in a public repository.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a strong methods paper with convincing exact benchmarks for the Ising and random-Ising cases and a clean ensemble-SR formulation. The main revision needed is an independent validation of the J1-J2-J3 fidelity-susceptibility boundaries; given the authors' access to exact diagonalization and the availability of QMC and tensor-network references for this model, such a check is feasible. If provided, I would support acceptance. The 'foundation model' framing should also be tempered in the abstract and introduction to avoid overstating out-of-distribution generalization, which the authors themselves show fails across phase boundaries."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuine methods advance. A single Transformer wave function with coupling embeddings, trained with the ensemble-SR variant, can approximate ground states across a family of spin Hamiltonians at once, and two of the three applications are checked against exact or effectively exact references. The J1-J2-J3 fidelity-susceptibility phase diagram is the weak part: it is validated internally, not independently, so treat that section as suggestive.\n\nWhat is new: the P(γ)-weighted S matrix that keeps ensemble SR stable as the number of systems grows, the two coupling-embedding schemes for O(1) and O(N) couplings, and the use of automatic differentiation to get the coupling-space QGT from the variational state. The benchmarks are the strongest part. On the transverse-field Ising chain, relative energy errors reach 1e-7 to 1e-9, and increasing R from 5 to 2000 at fixed total batch size does not degrade accuracy; the fidelity-susceptibility data collapse gives h_c = 1.00(1), nu = 1.00(2). On the random transverse-field Ising chain, disorder-averaged correlations and the full distribution of squared magnetizations on 1000 unseen realizations match free-fermion results for L = 16, 32, 64. Public HuggingFace models and NetKet examples are a real plus.\n\nSoft spots, in proportion. First, Eq. (22) is the QGT of the variational state, not the exact ground state. For the Ising chain the proxy is validated against the exact result, but for the J1-J2-J3 model the claimed phase boundaries are compared only with order parameters from the same wave function. That consistency check is not an independent test, and the SI's own result—FNQS trained only in one phase does not generalize across the boundary—shows the ansatz's coupling dependence is not automatically faithful to the true ground-state manifold. The boundaries may well be right, but the paper should either benchmark against existing tensor-network or series results for J1-J2-J3 or present the unsupervised claim as provisional.\n\nSecond, 'first calculation of fidelity susceptibility for a system with more than one coupling' is too strong. Multi-coupling fidelity susceptibility is standard in the cited literature; the contribution here is doing it with a single differentiable ansatz. That is worth saying, just not as a 'first'. Third, 'computational complexity equivalent to single-system optimization' is true only in the fixed-total-batch-size sense; a sentence on how accuracy depends on M/R would tighten it.\n\nBottom line: this is honest, careful numerical work and it deserves a serious referee. The multi-system training and the disorder results should be publishable; the J1-J2-J3 section needs either an independent benchmark or a downgraded claim.","headline":"One Transformer-based variational state conditioned on Hamiltonian couplings is a real step forward in multi-system NQS, and the disorder results are excellent; the J1-J2-J3 phase-diagram claim relies on self-referential validation and should be read as provisional.","tokens_in":25277,"tokens_out":3194,"would_cite":true,"duration_ms":29362,"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":"The paper claims that one fixed Transformer wave function, trained simultaneously across many Hamiltonians with ensemble Stochastic Reconfiguration, can reproduce per-system ground states and generalize to unseen couplings, with the…","keywords":["foundation model","neural-network quantum states","Transformer wave function","fidelity susceptibility","quantum phase transitions","disordered systems","stochastic reconfiguration","variational Monte Carlo"],"falsifier":"Compute the FNQS fidelity-susceptibility matrix for the $J_1$-$J_2$-$J_3$ model on a small cluster where exact diagonalization gives the phase boundaries: if the peak locations and eigenvector directions do not match the exact order-parameter transitions, the claim that Eq. (22) detects transitions without order parameters is falsified.","tokens_in":24283,"feed_emoji":"🧲","tokens_out":5756,"duration_ms":54442,"temperature":0.7,"pith_summary":"This paper proposes a single neural-network wave function that takes both the spin configuration and the Hamiltonian couplings as input, so that one variational state approximates the ground states of an entire family of Hamiltonians. The authors claim that ensemble Stochastic Reconfiguration optimizes this state across many systems with the same computational cost as optimizing one system and with no accuracy loss as the number of systems grows. If correct, this replaces many system-specific simulations with one pretrained model, making disorder-averaged observables and coupling-space fidelity susceptibility cheap to compute. The paper's own tests show that generalization works inside a phase but not across phase boundaries.","feed_headline":"One wave function learns many Hamiltonians at once","feed_subtitle":"A single Transformer-based quantum state matches per-system accuracy and finds phase transitions without order parameters.","key_machinery":"The machinery is a multimodal Vision Transformer: spin configurations are patched and embedded, and the couplings are either concatenated to every patch when there are $O(1)$ couplings, or patched and embedded with a separate matrix and then concatenated token-by-token when there are $O(N)$ couplings as in disorder, so attention mixes configuration and coupling information. Optimization uses ensemble Stochastic Reconfiguration, which solves $\\mathbf{S}\\dot{\\theta}=-\\tfrac{1}{2}\\mathbf{G}$ where both the geometric matrix $\\mathbf{S}$ and the gradient $\\mathbf{G}$ are averages over the coupling distribution $P(\\gamma)$, regularized by a diagonal shift $\\lambda$. The fidelity susceptibility $\\chi_{ij}(\\gamma)=\\Re\\{\\langle \\hat{O}_{\\gamma,i}^\\dagger \\hat{O}_{\\gamma,j}\\rangle_\\gamma-\\langle \\hat{O}_{\\gamma,i}^\\dagger\\rangle_\\gamma\\langle \\hat{O}_{\\gamma,j}\\rangle_\\gamma\\}$, with $\\hat{O}_{\\gamma,i}$ the diagonal operator of log-amplitude derivatives with respect to coupling $\\gamma^{(i)}$, is then available in closed form through automatic differentiation.","core_discovery":"The central claim is that the ground-state manifold $\\gamma \\mapsto |\\psi_0(\\gamma)\\rangle$ of a Hamiltonian family can be approximated by one fixed-capacity Transformer state $\\psi_\\theta(\\sigma|\\gamma)$ trained on the ensemble loss $\\mathcal{L}(\\theta)=\\int d\\gamma\\,P(\\gamma)\\,\\langle \\hat{H}_\\gamma\\rangle_\\gamma$. The authors demonstrate this on the transverse-field Ising chain, where a single network trained on five field values reproduces all five ground states and interpolates to unseen fields; on the $J_1$-$J_2$-$J_3$ Heisenberg model, where one network trained on 4000 coupling points maps the phase diagram; and on the random transverse-field Ising chain, where one network trained on 1000 disorder realizations matches exact results for disorder-averaged correlations and magnetization distributions. They further claim that the coupling-space geometric tensor, computed by automatic differentiation of the log-amplitude with respect to the couplings, gives a generalized fidelity susceptibility that detects phase transitions without order parameters, and that this is the first such calculation for a system with more than one coupling.","pith_inferences":["Editorial inference: if the coupling-to-ground-state map is smooth enough, the same ensemble-trained state could be differentiated repeatedly in coupling space to produce higher-order thermodynamic derivatives, such as specific heat or susceptibility surfaces, extending the paper's first-derivative demonstration into a surrogate equation of state.","Editorial inference: a natural stress test the paper does not run is to train only on one side of a first-order transition and then measure how the fidelity-susceptibility eigenvector directions degrade, which would quantify how much transition structure is genuinely learned rather than interpolated.","Editorial inference: because attention mixes coupling tokens with configuration tokens, the architecture should transfer to other parameterized Hamiltonian families, such as pressure-driven or electric-field-driven transitions, with no change to the ensemble loss."],"forward_implications":["One trained FNQS yields ground-state energies and correlation functions for many disorder realizations; the paper reports training on $R=1000$ realizations with only 10 Monte Carlo samples per realization and test error matching training error.","The coupling-space quantum geometric tensor, computable by automatic differentiation, gives a generalized fidelity susceptibility for Hamiltonians with several couplings, which the paper states is the first such calculation for more than one coupling.","Pretrained FNQS can be fine-tuned for specific systems, and the released checkpoints let other users start from a common model rather than training from scratch.","Training data must cover every phase of interest: the supplementary information shows that a model trained only in the paramagnetic phase of the transverse-field Ising chain cannot generalize into the ordered phase.","Increasing the number of Transformer layers systematically lowers the variational V-score on the $J_1$-$J_2$ Heisenberg model, so accuracy scales with network capacity even when one network is optimized across 1000 coupling values."],"supporting_citations":[{"why":"Introduced neural-network quantum states, the paradigm this work generalizes from single Hamiltonians to families of Hamiltonians.","marker":"[9]"},{"why":"Introduced Stochastic Reconfiguration, the optimizer that the authors generalize to an ensemble of systems.","marker":"[15]"},{"why":"Developed wave-function optimization in variational Monte Carlo that the ensemble SR update builds on.","marker":"[16]"},{"why":"Provides the variational Monte Carlo framework and the regularized SR update used in all simulations.","marker":"[17]"},{"why":"Introduced the Vision Transformer wave function that FNQS adapts by adding coupling embeddings.","marker":"[21]"},{"why":"Showed that pretrained neural quantum states can be fine-tuned, motivating the foundation-model reuse claim.","marker":"[25]"},{"why":"Established fidelity susceptibility as a computable probe of quantum phase transitions, the baseline the unsupervised detection claim extends.","marker":"[31]"},{"why":"Proved the scaling of geometric tensors at quantum critical points, justifying use of the coupling-space geometric tensor as fidelity susceptibility.","marker":"[32]"}],"fun_headline_variants":["One wave function, many Hamiltonians","Single neural network masters multiple quantum systems","Transformer quantum state generalizes beyond training set","Foundation wave function for any Hamiltonian","Universal quantum ansatz from one pretrained model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole scheme rests on assuming one fixed-capacity neural network can represent the coupling-to-ground-state map smoothly enough that amplitudes and coupling derivatives track the exact states, a property the paper verifies numerically but does not prove and that fails across phase boundaries.","fun_headline_variants_meta":{"raw":{"variants":["One wave function, many Hamiltonians","Single neural network masters multiple quantum systems","Transformer quantum state generalizes beyond training set","Foundation wave function for any Hamiltonian","Universal quantum ansatz from one pretrained model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000591,"raw_usage":{"total_tokens":2785,"prompt_tokens":971,"completion_tokens":1814,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":587,"completion_tokens_details":{"reasoning_tokens":1751}},"tokens_in":587,"tokens_out":1814,"duration_ms":14211,"temperature":1.0,"reasoning_tokens":1751,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T21:19:22.951243+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the FNQS fidelity-susceptibility matrix for the $J_1$-$J_2$-$J_3$ model on a small cluster where exact diagonalization gives the phase boundaries: if the peak locations and eigenvector directions do not match the exact order-parameter transitions, the claim that Eq. (22) detects transitions without order parameters is falsified.","supporting_citations":[{"cited_title":"Wang , author Y.-H","cited_arxiv_id":null,"evidence_quote":"Established fidelity susceptibility as a computable probe of quantum phase transitions, the baseline the unsupervised detection claim extends."}],"review_version":1}