{"id":"f96b7584-531b-4af1-9231-41aea4089af0","arxiv_id":"2506.02171","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"KAN-based trial wavefunctions reach about 1 percent ground-state energy accuracy for one-dimensional trapped bosons at roughly 10 times lower cost per training step than MLP-based wavefunctions, aided by a transferable two-body cusp term.","lead":"The paper tests Kolmogorov-Arnold networks, a type of neural network, as compact trial wavefunctions for quantum Monte Carlo simulations of interacting one-dimensional bosons, and finds they need about ten times less computation per training step than standard feed-forward networks. It also adds a two-particle cusp function that manages strong short-range interactions, a known pain point in quantum simulations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Tenfold efficiency claim rests on an uncontrolled MLP baseline: architecture, tuning, and steps-to-convergence are unreported, so total time-to-accuracy is not established.","rationale":"The variational energies for the tested one-dimensional systems appear credible and are validated against exact solutions and prior literature, so the correctness of the numerical method is not the main weakness. The cusp treatment is plausible, though its extension to higher dimensions remains untested. The load-bearing weakness is the efficiency comparison, because the paper's distinctive contribution is the claimed cost advantage over standard neural-network ansätze. That claim rests entirely on a single, under-specified MLP comparison with mismatched sampling protocols and no measured time-to-accuracy. Without a controlled benchmark, the central claim outruns the evidence and the paper reduces to a useful proof-of-concept. The reader's CONDITIONAL verdict is therefore appropriate, and no verdict change is needed.","tokens_in":10889,"tokens_out":8443,"duration_ms":85262,"concrete_test":"Rerun the benchmark of Eq. (12) for, e.g., N=8 at g=2 using the published KAN protocol and a documented MLP of matched capacity (state depth/width/activation), with both optimized to the same 1% energy-accuracy target over 10 independent seeds. Measure total wall-clock time and total FLOPs from initialization to target accuracy for both, including hyperparameter tuning effort. If the tuned MLP achieves the target in comparable or less total time, the claimed tenfold cost advantage is refuted; if the KAN remains roughly 10x cheaper under this matched protocol, the claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline assertion—that KANs are 'roughly 10 times cheaper computationally' than MLP-based ansätze—is not supported by the evidence presented. Section III reports the factor only as FLOPs per training step at 10,000 samples for 'a MLP and a KAN-based ansatz providing similar accuracy', but gives no MLP architecture (depth, width, activation), no hyperparameter search, no learning-rate schedule for the MLP, and no number of training steps needed to reach the accuracy target. Figure 2 shows the two methods run with very different protocols: the KAN uses 2,048–32,768 samples per step while the MLP uses 64,000–1,000,000, indicating the comparison was not matched in sampling budget or convergence criteria. Since total time-to-accuracy depends at least as much on steps and sample counts as on per-sample FLOPs, a properly tuned MLP could reach 1% accuracy in comparable or lower total cost; the abstract's central efficiency claim would then collapse. Section V itself concedes that 'it is impossible to define comparison criteria that are objective and useful,' which is an admission that the stated factor of 10 is not a controlled benchmark. No code or data are shipped, so the comparison cannot be independently reproduced.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Kolmogorov-Arnold network (KAN) trial wavefunctions for variational quantum Monte Carlo (VMC) simulations of one-dimensional bosonic systems. The ansatz is ψ = exp(-α Σ x_i^2) exp(-κ(x_1,...,x_N)), where κ is a bosonic KAN, and an additional pair-distance spline term (Eq. 10) is proposed to efficiently represent cusps. The authors validate against the exact solution of Eq. 7 for a solvable model and against prior literature for the delta-interaction trap model (Eq. 12) with N = 2 to 64, reporting 1% accuracy. They claim that KANs are roughly ten times cheaper computationally than MLP-based neural quantum states and that the cusp treatment generalizes to higher-dimensional short-range potentials.","tokens_in":11100,"tokens_out":3485,"duration_ms":35253,"significance":"If the efficiency claim and the cusp-transfer method hold, this would be a practically useful contribution to neural-network quantum Monte Carlo, particularly for nuclear and cold-atom systems with strong short-range interactions. The paper has genuine strengths: the energy minimization is a true variational upper bound, the one-dimensional benchmark results are validated against exact and literature values, and the idea of transferring a two-body cusp spline ν(x) from an N = 2 calculation to many-body systems is concrete and falsifiable. The central weaknesses are that the headline computational-efficiency claim rests on an uncontrolled MLP baseline, and the generalization to higher dimensions is asserted without numerical support.","major_comments":[{"comment":"The tenfold-efficiency claim is not supported by the evidence presented. The text reports only FLOPs per training step at 10,000 samples for 'a MLP and a KAN-based ansatz providing similar accuracy', but gives no MLP architecture (depth, width, activation), no hyperparameter search or learning-rate schedule for the MLP, and no number of steps to reach the target accuracy. Figure 2 shows very different sampling protocols: the KAN uses 2,048–32,768 samples per step while the MLP uses 64,000–1,000,000, so the comparison is not matched in sampling budget. Since total time-to-accuracy depends on steps and sample counts as well as per-sample FLOPs, a controlled benchmark is required. Section V's admission that 'it is impossible to define comparison criteria that are objective and useful' creates an internal tension with the abstract's unqualified '10 times cheaper'. I recommend either removing the quantitative claim or adding a matched benchmark with explicit MLP architecture, tuning, and total wall-clock time to the same accuracy.","section":"§III and Fig. 2"},{"comment":"The claim that the cusp treatment 'should be useful every time strong, short-ranged potentials exist, including in higher dimensions' is unsupported. All numerical tests are one-dimensional. The factorization property invoked for the ansatz is stated for x1 → x2 in the one-dimensional case, and the relationship between the fitted spline ν(x) and the actual Kato cusp condition in higher dimensions is not derived. If the authors wish to claim higher-dimensional applicability, they should either provide a proof that the same ν(x) construction satisfies the appropriate higher-dimensional two-body boundary condition or report a proof-of-principle calculation in 2D or 3D.","section":"§IV"},{"comment":"The claim that 'ν does not have to be trained at all, its value extracted from a N = 2 calculation being sufficient' is central to the transfer-learning advantage, but no quantitative evidence is shown. The text says this 'was verified to high accuracy' without presenting the comparison, error bars, or the actual ν(x) functions. I recommend adding a figure or table that quantifies the accuracy loss (or gain) when the frozen N = 2 ν(x) is used for N = 4, 8, 16, 32, and 64, compared with training ν from scratch.","section":"§IV, transfer of ν(x)"},{"comment":"The representational basis of the method relies on the 'reasonable assumption' that a bosonic KAN is a universal approximator of symmetric functions, which the paper explicitly states is unproven. Since all many-body results use this bosonic KAN, this is a non-trivial gap. I do not consider it grounds for rejection given the empirical validation, but the authors should either provide a proof or cited theorem, or add a test of representation quality on a non-symmetric target symmetric function beyond the specific models studied.","section":"§II, Eq. (5)"}],"minor_comments":[{"comment":"The denominator in the energy expression should be ∫ dx |ψ|^2, not ∫ dx ψ^2, and there are stray 'R' characters in the integral notation. Please correct these typos.","section":"Eq. (13)"},{"comment":"The caption does not specify the MLP architecture, activation function, number of hidden layers, or the convergence criterion for either curve. Adding these details is important for reproducibility.","section":"§III, Fig. 2 caption"},{"comment":"References [17] and [23] appear to be the same arXiv preprint; please consolidate or differentiate them.","section":"References"},{"comment":"The text says 'measured in floating point operations per second (FLOPs)' but FLOPs is a count of operations, not a rate; the rate is FLOPS (floating-point operations per second). Please correct the unit usage.","section":"§III, unit of cost"},{"comment":"The role of the function c(Δx) could be clarified: it is stated that c(Δx) can be subsumed into ν(x), but Eq. (10) uses ν(tanh(c(x_i - x_j))); please state the domain of the spline and how c(Δx) interacts with the tanh scaling.","section":"Eq. (10)-(11)"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable candidate for publication after major revision. The one-dimensional VMC results appear sound, and the cusp-transfer idea is interesting. The main issues are that the abstract's efficiency claim is overstated relative to the uncontrolled benchmark, and the higher-dimensional generalization is asserted without evidence. I would ask the authors to add a controlled efficiency comparison and to either soften or substantiate the higher-dimensional claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is a genuinely interesting first look at KAN wavefunctions for VMC, and the pair-cusp trick is the best part. The idea of using a per-pair spline that encodes the two-body cusp, then transferring it from N=2 to N=64 without retraining, is practical and clearly validated. The 1D energy results for the delta-interaction model are checked against exact and literature values, and the authors are upfront that the model in Eq. 6 is 'trivial' because the ansatz has the correct behavior baked in. I also appreciate them flagging that bosonic KAN universality is unproven.\n\nThe soft spot is the efficiency claim. The abstract says KANs are 'roughly 10 times cheaper' than other neural-network ansatze. In Section III that number comes from FLOPs per training step at 10,000 samples for one unspecified MLP. No MLP architecture, depth, width, activation, or tuning protocol is given, and Figure 2 shows KAN and MLP run with very different sampling budgets (2,048-32,768 versus 64,000-1,000,000 samples per step). That is not a controlled comparison. The conclusion even concedes that 'it is impossible to define comparison criteria that are objective and useful,' which directly undermines the abstract. This is the load-bearing flaw.\n\nThe higher-dimensional generalization is asserted, not demonstrated, and no code or data are shipped, so the comparison cannot be reproduced. The cusp term is conceptually close to standard Jastrow factors; the transferable aspect is the new part, and it should be positioned that way.\n\nVerdict: the 1D physics is likely sound, but the headline is not. Sent to a serious referee, the paper could be salvaged with a controlled benchmark (matched sample counts, full MLP details, time-to-accuracy) or a softened efficiency claim, plus at least one 2D/3D test. I would send it to review, but with major revision expected.","headline":"KAN wavefunctions: solid 1D results and a clever transferable cusp trick, but the 10x efficiency claim rests on an uncontrolled MLP baseline.","tokens_in":11687,"tokens_out":3872,"would_cite":false,"duration_ms":33926,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Kolmogorov-Arnold networks can serve as trial wavefunctions for quantum Monte Carlo that are roughly ten times cheaper to train than multilayer-perceptron ansatze, with a two-body spline term that captures short-range cusps.","keywords":["Kolmogorov-Arnold networks","variational Monte Carlo","neural quantum states","one-dimensional bosons","Kato's condition","short-range potentials","harmonic trap","transfer learning"],"falsifier":"Run a controlled comparison on the same one-dimensional Hamiltonians, optimizing an MLP ansatz with several architectures, activation functions, and learning-rate schedules to the same energy accuracy, and measure total wall-clock time to convergence; if a reasonably tuned MLP matches or beats the KAN cost, the tenfold efficiency claim fails. A second check is to apply the pair-distance cusp ansatz to a two- or three-dimensional system with strong short-range interactions and test whether the energy error stays near the 1% target without retraining the two-body spline.","tokens_in":10660,"feed_emoji":"⚛️","tokens_out":8035,"duration_ms":80281,"temperature":0.7,"pith_summary":"Kolmogorov-Arnold networks are proposed as a new class of trial wavefunctions for variational quantum Monte Carlo (VMC). The paper argues they represent ground states of one-dimensional trapped bosons accurately while costing roughly ten times fewer floating-point operations per training step than a multilayer-perceptron neural quantum state, at equal sample counts. It also argues that adding a spline function of each interparticle distance builds the short-distance cusp behavior, Kato's condition, directly into the ansatz, so strong short-range potentials do not require very high spline resolution. If these claims hold, VMC simulations of atomic and nuclear systems with short-range interactions become cheaper to run and easier to make accurate.","feed_headline":"KAN wavefunctions deliver VMC energies at roughly one-tenth the cost","feed_subtitle":"A Kolmogorov-Arnold network ansatz matches many-body energies with fewer parameters and cheaper training steps.","key_machinery":"The central object is the Kolmogorov-Arnold network: a representation of a multivariate function as sums of univariate spline functions, grounded in the Kolmogorov-Arnold representation theorem, with quadratic splines used for the line-functions. A bosonic KAN forces the first-layer line-functions to be identical, making the output symmetric under particle exchange. The ansatz is $\\psi = e^{-\\alpha \\sum x_i^2} e^{-\\kappa}$, where $\\kappa$ is either the KAN alone or the KAN plus the pair-distance cusp spline $\\nu(\\tanh(c(x_i-x_j)))$, designed so that the two-body factorization $\\psi \\to \\phi(x_1-x_2)\\Phi((x_1+x_2)/2,\\ldots)$ holds near contact and Kato's condition is satisfied with low-resolution splines. The machinery also includes staged training, gradually increasing the coupling $g$, doubling spline knots when training plateaus, and increasing Monte Carlo sample sizes, plus a reweighting formula that turns delta-function expectation values into computable averages.","core_discovery":"On the paper's own terms, the central discovery is that a Kolmogorov-Arnold network whose line-functions are quadratic splines, symmetrized by sharing the first-layer line-functions, is a workable positive bosonic wavefunction ansatz. Tested on $N=2$ through $N=64$ bosons in a one-dimensional harmonic trap with delta-function and linear interactions, the KAN ansatz reproduces known exact energies (the paper notes one corrected solubility condition) and matches the accuracy of a feed-forward MLP ansatz at roughly one-tenth the computational cost per training step. The paper further finds that appending a two-particle spline term $\\nu(\\tanh(c(x_i-x_j)))$ to the exponent makes the wavefunction satisfy Kato's condition with smooth, low-resolution splines; the same two-body function learned from an $N=2$ calculation works without retraining in the $N=64$ calculation, providing a transfer-learning shortcut that avoids solving the short-distance two-body problem inside each many-body run.","pith_inferences":["The authors leave implicit that the tenfold advantage should be tested as total wall-clock time to a target energy across multiple architectures and tuning schedules, not just FLOPs per step.","A testable extension is to apply the same two-body cusp transfer to fermionic systems or to non-positive wavefunctions, where antisymmetry may interfere with the pair-factorization argument.","If the transfer-learning property holds in higher dimensions, it points toward a modular neural ansatz in which short-distance correlations are fixed by a two-body solve and only long-distance physics is optimized."],"forward_implications":["KAN-based wavefunctions should reach roughly 1% accuracy on one-dimensional trapped-boson ground states with training cost about an order of magnitude below comparable MLP ansatze.","A cusp spline extracted from a two-body calculation transfers to many-body systems with no additional training, so the short-distance two-body problem does not have to be solved repeatedly.","Strong short-range potentials and wavefunction cusps can be represented accurately with low spline resolution, lowering the knot count and Monte Carlo statistics needed near contact.","The Kato-condition mechanism is intended to generalize to higher dimensions and to any system with strong short-range potentials, including realistic atomic and nuclear interactions."],"supporting_citations":[{"why":"Introduces the Kolmogorov-Arnold network architecture that the ansatz builds on.","marker":"[17]"},{"why":"Supplies the spline parameterization and practical KAN usage that the paper follows.","marker":"[23]"},{"why":"Provides the method for computing delta-function expectation values and the ordering procedure used for MLP symmetrization in comparisons.","marker":"[26]"},{"why":"Gives the exactly solvable trapped-boson model with linear plus delta interactions used to validate energies, with the paper's corrected solubility condition.","marker":"[28]"},{"why":"States Kato's condition, the short-distance relation the cusp spline is designed to satisfy.","marker":"[31]"},{"why":"Provides exact N=2 energies for the delta-interaction trap used in validation.","marker":"[33]"},{"why":"Defines the Tonks-Girardeau impenetrable-boson limit used as a strong-coupling check.","marker":"[41]"}],"fun_headline_variants":["KAN ansatz: 10x cheaper quantum Monte Carlo","KAN wavefunctions match many-body accuracy at 1/10 cost","KANs cut quantum Monte Carlo cost tenfold","KAN wavefunction: 10x cheaper than MLP in VMC","Transfer learning in KANs tackles short-range potentials"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The tenfold cost claim follows the paper's assumption that its particular MLP comparison is representative of other neural-network ansatze and that FLOPs per training step at 10,000 samples is a fair proxy for total time to accuracy.","fun_headline_variants_meta":{"raw":{"variants":["KAN ansatz: 10x cheaper quantum Monte Carlo","KAN wavefunctions match many-body accuracy at 1/10 cost","KANs cut quantum Monte Carlo cost tenfold","KAN wavefunction: 10x cheaper than MLP in VMC","Transfer learning in KANs tackles short-range potentials"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000319,"raw_usage":{"total_tokens":1760,"prompt_tokens":867,"completion_tokens":893,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":483,"completion_tokens_details":{"reasoning_tokens":806}},"tokens_in":483,"tokens_out":893,"duration_ms":7877,"temperature":1.0,"reasoning_tokens":806,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:29:19.175631+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a controlled comparison on the same one-dimensional Hamiltonians, optimizing an MLP ansatz with several architectures, activation functions, and learning-rate schedules to the same energy accuracy, and measure total wall-clock time to convergence; if a reasonably tuned MLP matches or beats the KAN cost, the tenfold efficiency claim fails. A second check is to apply the pair-distance cusp ansatz to a two- or three-dimensional system with strong short-range interactions and test whether the energy error stays near the 1% target without retraining the two-body spline.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the method for computing delta-function expectation values and the ordering procedure used for MLP symmetrization in comparisons."},{"cited_title":"Exactly-solvable system of one-dimensional trapped bosons with short and long-range interactions","cited_arxiv_id":"2009.03319","evidence_quote":"Gives the exactly solvable trapped-boson model with linear plus delta interactions used to validate energies, with the paper's corrected solubility condition."},{"cited_title":"Kato, Commun","cited_arxiv_id":null,"evidence_quote":"States Kato's condition, the short-distance relation the cusp spline is designed to satisfy."},{"cited_title":"Busch, B.-G","cited_arxiv_id":null,"evidence_quote":"Provides exact N=2 energies for the delta-interaction trap used in validation."},{"cited_title":"Tonks, Phys","cited_arxiv_id":null,"evidence_quote":"Defines the Tonks-Girardeau impenetrable-boson limit used as a strong-coupling check."}],"review_version":1}