{"id":"64dd1ef2-f986-41bc-8cf9-5d8043f0df60","arxiv_id":"2507.17810","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A variational autoencoder inverse mapper extracts the proton's gluon PDF from lattice QCD pseudo-Ioffe-time distributions, yielding results consistent with global fits.","lead":"This paper trains a machine-learning model to turn lattice QCD measurements of the proton's gluon structure into a gluon parton distribution function. The extracted gluon PDF matches existing global fits in the mid-to-high momentum region, demonstrating a generative AI-based route for combining lattice and phenomenological QCD analyses.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim may be an artifact of the four-parameter ansatz: the VAIM can only return PDFs of the form in Eq. (3), so agreement with global fits does not validate the extraction unless a misspecification test shows coverage.","rationale":"Agreement with reader: The reader's weakest assumption is exactly the one I identify: the four-parameter ansatz of Eq. (3) is the Achilles heel of the central claim. The reader's CONDITIONAL verdict is appropriate, and my stress test does not move it. Credit where due: the paper's external comparison to five global fits is genuine independent evidence, and the closure test is a useful sanity check that the trained VAIM reconstructs the training distribution; the jackknife-to-MC uncertainty propagation is a reasonable attempt at statistical propagation. However, the internal closure test cannot constrain the prior/parametric family, and the paper honestly flags this in Sec. 4.1. The proposed misspecification test directly checks whether the uncertainty band would contain a true PDF outside the family; if it passes, the concern is resolved and the central claim stands. If it fails, the central claim must be weakened to 'the lattice data are consistent with the best four-parameter fit of this form.' Either way, the current text overstates the model-independence of the 'decoding'.","tokens_in":11907,"tokens_out":9500,"duration_ms":107131,"concrete_test":"Misspecification coverage test: construct a synthetic truth PDF from a strictly more flexible family (e.g., a multi-spline or neural-net form consistent with NNPDF4.0 in the 0.2-0.7 x-range, but not representable by Eq. (3)), compute its RpITDs with Eq. (2), feed them through the trained VAIM, and check whether the truth lies within the 68% VAIM uncertainty band at all x in 0.2-0.7. Repeat with 10-20 such truths spanning the global-fit spread. If the empirical coverage is significantly below 68% (e.g., <50%), the quoted uncertainty band omits the dominant model discrepancy, and the central claim of consistency with global fits is not robust to the ansatz choice. A secondary check: refit the lattice RpITDs directly with Eq.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the VAIM-decoded gluon PDF agrees with global fits within uncertainties. This conclusion is load-bearing on the assumption that the true xg(x,2 GeV)/<x>g lies in the four-parameter family of Eq. (3). Because the training set is generated exclusively from Eq. (3) (Sec. 3.2, 'Training Data'), the decoder is structurally incapable of producing a PDF outside that family; the inverse map is a learned fit within a fixed functional form. The closure test (Fig. 2) is internal: it checks only that the predicted PDFs regenerate the lattice RpITDs through the same forward model Eq. (2) used to create the training data, so it cannot detect misspecification of Eq. (3). The paper's Sec. 4.1 explicitly lists 'functional parameterization choice' as an unquantified systematic. If the true gluon density has features not captured by x^alpha(1-x)^beta(1+gamma sqrt(x)+delta x)/norm in the 0.2<=x<=0.7 window, the posterior will be biased and the nominal band will undercover the truth, even though the band may still overlap smooth global fits because the prior is broad and the global fits themselves have sizeable uncertainties. Thus the external agreement, while encouraging, does not by itself establish that the lattice data have been 'decoded' rather than fitted to a predetermined form.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a variational autoencoder inverse mapper (VAIM) trained on synthetic reduced pseudo-Ioffe-time distributions (RpITDs) generated from a four-parameter gluon PDF ansatz (Eq. (3)) and the gluon pseudo-PDF matching relation (Eq. (2)). After training, the model is fed lattice RpITDs from the MSULat ensemble (a≈0.09 fm, Mπ≈310 MeV) and produces an ensemble of gluon PDF parameters. The resulting normalized gluon PDF xg(x,µ=2 GeV)/<x>g is compared visually with five global fits (CJ22, CT18, JAM22, MSHT20, NNPDF4.0) and claimed to be consistent within uncertainties, especially for 0.2≲x≲0.7. A closure test shows the predicted PDFs regenerate the lattice RpITDs within 2σ, and a Pearson correlation analysis is used to interpret the latent space.","tokens_in":12205,"tokens_out":6488,"duration_ms":65591,"significance":"If the extraction is unbiased, the paper demonstrates a new generative-AI bridge between lattice QCD and phenomenological PDFs, with potential to incorporate lattice data into global analyses. The method is transparent: the training-data generation is explicit, the architecture is described in detail, and the Monte Carlo ensembling for uncertainty propagation is clearly outlined. The paper is honest about unquantified systematics, explicitly listing functional parameterization choice, jackknife-to-MC conversion, and aleatoric/epistemic separation as future work in Sec. 4.1. However, the significance is currently limited by the absence of a misspecification test for the parameterization family and by the lack of any quantitative goodness-of-fit measure in the comparison with global fits.","major_comments":[{"comment":"The central claim of consistency with global fits is conditioned on the four-parameter ansatz of Eq. (3), because both the training data and the decoder output are confined to this family. The closure test in Fig. 2 only checks that the predicted PDFs regenerate the lattice RpITDs through the same forward model Eq. (2) used to create the training set, so it cannot detect whether the true gluon PDF lies outside the family. Since Sec. 4.1 explicitly lists \"functional parameterization choice\" as an unquantified systematic, the paper should either demonstrate that the global-fit PDFs (e.g., CJ22, NNPDF4.0) can be represented by Eq. (3) within the training ranges, or repeat the extraction with a more flexible basis (e.g., a neural-network PDF) and show that the result is stable. Without such a test, the agreement in Fig. 1 may be an artifact of the prior family rather than evidence that the lattice data have been decoded.","section":"3.2, Eq. (3), Fig. 2"},{"comment":"The claim that the VAIM result is \"consistent\" with CJ22, CT18, JAM22, MSHT20, and NNPDF4.0 is supported only by visual inspection of Fig. 1. Because the five global fits have different uncertainty definitions, a quantitative measure is needed, such as the chi-squared or average |pull| per x-bin between the VAIM band and each fit over the stated 0.2≲x≲0.7 range. Please provide these numbers, including the treatment of correlated uncertainties, or at minimum a table of pointwise deviations. This is load-bearing because \"within uncertainties\" is the paper's main conclusion.","section":"2, Fig. 1"},{"comment":"The jackknife-to-Gaussian conversion in Eq. (4) is not standard as written. The jackknife variance of the mean is (N−1)/N times the sum of squared deviations of the jackknife samples; if sigma^2_JK denotes this quantity, multiplying by N−1 overestimates the variance by a factor of order N, whereas if sigma^2_JK denotes the sum of squared deviations, the correct scaling would be (N−1)/N, not (N−1). Please clarify the definition of sigma^2_JK and justify the factor, since the width of the VAIM band directly enters the consistency claim.","section":"4.1, Eq. (4)"}],"minor_comments":[{"comment":"The statement that the average deviation is below 1σ and the maximum within 2σ needs a precise definition of σ (lattice statistical uncertainty only?) and the number of points; please provide the numerical values.","section":"2, Fig. 2"},{"comment":"The constraint 1+γ+δ>0 does not guarantee positivity of the factor 1+γ√x+δx for all x∈[0,1]; since the PDF must be positive, either enforce a stronger condition or verify positivity of all generated PDFs.","section":"3.2, Eq. (3) and decoder description"},{"comment":"The acronym is typeset as \"V AIM\" with a space in the abstract and elsewhere; this should be corrected to \"VAIM\".","section":"Abstract"},{"comment":"The claim of \"first decoding\" should be qualified to \"first VAIM-based decoding\" given previous ML-based gluon PDF extractions in Refs. [31,32].","section":"1, Introduction"},{"comment":"The uncertainty band is effectively statistical-only; this should be stated in the caption of Fig. 1 to avoid implying that the band includes known systematic uncertainties.","section":"4.1, Uncertainty quantification"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of Physics Letters B and the approach is interesting, but the central claim currently rests on an unquantified parameterization choice and a purely visual comparison with global fits. The authors should be asked to add a misspecification test and quantitative goodness-of-fit measures; with those additions the paper could become acceptable. No concerns about attribution or integrity."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a legitimate but narrow methods paper. The genuinely new piece is applying the VAIM to lattice gluon RpITDs — prior VAIM work was on quark PDFs in Mellin space and on Compton form factors. The authors describe the architecture, training data, and uncertainty propagation in enough detail that the work is reproducible in principle, and they are candid about the biggest limitation: the parameterization choice is listed as an unquantified systematic in Sec. 4.1.\n\nWhat the paper does well: the closure test in Fig. 2 is a useful sanity check, showing the decoded PDFs regenerate the lattice RpITDs within about 1σ, and the Pearson correlation analysis in Fig. 3 is a nice addition that gives some visibility into which matrix elements drive the result. The comparison to five global fits is encouraging, and the paper does not oversell the agreement — it says \"consistent within uncertainties,\" which is accurate at a visual level.\n\nThe soft spot is load-bearing: the VAIM is trained exclusively on PDFs of the form x^α(1-x)^β(1+γ√x+δx)/norm, so the decoder cannot represent anything outside that family. The closure test is internal — it validates that the model can invert the training family, not that the family contains the true gluon PDF. If the true PDF has features outside this form in the 0.2≲x≲0.7 window, the result is biased even if the band overlaps global fits. This is not a hidden flaw; the authors acknowledge it, but it means the title's \"decoding\" overpromises. The comparison to global fits is also visual; there is no χ² or similar quantitative measure. Other systematics, such as the RpITD range cut, the Jackknife-to-MC conversion, and the neglected quark-gluon mixing, are reasonable but unquantified.\n\nI do not see a fabricated or circular claim in the strong sense: the agreement with global fits is external and meaningful. But the extraction is a fit to a predetermined functional form, not an ab initio decoding. For a proof-of-principle, that is acceptable; for a headline result, it needs a misspecification test.\n\nWho this is for: lattice QCD practitioners and people working on ML-based inverse problems in hadron structure. It deserves a serious referee — the method is a new bridge and the execution is mostly clean — but I would ask the authors to add a quantitative goodness-of-fit to the global fits and at least one robustness check against a broader parameterization or a nonparametric baseline.","headline":"A narrow but honest VAIM demonstration for gluon PDFs; the four-parameter ansatz means the result is a fit within a family, not a full decoding, but the paper deserves serious review as a proof-of-principle.","tokens_in":12776,"tokens_out":1651,"would_cite":false,"duration_ms":19371,"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":"A generative inverse mapper decodes the proton's gluon density from lattice QCD and matches global fits within uncertainties.","keywords":["gluon PDF","lattice QCD","variational autoencoder inverse mapper","pseudo-PDF method","Ioffe-time distribution","inverse problem","generative machine learning","proton structure"],"falsifier":"Generate synthetic RpITDs from a gluon PDF deliberately outside the four-parameter family, such as one with an additional bump at moderate $x$, feed them to the trained VAIM, and check whether the reconstructed RpITDs and predicted PDF reproduce the input within the reported uncertainties; a significant discrepancy would localize the bias in the assumed parameterization rather than the network.","tokens_in":11688,"feed_emoji":"⚛️","tokens_out":9446,"duration_ms":87878,"temperature":0.7,"pith_summary":"The paper is trying to establish that generative machine learning can solve the inverse problem of extracting the proton's gluon parton distribution function from lattice QCD data. It trains a variational autoencoder inverse mapper to convert reduced pseudo-Ioffe-time distributions (RpITDs), computed on a lattice ensemble with spacing about 0.09 fm and pion mass about 310 MeV, into the parameters of a normalized gluon PDF. The decoded distribution is then compared with five phenomenological global fits and is consistent with them within uncertainties, especially for momentum fractions $0.2 \\lesssim x \\lesssim 0.7$ where the lattice signals are strongest. If correct, this opens a route for lattice information to enter global analyses directly through a generative, data-driven framework rather than through hand-picked fit forms.","feed_headline":"Machine learning decodes the proton's gluon density from lattice QCD","feed_subtitle":"A generative network turns lattice Ioffe-time data into a gluon PDF matching global fits in the intermediate-x region.","key_machinery":"The central object is the variational autoencoder inverse mapper (VAIM), a variational autoencoder whose latent space is split into a 35-dimensional observable channel and a 256-dimensional generative channel. The observable channel carries the reduced pseudo-Ioffe-time distributions (RpITDs), double ratios of Wilson-line matrix elements normalized to one at zero Ioffe time. Training data are synthetic: PDF parameters $\\alpha,\\beta,\\gamma,\\delta$ are drawn from uniform ranges, the four-parameter ansatz $xg(x)/\\langle x\\rangle_g = x^\\alpha(1-x)^\\beta(1+\\gamma\\sqrt{x}+\\delta x)/N$ is formed, and RpITDs are generated through the gluon pseudo-PDF matching relation $\\mathcal{M}(\\nu,z^2)=\\int_0^1 dx\\,[xg(x)/\\langle x\\rangle_g]\\,R_{gg}(x\\nu,z^2\\mu^2)$, with the quark-gluon mixing kernel neglected. After training, lattice RpITDs enter the observable channel and the decoder outputs an ensemble of parameter sets sampled from the latent posterior; that ensemble defines the PDF uncertainty. The network uses four residual blocks with skip connections per encoder and decoder, and its three output heads enforce the physical ranges of $\\alpha$, $\\beta$, and $1+\\gamma+\\delta>0$.","core_discovery":"The central discovery claimed is that a trained VAIM, supplied with the 35 lattice RpITDs, produces a normalized gluon PDF $xg(x,\\mu=2\\,\\mathrm{GeV})/\\langle x\\rangle_g$ whose central value lies within the uncertainty bands of the major phenomenological global fits over most of the $x$-range, with the best constraints at $0.2 \\lesssim x \\lesssim 0.7$. The claim is supported by a closure test: RpITDs reconstructed from the predicted PDF parameters track the lattice input with an average point-by-point deviation below $1\\sigma$ and a maximum deviation within $2\\sigma$. The paper further claims that Pearson correlations between the 35 input RpITDs and the $x$-dependent PDF reveal a learned latent representation rather than trivial memorization, and that the uncertainty band widens at small $x$ where the lattice constraints are weak.","pith_inferences":["Beyond the paper, replacing the four-parameter ansatz with a wider nonparametric basis would turn the stated unquantified parameterization systematic into a measurable quantity: how much of the decoded shape comes from the assumed functional form versus the lattice data.","Beyond the paper, the same observable-to-parton inverse mapping could be applied wherever a perturbative matching relation exists, including polarized gluon distributions, quark helicity distributions, or fragmentation functions.","Beyond the paper, a natural endgame is to feed lattice RpITDs directly into a global fit through a covariance-aware likelihood using the trained VAIM as a fast surrogate, rather than comparing decoded PDFs after the fact.","Beyond the paper, the small-$x$ widening of the uncertainty band is a concrete prediction that future high-energy electron-ion collider data, which probe that region, could shrink or challenge."],"forward_implications":["Lattice RpITDs alone, without any experimental input, can produce a gluon PDF whose shape agrees with global fits in the intermediate-to-large-$x$ window, establishing lattice QCD as an independent constraint on the gluon.","The VAIM uncertainty band is comparable to phenomenological fits where the lattice data are informative and expands where they are not, so the method reports where the gluon is actually known.","The correlation analysis identifies which combinations of Wilson-line length and hadron momentum carry independent information, which can guide future lattice ensembles toward the most constraining kinematic points.","Once epistemic uncertainties such as parameterization choice are quantified, improvements in lattice statistics should translate directly into shrinking $x$-dependent PDF uncertainties."],"supporting_citations":[{"why":"Supplies the lattice RpITD matrix elements that are fed into the trained VAIM.","marker":"[24]"},{"why":"Defines the gluon pseudo-PDF matching kernel used to generate training RpITDs from PDF parameters.","marker":"[18]"},{"why":"Introduces the variational autoencoder inverse mapper architecture used throughout.","marker":"[35]"},{"why":"Establishes the VAIM approach for PDFs and supplies the four-parameter functional form.","marker":"[37]"},{"why":"Global fit used as a comparison baseline for the decoded gluon PDF.","marker":"[1]"},{"why":"Global fit used as a comparison baseline for the decoded gluon PDF.","marker":"[2]"},{"why":"Global fit used as a comparison baseline for the decoded gluon PDF.","marker":"[3]"},{"why":"Global fit used as a comparison baseline for the decoded gluon PDF.","marker":"[4]"},{"why":"Global fit used as a comparison baseline for the decoded gluon PDF.","marker":"[5]"}],"fun_headline_variants":["Generative AI decodes proton's gluon density from lattice QCD","Variational autoencoder maps lattice data to gluon PDF","Lattice QCD meets AI: gluon PDF matches global fits","Closure test: AI-produced gluon PDF consistent with lattice"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption, which the paper itself flags as an unquantified systematic, is that the true gluon PDF lies inside the four-parameter family $x^\\alpha(1-x)^\\beta(1+\\gamma\\sqrt{x}+\\delta x)/N$, because the network only sees training data generated from that form and can never return a PDF outside it.","fun_headline_variants_meta":{"raw":{"variants":["Generative AI decodes proton's gluon density from lattice QCD","Variational autoencoder maps lattice data to gluon PDF","Lattice QCD meets AI: gluon PDF matches global fits","Closure test: AI-produced gluon PDF consistent with lattice"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000279,"raw_usage":{"total_tokens":1623,"prompt_tokens":878,"completion_tokens":745,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":494,"completion_tokens_details":{"reasoning_tokens":672}},"tokens_in":494,"tokens_out":745,"duration_ms":7922,"temperature":1.0,"reasoning_tokens":672,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:39:27.215039+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate synthetic RpITDs from a gluon PDF deliberately outside the four-parameter family, such as one with an additional bump at moderate $x$, feed them to the trained VAIM, and check whether the reconstructed RpITDs and predicted PDF reproduce the input within the reported uncertainties; a significant discrepancy would localize the bias in the assumed parameterization rather than the network.","supporting_citations":[],"review_version":1}