{"id":"ecce67c6-928e-4fec-8de4-5dc12047f351","arxiv_id":"1909.01359","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A least-squares GAN trained on Lund jet plane images reproduces the simulated jet substructure distribution to within a few percent, and a CycleGAN maps between jet categories such as parton-level vs detector-level or QCD vs W jets.","lead":"Neural networks are trained to generate and translate images of the radiation patterns inside particle jets. The generated images match the average simulated jet structure within a few percent, and a cycle-consistent network can convert one jet category into another.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Independent pixel-wise sampling from averaged Lund images discards inter-pixel correlations, so the few-percent claim supports only the marginal density, not event-level jet substructure; the paper's Section 4 concession makes this explicit.","rationale":"The reader's weakest assumption is exactly the load-bearing point: the generative pipeline's only mechanism for producing discrete jets is independent pixel sampling from averaged images, which cannot encode jet-defining correlations. The paper is honest about this in Section 4, but the limitation directly affects the scope of the abstract's claims. The demonstrated few-percent agreement applies to the average Lund density; it does not establish that generated images are faithful individual jets, and the CycleJet mapping is validated on averaged images rather than per-event transformations. A two-point correlation or pixel-covariance test on sampled generated images would settle whether the loss of correlations is numerically important. If it is, the central claim should be read narrowly as a statement about the average two-dimensional distribution, which is what the paper literally claims. This supports the reader's conditional verdict without requiring a change: the paper's primary quantitative claim is credible, but its broader event-level and data-augmentation implications are not yet demonstrated.","tokens_in":12973,"tokens_out":5818,"duration_ms":65049,"concrete_test":"Generate 100k event images from the trained gLund model using the Section 2.3 sampling step (independent Bernoulli draws per pixel from the generated average image). On a held-out reference sample of 100k Pythia+Delphes Lund images, compute the pixel-pair covariance C(i,j)=<p_i p_j>-<p_i><p_j> and the binned two-point density over pixel separations (Delta ln kt, Delta ln 1/Delta_ab). Compare these with the same quantities for the generated sampled images, including bootstrap uncertainties. If the normalized difference exceeds a few percent, the model reproduces only the marginal density and the independent-sampling assumption fails; if they match, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central numerical claim is agreement to within a few percent, but the evidence establishes this for the average Lund-plane density, not for individual jets. In Section 2.3, the LSGAN is trained on averages of navg=32 sparse images, and each generated image is converted into an 'event' by sampling every pixel independently with probability equal to the averaged pixel value. Independent Bernoulli sampling preserves the per-pixel marginal rate by construction, but it cannot reproduce correlations between emissions that are well separated in the Lund plane. Real Lund images have structured activation patterns inherited from the clustering history; independently sampled images will instead have approximately Poisson-binomial counts and no inter-pixel correlation structure. The paper's own Section 4 concedes that the preprocessing 'looses information on correlations between emissions at wide angular and transverse momentum separation' and that 'it is difficult to evaluate or improve the formal logarithmic accuracy of the generated samples.' The CycleJet validation in Eq. (11) also compares averaged reference and transformed images, so the claim that one can 'retroactively change simulation settings' on an existing event sample is likewise supported only at the level of averaged densities. Therefore the few-percent agreement constrains the one-point density, not the joint distribution that defines event-level jet substructure.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces two neural-network-based tools for Lund jet images. The first, gLund, is an LSGAN trained on 500k Pythia+Delphes QCD jets. To handle sparsity, training images are produced by averaging navg=32 individual Lund images, and generated images are converted back to discrete images by independent per-pixel Bernoulli sampling. The paper reports agreement with the reference sample within a few percent in the bulk of the Lund plane, and compares gLund with a VAE and a WGAN-GP on density slices and derived observables. The second tool, CycleJet, is a CycleGAN trained to map between parton-level and detector-level images and between QCD and boosted-W jets, with validation based on comparisons of average images. Code and data are released openly.","tokens_in":13243,"tokens_out":3425,"duration_ms":37319,"significance":"If the reported accuracy extends to event-level jet substructure, the paper would provide a fast, storage-efficient alternative to full Monte Carlo simulation and a new way to reinterpret event samples. The strengths of the work are its open-source implementation, reproducible data release, systematic hyperparameter optimization, and a useful benchmark comparison among LSGAN, WGAN-GP, and VAE. However, as presented, the evidence supports reproduction of the one-point Lund-plane density, not the joint distribution of activated pixels that defines individual jet radiation patterns. The authors explicitly concede in Section 4 that their preprocessing loses wide-separation correlations and that the formal logarithmic accuracy of generated samples is hard to evaluate. The paper is therefore a useful contribution to fast approximate simulation of averaged jet densities, but its central event-level claims need either additional validation or a careful reframing.","major_comments":[{"comment":"The stochastic generation step treats each pixel as an independent Bernoulli random variable with probability equal to the averaged pixel value. This preserves per-pixel marginal rates but cannot reproduce correlations between emissions that are well separated in the Lund plane; real Lund images have structured activation patterns inherited from the clustering history. The few-percent agreement shown in Figures 7 through 10 therefore constrains only the average density, not the joint distribution over individual jet images. The paper's own Section 4 concession, that the preprocessing 'looses information on correlations between emissions at wide angular and transverse momentum separation' and that it is 'difficult to evaluate or improve the formal logarithmic accuracy of the generated samples,' confirms this limitation. The authors should either validate event-level observables sensitive to inter-pixel correlations or explicitly restrict the central claim to reproduction of the averaged Lund-plane density.","section":"Section 2.3 and Section 4"},{"comment":"No train/test split is described anywhere in the experimental setup. Hyperparameter selection is performed using the loss Lh in Eq. (3), which compares generated images with the reference preprocessed images, and the final validation in Figures 7-10 again compares generated samples against the same reference sample used for training. Without an independent test set and without statistical uncertainties on the ratio plots, the reported 'few percent' accuracy may reflect memorization or overfitting rather than generalization. The authors should specify the split between training, validation, and test samples, and report quantitative errors with uncertainties for Figures 7-10.","section":"Sections 2.4 and 2.5"},{"comment":"The CycleJet validation is based on the loss Lh = ||RA - P_{B->A}|| + ||RB - P_{A->B}||, which compares only the average reference images with the average transformed images. Consequently, the claim that one can 'retroactively change simulation settings or the underlying process on an existing sample' is demonstrated only at the level of mean densities. Individual event mappings may not preserve the correlations required for physics applications. The authors should add per-event validation metrics, such as distributions of pixel counts, two-point correlations, or downstream substructure observables evaluated on individual transformed events, or substantially weaken the corresponding claims in the abstract and Section 3.","section":"Section 3.2, Eq. (11)"},{"comment":"The abstract states that the model 'retrieves the underlying two-dimensional distribution to within a few percent,' but Figure 7(c) shows deviations well beyond a few percent near the boundaries of the Lund plane, and the 'bulk' region is never quantitatively defined. The authors should define the region over which the few-percent claim holds and report a concrete aggregate metric, e.g., a maximum or average relative deviation over that region, rather than relying on visual inspection of ratio plots.","section":"Figure 7(c) and abstract"}],"minor_comments":[{"comment":"The sentence 'loosing information on correlations' should read 'losing information on correlations.'","section":"Section 4"},{"comment":"The caption of Figure 12 refers to 'delphes-level sample'; the capitalization should be 'Delphes-level' for consistency with the rest of the text.","section":"Section 3.2"},{"comment":"The quantities z(i) and Delta(i) in Eq. (6) are not defined in the text; please define them explicitly or refer to the original mMDT definition.","section":"Equation (6)"},{"comment":"The hyperparameter tables would be more reproducible if they included the random seed and the number of evaluation samples used for each reported metric.","section":"Appendix A"},{"comment":"Reference [40] is incomplete: it lists only '(2019), 1909.04451' without authors or title. Please complete the bibliographic entry.","section":"References"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know first: this is a solid, honest application paper, not a big physics or ML conceptual leap. What it does that is genuinely new is to bring GANs to the Lund jet plane representation – no prior GAN/VAE paper for LHC simulation used this image basis – and to show a CycleGAN mapping between jet categories (parton-to-detector, QCD-to-W). That mapping for retroactively changing simulation settings is a real and potentially useful idea. The authors also ship code and data, which makes the work reproducible and worth engaging with on that basis alone.\n\nWhat the paper does well: the density comparisons in Figs. 7–9 support the few-percent claim for the bulk of the Lund plane, and the comparison against VAE and WGAN-GP is a fair benchmark. Fig. 10 goes beyond average densities to show event-level observables like soft-drop multiplicity and groomed mass, and those look decent. The authors openly state in Section 4 that the averaging-plus-independent-sampling preprocessing loses correlations between emissions at wide separation, and that the formal logarithmic accuracy is difficult to evaluate. That concession is in the paper itself, so the stress-test concern is real but not hidden.\n\nThe soft spots are in proportion. First, there is no stated train/validation/test split; hyperopt selects the best model on a loss that compares to the reference sample, so the reported accuracy is partly tuned, not an independent test. Second, the few-percent claim is for average densities; the independent pixel-wise sampling from averaged images indeed cannot reproduce inter-pixel correlations, and the CycleJet validation in Eq. (11) compares average images, not event-level mapping quality. The paper would be stronger with explicit separation, error bars on the ratio plots, and a quantitative check of CycleJet on event-level observables after mapping. These are fixable weaknesses, not fatal ones.\n\nWho is this for: anyone working on fast simulation or data augmentation for jet substructure will get value from this. It deserves a serious referee; the experiments are clear enough to check and the code is available. I would not cite it next to a physics result, but I would cite it as a methods reference. Take the few-percent wording as applying to averages, and you have a fair and useful paper.","headline":"A genuinely useful and honest paper: gLund and CycleJet bring GANs to the Lund plane with open code and data, and the few-percent claim holds for average densities, with the authors upfront that event-level correlations are a known limitation.","tokens_in":13758,"tokens_out":1223,"would_cite":true,"duration_ms":14623,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["12.38.-t","07.05.Mh"],"model":"deepseek-v4-flash","headline":"This paper claims that a neural generative model trained on Lund jet images reproduces the two-dimensional radiation-pattern density within a few percent, and that a cycle-consistent network can map images between jet categories.","keywords":["generative adversarial networks","Lund jet plane","jet substructure","cycle-consistent adversarial networks","fast simulation","data augmentation","deep learning in particle physics"],"falsifier":"Compute the joint probability that two pixels separated widely in $\\ln k_t$ or $\\ln(1/\\Delta_{ab})$ are simultaneously activated, for both generated and reference samples; if the single-pixel densities agree to a few percent but the joint probabilities disagree by more than the statistical uncertainty, the model's independent-pixel sampling is falsified as a description of individual jets.","tokens_in":12760,"feed_emoji":"⚛️","tokens_out":10101,"duration_ms":88781,"temperature":0.7,"pith_summary":"This paper is trying to establish that neural generative models can serve as fast, accurate stand-ins for Monte Carlo simulation of jet substructure. The authors encode each jet's radiation pattern as a Lund jet plane image, train a Least-Squares GAN on probabilistically averaged images, and report that the generated samples reproduce the average reference density to within a few percent in the bulk of the image. They also train a cycle-consistent adversarial network to translate Lund images between jet categories, showing that parton-level images can be promoted to detector-level images and that QCD jets can be mapped to boosted W jets (and back). If these claims hold, event generation for jet-substructure studies could be replaced by near-instant neural inference, and existing simulated samples could be repurposed to new processes or simulation settings without regeneration.","feed_headline":"Generative networks reproduce jet radiation patterns to a few percent","feed_subtitle":"A second model rewrites existing samples to new processes or simulation settings, cutting simulation cost.","key_machinery":"The central object is the primary Lund jet plane: reclustering a jet's constituents with the Cambridge/Aachen algorithm and stepping down the hardest branch assigns each emission a pixel at coordinates ($\\ln(1/\\Delta_{ab})$, $\\ln k_t$), producing a 24 by 24 binary image of the jet's radiation pattern. The paper's training trick is to average $n_{\\rm avg}$ such sparse images and treat each averaged pixel as an activation probability, which makes the discrete images learnable by a Least-Squares GAN, a generator/discriminator pair with a quadratic loss and a minibatch-discrimination layer. For the mapping half, a cycle-consistent adversarial network learns paired forward and inverse translations between two jet-image domains, enforcing that translating an image to the other domain and back recovers the original.","core_discovery":"The central discovery claimed is twofold. First, with a stochastic probabilistic interpretation of Lund images, averaging $n_{\\rm avg}$ input images and sampling each pixel as a probability, an LSGAN with a minibatch-discrimination layer and ZCA whitening reproduces the average Lund jet plane density of detector-level QCD jets to within 3 to 5 percent in the bulk region, matching a WGAN-GP and clearly outperforming a VAE whose accuracy saturates around 20 percent because of posterior collapse. Second, a CycleGAN-based model, CycleJet, learns unpaired translations between parton-level and detector-level Lund images and between QCD-jet and W-jet images, with the translated average images matching their target domains well enough to be used for retroactive changes of simulation settings or underlying process.","pith_inferences":["The paper's per-pixel independence assumption implies that event-level observables sensitive to correlations between widely separated emissions are not guaranteed by the few-percent density agreement; testing such observables would probe the boundary of the claim.","The same cycle-consistency recipe could be applied to other domain pairs, such as quark- versus gluon-initiated jets or different pileup conditions, provided the two domains share the Lund-plane support; the paper's two examples motivate but do not establish that generality.","Because CycleJet is trained on averages and then samples pixels, applying it to real data would require the preprocessing to be invertible or the mapping to be trained directly on individual events.","If the few-percent density accuracy persists at higher jet transverse momenta or in rarer kinematic corners, fast inference could be combined with traditional generators in a hybrid scheme, using the network for bulk phase space and the generator for tails."],"forward_implications":["Jet-substructure samples for the bulk of phase space can be produced by fast network inference instead of full Monte Carlo generation, reducing simulation time and storage.","Parton-level samples can be upgraded to detector-level Lund images with a learned mapping, avoiding a full detector-simulation pass.","An existing QCD dijet sample can be reinterpreted as a sample of boosted hadronically decaying W jets (and vice versa), enabling process remapping without new event generation.","Among the tested alternatives, the LSGAN is the practical choice: the VAE is limited to roughly 20 percent accuracy by posterior collapse, while the WGAN-GP matches distributions but produces less realistic individual images.","Groomed observables reconstructed from generated images, such as soft-drop multiplicity and jet mass, follow the reference distributions, supporting the use of these samples in physics analyses."],"supporting_citations":[{"why":"Defines the primary Lund jet plane whose $\\ln(1/\\Delta_{ab})$ versus $\\ln k_t$ coordinates turn a jet's clustering history into the images used throughout.","marker":"[22]"},{"why":"Introduces the least-squares GAN objective that stabilises training of the gLund generator.","marker":"[42]"},{"why":"Generates the 500k-jet dijet reference sample used for training and all fidelity comparisons.","marker":"[43]"},{"why":"Provides the fast detector simulation and particle-flow reconstruction that define the detector-level target images.","marker":"[44]"},{"why":"Supplies the Wasserstein GAN baseline whose distribution accuracy is compared with gLund.","marker":"[45]"},{"why":"Introduces the cycle-consistency adversarial framework that CycleJet uses to translate between jet categories.","marker":"[46]"},{"why":"Defines the variational autoencoder baseline, whose limited accuracy of around 20 percent frames the improvement claimed for gLund.","marker":"[30]"},{"why":"Supplies the minibatch-discrimination layer used to prevent generator collapse in the LSGAN.","marker":"[54]"},{"why":"Provides the automated hyperparameter search procedure that selected the reported model configurations.","marker":"[56]"}],"fun_headline_variants":["GANs hit few-percent accuracy on jet radiation patterns","CycleGAN rewrites jet simulation settings retroactively","Lund plane GANs match detector-level jets to 3%","AI generates jet images with few-percent fidelity","Generative networks translate between jet categories"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that averaging $n_{\\rm avg}$ images and sampling each pixel independently as a probability gives a faithful model of individual jet radiation patterns, not just of the average density.","fun_headline_variants_meta":{"raw":{"variants":["GANs hit few-percent accuracy on jet radiation patterns","CycleGAN rewrites jet simulation settings retroactively","Lund plane GANs match detector-level jets to 3%","AI generates jet images with few-percent fidelity","Generative networks translate between jet categories"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000139,"raw_usage":{"total_tokens":1087,"prompt_tokens":803,"completion_tokens":284,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":419,"completion_tokens_details":{"reasoning_tokens":222}},"tokens_in":419,"tokens_out":284,"duration_ms":3179,"temperature":1.0,"reasoning_tokens":222,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:20:07.272268+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the joint probability that two pixels separated widely in $\\ln k_t$ or $\\ln(1/\\Delta_{ab})$ are simultaneously activated, for both generated and reference samples; if the single-pixel densities agree to a few percent but the joint probabilities disagree by more than the statistical uncertainty, the model's independent-pixel sampling is falsified as a description of individual jets.","supporting_citations":[{"cited_title":"Arjovsky, S","cited_arxiv_id":null,"evidence_quote":"Supplies the Wasserstein GAN baseline whose distribution accuracy is compared with gLund."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the cycle-consistency adversarial framework that CycleJet uses to translate between jet categories."},{"cited_title":"Bergstra, D","cited_arxiv_id":null,"evidence_quote":"Provides the automated hyperparameter search procedure that selected the reported model configurations."}],"review_version":1}