{"id":"ad194065-052b-4221-bff5-d8b2b77e0859","arxiv_id":"2412.02468","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A two-stage neural network that separates single and double resonant top production improves the expected 95% CL limit on the right-handed Wtb coupling f_R^V from 0.21 to 0.17.","lead":"The paper trains two stages of neural networks to separate Standard Model and anomalous right-handed signals in the Wtb vertex for top quark production, and reports that splitting single and double resonant events tightens the expected constraint on the anomalous coupling from |f_R^V| < 0.21 to |f_R^V| < 0.17. The result is a Monte Carlo demonstration only, with no experimental data or released code.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed 0.21→0.17 improvement conflates the NN phase-space split with a change in the assumed f_R^V signal-rate scaling, so the comparison is not apples-to-apples and the f^4 scaling is unvalidated.","rationale":"The reader's weakest assumption points to the signal-rate scaling, and I agree and sharpen it. The central claim is about an improvement, and an improvement claim requires comparing like with like. Here the split analysis differs from the unsplit analysis not only in using separate NNs but also in using a different functional dependence on the parameter of interest. That confound directly undermines the causal attribution that splitting leads to stricter constraints. The f^4 scaling is not derived in the text; it is imported from [22] with no validation against the actual MC event sets, which is especially problematic because the full process is a mixture of single- and double-resonant diagrams. The 0.9 threshold adds another degree of freedom. Until the split and unsplit analyses are run with identical signal scaling, the 0.21 to 0.17 gain is not established. The statistical derivation's internal errors and missing yield inputs further support a conditional verdict, but the scaling confound is the most directly load-bearing issue for the central claim.","tokens_in":6957,"tokens_out":14413,"duration_ms":156523,"concrete_test":"Test the scaling confound directly: recompute the unsplit and split limits from the same MC histograms with a common signal-rate model, e.g., set μ = f_R^V^2 in every region, and also try a mixed polynomial μ = a f_R^V^2 + b f_R^V^4 for the unsplit sample, keeping the first- and second-level NN classifiers fixed. If the split limit is no longer strictly better than the unsplit limit, the claimed improvement is due to the f_R^V^4 assumption, not to the NN separation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claim is that phase-space splitting of tWb events into single- and double-resonant classes improves the expected 95% CL limit on |f_R^V| from 0.21 to 0.17. The comparison, however, changes two variables at once: the event classification and the assumed signal-rate scaling. The unsplit limit uses μ = f_R^V^2 for the full tWb sample; the split limit uses μ = f_R^V^2 for the t-tbar region and μ = f_R^V^4 for the single-resonant region. Since the full tWb sample contains both single- and double-resonant diagrams, its cross section is not purely quadratic in f_R^V; and the f^4 assignment for the single-resonant region is asserted from ref. [22] without derivation or validation in this paper. The 0.9 first-level NN threshold defining the split is also arbitrary. Thus the reported improvement may be an artifact of the assumed signal parameterization rather than of the NN separation itself. Additionally, the printed statistical derivation is internally inconsistent (Eq. (8) with the defined S1 gives an imaginary Z; Eq. (5) has incorrect derivatives) and no s_i,b_i yields or luminosities are given, so the numerical limits are not reproducible as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a neural-network-based strategy for separating single- and double-resonant top quark production in the tWb final state, and then applying dedicated second-level classifiers to distinguish Standard Model left-handed from anomalous right-handed vector Wtb couplings in each phase-space class. The authors report that this splitting improves the expected 95% CL constraint on the anomalous right-handed vector coupling from |f_R^V| < 0.21 (unsplit) to |f_R^V| < 0.17 (split), using a simplified likelihood-based statistical model with a 20% systematic uncertainty. The central quantitative claim is therefore that phase-space separation of the two resonance topologies increases sensitivity to this anomalous coupling.","tokens_in":7308,"tokens_out":4043,"duration_ms":41693,"significance":"If the claimed improvement is genuine, the approach would be useful because it avoids the common practice of removing part of the diagram set to separate single- and double-resonant contributions, instead using neural networks to classify events and treating the interference as smeared across classes. The paper also uses a complete gauge-invariant set of diagrams and provides Monte Carlo distributions showing that the second-level classifiers achieve visible separation between left- and right-handed vector operator hypotheses. However, the statistical derivation in Section 4 is not internally consistent, the numerical limits are not reproducible from the given information, and the reported improvement may be an artifact of the assumed f_R^V scaling rather than of the neural-network split. The potential significance is therefore not yet established by the manuscript as written.","major_comments":[{"comment":"The statistical model is internally inconsistent. In Eq. (4) the Poisson term is printed as n_i · ln( (μ s_i + b_i)/n_i! ) − μ s_i − b_i, which is missing the exponent n_i in the numerator and is not a valid log-likelihood. Eq. (5) mixes i and j indices and the derivative with respect to b_j is incorrect: the control-region derivative should be m_j/b_j − τ, not m_j/(τ b_j) − τ. In Eq. (8), the significance Z is written as sqrt(2(S1 + S2)) using S1 and S2 from Eqs. (6)–(7), but with Asimov data n_i = b_i and b̂_i = b_i, S1 can be negative, so Z is not guaranteed real. Since the limits 0.21 and 0.17 are the central quantitative result, these formulas must be corrected and the calculation repeated.","section":"Section 4, Eqs. (4)–(8)"},{"comment":"The comparison between the unsplit and split cases changes two assumptions at once. The unsplit limit uses μ = f_R^V^2 for the full tWb sample, while the split limit uses μ = f_R^V^2 for the double-resonant region and μ = f_R^V^4 for the single-resonant region, with the f^4 scaling asserted from reference [22] without derivation or validation. The full tWb sample contains both single- and double-resonant diagrams, so its cross-section dependence on f_R^V is not necessarily purely quadratic; the reported 0.21 → 0.17 improvement could therefore be driven by the assumed signal parameterization rather than by the neural-network phase-space split. The authors should test the split with a consistent scaling for both samples or provide a derivation of the appropriate scaling of the single-resonant region.","section":"Section 4, paragraph after Eq. (8)"},{"comment":"The numerical limits are not reproducible as written. No yields s_i and b_i, no luminosity, no binning parameters N and M, and no value for the 20% systematic Δ are given. The quantities δ_i and τ are defined circularly: the text introduces δ_i = 1/sqrt(τ b_i) with τ = Δ^2 * Σ b_i, and later substitutes τ = 1/(δ_i^2 b_i). This leaves the normalization of the control region unspecified. The authors should provide the input histograms, binning, and either a public implementation or enough detail to reproduce the 0.21 and 0.17 numbers.","section":"Section 4, input values"},{"comment":"The split between the double-resonant and single-resonant classes is defined by a fixed threshold of 0.9 on the first-level neural-network discriminator, but this threshold is not justified or optimized. The sensitivity result will depend on this arbitrary cut, and the paper should either scan the threshold or at least demonstrate that the qualitative conclusion is stable within a reasonable range.","section":"Section 3, first-level NN threshold"}],"minor_comments":[{"comment":"Several axis labels in Fig. 1 are garbled, e.g., 'cos(D,l)_RF(l,0' and 'cos(l,u)_RF(l,0', which makes it impossible to identify the plotted observables. Please fix the labels.","section":"Figure 1"},{"comment":"The notation 'tWb' is used both for the full final state (including double-resonant diagrams) and for the single-resonant phase-space region in the sentence 'for tWb- μ = f_R^V^4'. This is confusing and should be disambiguated, e.g., by calling the single-resonant region 'tW' or 'single-resonant'.","section":"Section 4"},{"comment":"There are numerous typographical and readability issues, such as 'splitted' instead of 'split', 'a object' instead of 'an object', and the phrase 'Wtb vertex into a Wtb vertex'. A thorough language edit is needed.","section":"General"},{"comment":"Reference [15] is cited in the text as a source of high-level observables but appears to be about a different topic; please verify that the reference list matches the citations.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is a phenomenological study with a potentially interesting idea, but the statistical analysis in Section 4 is not sound as written and the central numerical claim is not reproducible. The issues are substantial but in principle fixable by redoing the statistical calculation, providing the inputs, and validating the signal-rate scaling. I would not reject outright, but the manuscript needs major revision before it can be considered for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe paper extends the authors' earlier work on separating single- and double-resonant tWb events with a neural network, adding a second level of classifiers that distinguish SM left-handed from anomalous right-handed vector Wtb couplings within each event class. That is a genuine, modest extension: the plots in Sec. 3 show the two classes separate well for both the ttbar-like and single-resonant phase space regions, and the authors are careful to use the full gauge-invariant diagram set rather than removing diagrams. The two-stage design is sensible and the MC training setup is described in enough detail to reproduce the shapes in principle.\n\nThe soft spots are concentrated in Sec. 4. The central comparison, 0.21 without splitting vs 0.17 with splitting, changes two things at once: the phase-space split and the assumed dependence of the signal rate on f_R^V. The unsplit limit uses mu = f_R^V^2 for the whole tWb sample, while the split analysis uses mu = f_R^V^2 for the ttbar region and mu = f_R^V^4 for the single-resonant region. If the full tWb sample is a mixture of both diagram classes, its cross section is not purely quadratic in f_R^V, so the two limits are not directly comparable. The f^4 scaling is asserted from ref [22] without derivation or cross-check here, and the 0.9 threshold on the first-level NN that defines the split is not justified. If the true scaling differs, the size of the gain—and perhaps its sign—would change.\n\nThe statistical model itself is hard to follow, and not just in the usual phenomenological shorthand. Some formulas appear garbled: Eq. (8) with the defined S1 would give an imaginary Z, the derivative in Eq. (5) does not look right, and quantities like delta_i and tau are introduced in a way that leaves the reader guessing. No yields, bin counts, or luminosity are given, so the quoted limits are not reproducible from the text. That is a real problem for a paper whose main result is a pair of numbers.\n\nWhat the paper does well is the phenomenological setup and the demonstration that operator-level separation within single/double resonant classes is feasible. That part earns attention. But the quantitative claim needs a clean statistical derivation, a validated or at least defended scaling assumption, and a matched comparison where only the splitting changes.\n\nFor a top-quark EFT audience, this is worth a serious referee, but heavy revision is required before the limits can be believed. I would not cite it in its current form.","headline":"The NN separation itself looks plausible, but the reported 0.21→0.17 sensitivity gain is confounded by an unvalidated change in signal-scaling assumptions and a statistical derivation that is not reproducible as written.","tokens_in":7814,"tokens_out":2639,"would_cite":false,"duration_ms":25059,"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":"Splitting top quark events by resonance class tightens the bound on an anomalous Wtb coupling from 0.21 to 0.17.","keywords":["top quark","Wtb vertex","anomalous couplings","effective field theory","neural network","single top production","top pair production","LHC"],"falsifier":"Compute the expected upper limit on f_R^V using a full event generator at several fixed values of f_R^V (for example 0, 0.1, 0.2, 0.3) without assuming a power-law scaling, and compare the single-resonant and double-resonant signal rates; if the single-resonant rate does not scale approximately as f_R^$V^{4}$, the reported 0.17 versus 0.21 comparison will not reproduce.","tokens_in":6827,"feed_emoji":"⚛️","tokens_out":5337,"duration_ms":54209,"temperature":0.7,"pith_summary":"The paper argues that sensitivity to an anomalous right-handed vector coupling at the Wtb vertex is improved by first splitting top quark production events into single-resonant and double-resonant classes with a neural network, then running a separate classifier inside each class. The two classes depend on the anomalous coupling differently, so analyzing them separately gives a tighter expected upper limit on the coupling magnitude |f_R^V|: below 0.17 with the split versus 0.21 without it. This matters because searches for new physics in the top quark sector may be able to use event topology rather than diagram removal to separate contributions, and single-resonant production, though smaller, is more sensitive to right-handed couplings. The paper also presents a statistical model based on asymptotic likelihood formulas to convert the neural network outputs into constraints.","feed_headline":"Splitting top events tightens Wtb coupling bound to 0.17","feed_subtitle":"Separate neural networks for single and double top production tighten limits on a right-handed Wtb coupling.","key_machinery":"The machinery is a two-level deep neural network architecture. The first-level network separates events into double resonant (ttbar) and single resonant (tWb) production based on kinematic observables; the second level consists of two separately trained networks, one for each production class, that discriminate between the left-handed vector operator (SM) and the anomalous right-handed vector operator (RV). The statistical engine is an asymptotic Poisson likelihood with profiled background normalization, using Wilks and Wald formulas and Asimov data to derive one-sided 95% CL limits on the coupling f_R^V. The Wtb vertex is the interaction of the top quark, the W boson, and the bottom quark.","core_discovery":"The central claim is that a two-stage neural network analysis of pp -> l nu b bbar q qbar' events with a tWb final state can separate the contributions of the left-handed vector operator (SM-like) and the anomalous right-handed vector operator at the Wtb vertex more effectively when single and double resonant top quark production are treated as separate classes. A first-level network classifies each event as double resonant (ttbar) or single resonant (tWb); two second-level networks then discriminate LV from RV events in each class. Using a Poisson likelihood with a nuisance parameter for background normalization and Asimov data, the expected 95% CL upper limit on |f_R^V| improves from 0.21 for the combined phase space to 0.17 when the phase space is split, assuming the single-resonant signal rate scales as f_R^$V^{4}$ and the double-resonant rate as f_R^$V^{2}$.","pith_inferences":["The size of the gain (0.21 to 0.17) depends on the assumed f_R^V^4 scaling of the single-resonant signal; if a full matrix-element calculation shows a different scaling, the gain from splitting could shrink or vanish.","The same two-stage separation could be applied to tensor operators at the Wtb vertex, where single-resonant production is also expected to be the more sensitive class.","A direct cross-check would be to generate Monte-Carlo samples at several f_R^V values and fit the signal strength in each class, rather than relying on the cited f_R^V^2 and f_R^V^4 power-law dependencies.","If the splitting gain persists under realistic systematic uncertainties, the approach could become a standard way to structure searches for anomalous top quark couplings at the LHC."],"forward_implications":["A combined analysis of single and double resonant top quark events, without removing diagrams from the gauge-invariant set, can yield a stricter constraint on the anomalous right-handed vector operator than analyzing the full tWb final state as one class.","The two neural network levels provide separable discriminators that can be reused for event classification in experimental analyses of top quark final states.","The statistical procedure, with a Poisson likelihood and a profiled nuisance parameter, gives a practical route from classifier outputs to 95% CL upper limits on anomalous couplings.","Separate limits can be quoted for the double-resonant and single-resonant phase space regions, reflecting their different sensitivities to anomalous right-handed couplings.","The method offers a way to account for interference between single and double resonant production by classifying events rather than by removing part of the matrix element set."],"supporting_citations":[{"why":"Provides the method for separating single and double resonant top quark production in phase space using deep neural networks, which the current work builds on.","marker":"[6]"},{"why":"Supplies the assumed dependence of the signal strength on f_R^V (f_R^V^2 for double resonant, f_R^V^4 for single resonant) used to compute the limits.","marker":"[22]"},{"why":"Gives the asymptotic formulas for upper limits using the Wilks and Wald approximations employed by the paper.","marker":"[17]"},{"why":"Defines the general Wtb vertex Lagrangian with left- and right-handed vector and tensor form factors used to set up the (LV,RV) scenario.","marker":"[13]"},{"why":"Supplies the construction of high-level observables used as input variables for the neural networks.","marker":"[14]"},{"why":"Describes the CompHEP package used for Monte-Carlo event generation and numerical calculations.","marker":"[7]"}],"fun_headline_variants":["Splitting top events by resonance type tightens Wtb bound to 0.17","Two-stage AI divides top events to refine Wtb coupling limit","Neural network split of top production sharpens Wtb right-handed bound"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole comparison rests on the assumed rate scaling for the anomalous right-handed coupling: the single-resonant class is taken to grow as f_R^$V^{4}$ and the double-resonant class as f_R^$V^{2}$, a dependence cited from reference [22] and not validated here; if that scaling is wrong, the quoted limits and the size of the splitting gain change.","fun_headline_variants_meta":{"raw":{"variants":["Splitting top events by resonance type tightens Wtb bound to 0.17","Two-stage AI divides top events to refine Wtb coupling limit","Neural network split of top production sharpens Wtb right-handed bound"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001026,"raw_usage":{"total_tokens":4304,"prompt_tokens":906,"completion_tokens":3398,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":522,"completion_tokens_details":{"reasoning_tokens":3334}},"tokens_in":522,"tokens_out":3398,"duration_ms":27499,"temperature":1.0,"reasoning_tokens":3334,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:24:36.128150+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the expected upper limit on f_R^V using a full event generator at several fixed values of f_R^V (for example 0, 0.1, 0.2, 0.3) without assuming a power-law scaling, and compare the single-resonant and double-resonant signal rates; if the single-resonant rate does not scale approximately as f_R^$V^{4}$, the reported 0.17 versus 0.21 comparison will not reproduce.","supporting_citations":[{"cited_title":"Secondary particles spectra in the decay of a polarized top quark with anomalous $tWb$ coupling","cited_arxiv_id":"hep-ph/0601155","evidence_quote":"Supplies the assumed dependence of the signal strength on f_R^V (f_R^V^2 for double resonant, f_R^V^4 for single resonant) used to compute the limits."},{"cited_title":"Effective operators in top physics","cited_arxiv_id":"1008.3225","evidence_quote":"Defines the general Wtb vertex Lagrangian with left- and right-handed vector and tensor form factors used to set up the (LV,RV) scenario."}],"review_version":1}