{"id":"0210dd4b-bc8f-4d91-b39d-8f09cb36360e","arxiv_id":"2505.07769","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A GNN-based search could give the HL-LHC an exclusion reach near 2.4 TeV for vectorlike B quarks decaying fully hadronically through b plus a singlet scalar, with performance comparable to semileptonic searches.","lead":"This paper simulates LHC collisions to test whether a graph neural network can pick out rare, fully hadronic decays of a hypothetical heavy bottom-like quark B. It matters because it suggests the High-Luminosity LHC could search this difficult all-jet channel out to masses around 2.4 TeV, close to what semileptonic searches can do.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"QCD multijet background is asserted to be 'essentially eliminated' without any simulation, cross-section, or row in Table II; if QCD events survive C1-C5 at a non-negligible rate, the Fig. 9 reach is optimistic.","rationale":"I read the paper as a simulation-based projection study whose central claim is that a GNN-based analysis of the fully hadronic 2b+4j/6b channel can reach MB ~ 1.8 TeV for discovery and ~2.4 TeV for exclusion at HL-LHC with BR(B->bPhi)=100%. For this claim to hold, the background model must be complete and the GNN must perform as simulated. The weakest link is the treatment of the QCD multijet background: it is the largest-rate process at a hadron collider, yet it is excluded from Table I and Table II, with only a qualitative assertion that it is 'essentially eliminated' by selection cuts. No quantitative estimate is given, so the background normalization is potentially underestimated. This concern is load-bearing because the signal yields are small in the high-mass region where the reach is claimed, and even a modest additional background component could move the contours significantly. The reader identified the same weakest assumption, and I agree. The reader's conditional verdict is appropriate: the paper should be accepted only if the authors provide a QCD multijet estimate, a systematic uncertainty treatment, and a held-out threshold selection protocol, or clearly label the results as optimistic upper bounds. I therefore leave the verdict unchanged.","tokens_in":18665,"tokens_out":11715,"duration_ms":117748,"concrete_test":"Generate a QCD multijet sample (pp->jj, pp->jjj, pp->bb+jets) at 14 TeV using the same MadGraph+Pythia8+Delphes setup with MLM matching, apply the C1-C5 selection, and run the surviving events through the pretrained and finetuned GNN classifiers. Compute the QCD yield at L=3 ab^-1. If the post-GNN QCD yield exceeds roughly 10^4 events (about 1% of the 1.58e6 background after C5), add it to NB and recompute ZD/ZE for the Fig. 9 grid; if the MB reach drops by more than ~100 GeV at any BR(B->bPhi), the stated HL-LHC reach is not robust and the paper should present an explicit QCD estimate or label the results as optimistic upper bounds.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central reach claim (discovery to ~1.8 TeV, exclusion to ~2.4 TeV at HL-LHC for BR(B->bPhi)=100%) rests on the background model in Table I and the cut flow in Table II, which contains no QCD multijet process. Section III A states that 'the strong pT cuts and the demand for b-tagged jets essentially eliminate the QCD multijet background,' but no QCD sample is generated, no cross-section is quoted, and Table II has no QCD row. After C5 the total SM background is 1.58e6 events at L=3 ab^-1; the signal at the claimed exclusion edge is small (e.g., 571 events after C5 for MB=2 TeV, 14 for 2.8 TeV). An unsuppressed QCD multijet component, either real b-jets from gluon splitting or mistagged light jets, could easily contribute at a level comparable to or exceeding the included backgrounds, because the inclusive jet cross section is orders of magnitude larger than the processes listed. Because QCD events were also absent from GNN training, the classifier has never seen this class and its rejection is unknown. If the post-selection plus GNN QCD yield is even a few percent of the 1.58e6 background, the 5-sigma and 2-sigma contours in Fig. 9 shift to lower MB, undermining the claim that the fully hadronic mode is competitive with semileptonic searches.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a search strategy for pair-produced vectorlike B quarks decaying through the exotic mode B -> b Phi with fully hadronic final states (2b + 4j or 6b), where Phi is a new gauge-singlet scalar or pseudoscalar decaying mainly to gg or bb. The authors generate signal and Standard Model background events at sqrt(s) = 14 TeV with MadGraph/Pythia/Delphes, apply a five-stage cut flow (C1-C5), and then train a graph neural network followed by a deep neural network to separate signal from background. They quote projected HL-LHC discovery and exclusion reaches on the (M_B, M_Phi) plane for branching fractions beta_bPhi = 0.4, 0.7, 1.0, with the headline claim that at L = 3000 fb^-1, MB up to about 2.4 TeV can be excluded and a discovery reach around 1.8 TeV can be achieved for BR(B -> bPhi) = 100%. The paper also translates the exclusion contours into the BR(B -> bPhi) versus MB plane for singlet and doublet B + Phi models, comparing with current LHC limits.","tokens_in":18964,"tokens_out":3851,"duration_ms":42657,"significance":"If the projected reach is reliable, the paper would be a useful contribution: it demonstrates that a dedicated fully hadronic search with a GNN classifier can compete with semileptonic searches for exotic vectorlike B decays, and it provides a concrete analysis pipeline, including public model files. The event-generation setup, cut definitions, and GNN architecture are described in enough detail to be reproduced, and the classifier training strategy is clearly explained. However, the central reach claims rest on two incompletely justified assumptions: that the QCD multijet background is negligible without a quantitative estimate, and that systematic uncertainties can be ignored in the significance formulas. These gaps affect the headline numbers and need to be addressed before the projected reach can be taken at face value.","major_comments":[{"comment":"The QCD multijet background is asserted to be 'essentially eliminated' by the strong pT cuts and the b-tagging requirement, but no QCD multijet sample is generated, no cross-section is quoted, and Table II contains no multijet row. The inclusive multijet cross-section is orders of magnitude larger than the processes listed, and neither real b-jets from gluon splitting nor mistagged light jets are excluded at the rates required to make this statement quantitative. Because the GNN was trained without any QCD multijet class, its rejection power for this background is unknown. Please provide a quantitative estimate from a generated QCD multijet sample passing C1-C5 (and, ideally, the GNN classifier), or a data-driven sideband estimate, and include the result in the background yield used in Eq. (16).","section":"III A / Table II"},{"comment":"The discovery and exclusion significances are computed from the Poisson counting formulas of Eqs. (16) and (17), which contain no systematic uncertainties. With the total background after C5 being 1.58e6 events and the surviving signal being a few hundred events at the exclusion edge, even a few percent uncertainty on the background normalization or on the b-tagging efficiency can shift the 5-sigma and 2-sigma contours in Fig. 9 substantially. The projected reach should be recomputed with nuisance parameters (for example, log-normal background uncertainties) or the authors should demonstrate explicitly that such systematics are negligible for this analysis.","section":"V, Eqs. (16)-(17)"},{"comment":"The abstract states a discovery reach of about M_B = 1.8 TeV for BR(B -> bPhi) = 100%, while Section V states that 'it is possible to attain a discovery significance score of 5 sigma at values M_B > 2 TeV.' These two statements are inconsistent, and the exact mass value corresponding to the beta_bPhi = 1.0 contour in Fig. 9(a) should be identified. Please reconcile the abstract with the body of the paper.","section":"Abstract / V"},{"comment":"The text in Section IV C says that the threshold on the classifier response is scanned and the value maximizing the discovery sensitivity in Eq. (16) is selected, but it does not specify whether this optimization is performed on a validation set that is independent from the events used to compute the quoted N_S and N_B. If the same events are used both to choose the threshold and to evaluate the significance, the projected Z_D is biased upward. Please clarify the train/validation/test split and, if a separate validation set was used, state the resulting threshold selection procedure explicitly.","section":"IV C"}],"minor_comments":[{"comment":"The background labels such as 'thth', 'ththH', 'ththWh', and 'ththZh' are not defined in the text or captions; please clarify which SM processes these abbreviations denote and how they map to the processes listed in Table I.","section":"Table II"},{"comment":"The selection efficiency in Fig. 5 is quoted without statistical uncertainties; adding them would help assess whether the small differences across the parameter grid are meaningful.","section":"III A / Fig. 5"},{"comment":"The notation for the Higgs doublet H and its vev is standard, but the sign conventions in Eqs. (2)-(4) should be stated more explicitly so that the mixing angles and the signs of the off-diagonal couplings are unambiguous.","section":"II A, Eq. (2)"},{"comment":"The comparison with the previous semileptonic result in Ref. [27] is stated only in the text; showing the corresponding exclusion curve in Fig. 10 would make the claimed improvement easier to verify.","section":"V / Fig. 10"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the journal's scope and the authors are transparent about their simulation choices, which I appreciated. The central concern is not circularity but completeness: the missing QCD multijet background and the absence of systematic uncertainties are load-bearing for the quoted reach. If the authors can provide a quantitative QCD estimate and a systematics-inclusive significance calculation, the paper could become a solid contribution. I would also encourage the editor to check that the abstract and Section V are reconciled before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read the fully hadronic B→bΦ paper. The punchline: this is a serious, competently executed simulation study that deserves peer review, but the central reach claim—2.4 TeV exclusion at HL-LHC—rests on the QCD multijet background being waved away rather than estimated. That is the soft spot that matters.\n\nWhat's genuinely new: the channel itself. The 2b+4j/6b fully hadronic final state with a GNN event classifier is not in the cited literature, and the sequential graph construction with shared-attribute nodes is a real adaptation, not a repackaging. The event generation is careful: FeynRules/MadGraph/Pythia/Delphes with MLM matching, NNLO signal cross sections, and high-order background cross sections for everything listed. The two-step training strategy (pre-train across the mass grid, fine-tune per point) is sensible and clearly documented. Model files are public. This is reproducible, formal work in the sense that matters.\n\nNow the load-bearing flaw. Section III A asserts that strong pT cuts and b-tagging 'essentially eliminate' the QCD multijet background, but there is no QCD sample, no cross-section quoted, and no row in Table II. After C5, the total SM background is 1.58e6 events at 3 ab⁻¹. The inclusive jet cross section is orders of magnitude larger than the processes shown. Even a fraction of a percent survival—real b-jets from gluon splitting or mistagged light jets—would add a non-negligible term, and because QCD events were never sent through the GNN, the classifier has no learned rejection for them. This is exactly the kind of thing that shifts 2σ contours down by hundreds of GeV. I don't think it invalidates the method, but it does mean the 2.4 TeV exclusion should be presented as optimistic until a QCD estimate is supplied.\n\nSecondary issues, in proportion: the significance formulas are pure Poisson counting, with no systematic uncertainty term; the threshold is selected per mass point on the evaluated samples, which biases significance upward; and the abstract says discovery reach to about 1.8 TeV while Section V says 5σ above 2 TeV—that inconsistency should be resolved. The concluding paragraph calls the analysis 'conservative' because it ignores single production, but omitting QCD background is not conservative in that direction.\n\nWho is this for? Collider phenomenologists working on VLQ searches and people who want an example of GNN event classification done with care. It deserves a serious referee, and I would send it out—but with a clear instruction that the authors must either generate a QCD multijet sample or give a quantitative argument for its suppression, add systematics, and move threshold selection to a held-out set. As it stands, the paper is a step in the right direction with an overconfident headline.","headline":"A credible, well-executed GNN-based collider study whose headline reach depends on an unquantified assertion that the QCD multijet background is negligible.","tokens_in":19581,"tokens_out":2029,"would_cite":false,"duration_ms":21999,"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":"The paper projects that a graph-neural-network search can reach 1.8 TeV discovery and 2.4 TeV exclusion for fully hadronic vectorlike B decays at the HL-LHC.","keywords":["vectorlike B quark","exotic decay","fully hadronic final state","graph neural network","singlet scalar","HL-LHC reach","bottom-quark tagging","LHC search strategy"],"falsifier":"Generate a large QCD multijet sample through the same detector simulation and selection C1–C5 with realistic b-tagging, apply the trained GNN, and add the survivors to $N_B$ in Eq. (16); if the resulting $5\\sigma$ discovery contour drops below roughly 1.5 TeV for $BR(B\\to b\\Phi)=100\\%$, the paper's central reach claim is contradicted.","tokens_in":18408,"feed_emoji":"⚛️","tokens_out":10836,"duration_ms":97829,"temperature":0.7,"pith_summary":"This paper argues that the fully hadronic decay chain $pp \\to B\\bar{B} \\to (b\\Phi)(\\bar{b}\\Phi)$, with the singlet scalar or pseudoscalar $\\Phi$ decaying to $gg$ or $b\\bar{b}$, can be brought within reach of the HL-LHC even though the final state has no leptons to trigger on. The authors build a hybrid classifier that represents each collision event as a graph of jets, fatjets, and event-level features, passes it through a graph neural network, and finishes with a deep neural network. With this pipeline and 3000 fb$^{-1}$ of data, they project that a search in the $2b+4j/6b$ final state could exclude vectorlike $B$ masses up to about 2.4 TeV and discover them near 1.8 TeV when $B\\to b\\Phi$ saturates the branching ratio. If correct, this would make a fully hadronic search competitive with semileptonic searches for the same process.","feed_headline":"Graph network search pushes hadronic B-quark reach to 2.4 TeV","feed_subtitle":"Fully hadronic decays of pair-produced vectorlike B quarks could match semileptonic searches at the HL-LHC.","key_machinery":"The central machinery is an event graph with heterogeneous node types: shared-attribute nodes carry the four-momenta of every reconstructed jet and fatjet, auxiliary jet and fatjet nodes carry substructure and b-tag information, a global node carries event-level variables, and a CLS token aggregates the embedding for classification. The graph is built sequentially, first connecting kinematic nodes in a clique, then linking each object to its attributes, then connecting nearby objects with a $k$-nearest-neighbour rule in the $\\eta$-$\\phi$ plane, and finally connecting everything to the CLS token. Attention-based graph-convolution message passing updates the embeddings, and a five-layer deep neural network performs the final signal-versus-background classification after a bias-adjusted loss is used in pre-training on all mass points and fine-tuning on each point.","core_discovery":"The paper's central claim is that a hybrid deep-learning classifier, a graph neural network that builds an event-level embedding from every reconstructed jet, fatjet, and global event feature followed by a deep neural network, can separate the fully hadronic $pp\\to B\\bar B\\to (b\\Phi)(\\bar b\\Phi)$ signal, with $\\Phi\\to gg/b\\bar b$, from Standard Model backgrounds. Applied to 14 TeV proton collisions at the HL-LHC luminosity of 3000 fb$^{-1}$, the analysis projects a $5\\sigma$ discovery reach around $M_B \\simeq 1.8$ TeV and a $2\\sigma$ exclusion reach up to about 2.4 TeV when $BR(B\\to b\\Phi)=100\\%$, with the exclusion still reaching about 1.8 TeV at 40% branching. The same reach applies to singlet $B+\\Phi$ and doublet $(T,B)+\\Phi$ models, and it makes the fully hadronic search competitive with the semileptonic search that previously set the benchmark for this process.","pith_inferences":["The paper's zero-multijet assumption can be tested before the LHC run: a data-driven sideband that measures how often the selection cuts pass ordinary QCD events would directly bound how much of the 2.4 TeV exclusion is real.","The pretrain-then-finetune recipe suggests the event-graph encoder could be reused as a single pretrained backbone for several hadronic resonance searches, with only the final classifier retrained for a given mass point.","Because single-$B$ production can become competitive with pair production above about 2 TeV, including single-$B$ and $pp\\to B\\Phi$ contributions could extend the reach beyond the pair-production-only contours shown here."],"forward_implications":["At 3000 fb$^{-1}$, masses up to about 2.4 TeV could be excluded for $BR(B\\to b\\Phi)\\simeq 100\\%$, with discovery sensitivity around 1.8 TeV; even at 40% branching the exclusion still reaches about 1.8 TeV.","The fully hadronic $2b+4j/6b$ channel becomes competitive with the semileptonic $b\\Phi$ search, which had previously set the benchmark for this decay mode.","The same exclusion contours apply to both singlet $B+\\Phi$ and doublet $(T,B)+\\Phi$ models, extending beyond current recast LHC limits and beyond the monoleptonic reach of the authors' earlier study.","Because the signal yield scales as $BR(B\\to b\\Phi)^2$, the plotted contours can be rescaled to estimate sensitivity for any intermediate branching ratio or for additional $B$ decay modes.","Section V further states that, for part of the parameter plane, a $5\\sigma$ discovery significance can be reached at $M_B$ above 2 TeV."],"supporting_citations":[{"why":"Defines the singlet and doublet $B+\\Phi$ models, the parameter space, and the recast LHC constraints that motivate the fully hadronic search.","marker":"[25]"},{"why":"The authors' earlier machine-learning search for the same $B$ quark in the semileptonic mode; supplies the performance benchmark this paper aims to match and the bias-adjusted loss strategy.","marker":"[27]"},{"why":"Supplies the b-tagging working point, including efficiency and mistag rates, used in the event selection.","marker":"[33]"},{"why":"Source of the NNLO signal cross section and the QCD $K$-factor used to normalize the pair-production signal.","marker":"[35]"},{"why":"Provides the message-passing formalism for graph neural networks in particle physics that the event-graph classifier is built on.","marker":"[45]"},{"why":"Provides the asymptotic significance formulas, $Z_D$ and $Z_E$, used to convert classifier output and event counts into discovery and exclusion reaches.","marker":"[66]"}],"fun_headline_variants":["GNN tagger probes hadronic B' to 1.8 TeV discovery","Hadronic B' search reaches 2.4 TeV exclusion via GNN","Graph network tags hadronic B' decays at HL-LHC","GNN makes hadronic B' search match semileptonic","Fully hadronic B' reach extended by GNN to 2.4 TeV"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that ordinary QCD multi-jet events essentially vanish after the selection cuts and b-tagging requirements, so they are left out of both the background count and the training data; if even a small fraction survive, the projected discovery and exclusion masses are too high.","fun_headline_variants_meta":{"raw":{"variants":["GNN tagger probes hadronic B' to 1.8 TeV discovery","Hadronic B' search reaches 2.4 TeV exclusion via GNN","Graph network tags hadronic B' decays at HL-LHC","GNN makes hadronic B' search match semileptonic","Fully hadronic B' reach extended by GNN to 2.4 TeV"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000868,"raw_usage":{"total_tokens":3803,"prompt_tokens":1030,"completion_tokens":2773,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":646,"completion_tokens_details":{"reasoning_tokens":2671}},"tokens_in":646,"tokens_out":2773,"duration_ms":20904,"temperature":1.0,"reasoning_tokens":2671,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:09:30.478404+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a large QCD multijet sample through the same detector simulation and selection C1–C5 with realistic b-tagging, apply the trained GNN, and add the survivors to $N_B$ in Eq. (16); if the resulting $5\\sigma$ discovery contour drops below roughly 1.5 TeV for $BR(B\\to b\\Phi)=100\\%$, the paper's central reach claim is contradicted.","supporting_citations":[{"cited_title":"Vector boson pro- duction at hadron colliders: A fully exclusive qcd calcula- tion at next-to-next-to-leading order,","cited_arxiv_id":null,"evidence_quote":"Provides the message-passing formalism for graph neural networks in particle physics that the event-graph classifier is built on."}],"review_version":1}