{"id":"43e90797-f202-448c-958d-a873f5f2633b","arxiv_id":"2411.16361","paper_version":2,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"ATLAS observes no excess in H to Z plus light hadronic resonance decays and sets 95% CL branching-fraction limits near 10% for low resonance masses, improving on the previous search.","lead":"ATLAS searched for Higgs bosons decaying into a Z boson and a very light particle that decays into hadrons, using the full Run 2 LHC dataset. No signal was found, and the new limits are about twice as strong as the previous search for this decay.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Background reweighting extrapolation into the blinded signal region is not directly validated; the provided closure test covers only adjacent bands and excludes the post-classification tail.","rationale":"The reader's weakest assumption identifies the same critical point: the background reweighting NN is trained on sidebands and applied to the blinded signal region, with a closure test that does not cover the exact SR or the post-classification selection. This is the most load-bearing assumption because all reported limits are derived from a binned fit to m_llj where the background shape comes from this reweighted MC; a bias here would propagate directly to the signal-strength extraction. However, the concern does not warrant changing the verdict. The paper explicitly adds two RW performance systematic uncertainties (a control-region shape comparison and a bootstrap method) that are designed to cover this extrapolation, and the final limits are consistent with the expected background-only hypothesis with no significant excess. The analysis follows standard ATLAS practice, is published in Phys. Lett. B, and the central null result is robust to moderate background-model variations. The proposed closure test using Sherpa as pseudo-data would provide a direct, quantitative validation of the extrapolation, and would either confirm that the quoted systematics are sufficient or reveal an additional bias. Until such a test is performed, the concern remains a caveat rather than a demonstrated flaw, so the original ACCEPT verdict is unchanged.","tokens_in":50221,"tokens_out":5203,"duration_ms":52574,"concrete_test":"Use the alternative Sherpa Z+jets sample as pseudo-data. Apply the exact analysis chain to the Powheg+Pythia sample: train the reweighting NN on the sidebands only, compute weights, then pass events through the classification NN and apply the 0.93 threshold. Compare the resulting m_llj distribution in the 120–140 GeV range, and the classification-score distribution above 0.93, between the reweighted Powheg and Sherpa. If the two agree within the quoted RW performance systematic (CR and BS), the extrapolation is validated; a significant disagreement would show the SR background model is not reliable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that no excess is observed and the B(H→Za) limits are valid—depends on the background prediction in the 120–140 GeV signal region. This prediction is obtained by training a reweighting NN on the m_llj sidebands (100–120 and 140–180 GeV) and applying the resulting per-event weights to MC events in the SR. The closure test described in Sec. 5 is not a test of this extrapolation: a separate NN trained with a wider excluded region (115–145 GeV) is compared with data only in the adjacent intervals [115–120] and [140–145] GeV, not the exact SR, and the comparison is performed on the input variables before the classification-NN requirement. The SR background is then reduced by a classification threshold at score >0.93, which selects a tail of the distribution where the reweighted MC has very few events; the validation of the reweighted background in Figure 2 is also done before the classification cut. The CR and bootstrap systematic uncertainties in Sec. 7 partially cover this, but they rely on the assumption that the data/MC ratio in the CR (score 0.883–0.93) equals that in the SR and that the bootstrap variations have the correct scale. If the reweighting extrapolates incorrectly across the boundary of the blinded region or mis-models the high-score tail, the background shape in the SR would be biased, directly affecting the fitted signal strength and hence the reported limits.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a search for the Higgs boson decay H→Za, where a is a light hadronically decaying resonance with mass 0.5–3.5 GeV, using the full Run 2 dataset of 140 fb⁻¹ of proton–proton collisions at √s = 13 TeV recorded by ATLAS. The analysis selects events with a leptonically decaying Z boson and a highly boosted single jet containing the resonance decay products. The background is modelled by reweighting Monte Carlo simulation with a neural network trained on sidebands of the m_ℓℓj distribution, and two additional neural networks (regression and classification) are used to suppress background. A binned profile-likelihood fit to the m_ℓℓj distribution finds no significant excess over the background prediction, and 95% CL upper limits are set on B(H→Za) for both gluon and quark decay modes, as well as on the effective ALP coupling C^eff_ZH/Λ. The limits improve on the previous ATLAS result by up to a factor of two to three.","tokens_in":50487,"tokens_out":6512,"duration_ms":62078,"significance":"The result represents a substantial improvement over the earlier ATLAS search for H→Za in the hadronic final state, both in the data-driven background modelling and in the resulting branching-fraction limits. The analysis includes a careful treatment of systematic uncertainties, including dedicated uncertainties for the reweighting performance (control-region and bootstrap methods), alternative Monte Carlo generators, and experimental and theoretical signal uncertainties. The interpretation in terms of an ALP coupling provides new constraints on a well-motivated class of beyond-the-Standard-Model scenarios. The central claim of no excess is supported by the observed data being compatible with the background-only hypothesis across the mass range, with the largest local significance at about 1.5σ.","major_comments":[],"minor_comments":[{"comment":"The statement that the reweighting NN is able to reweight events in the excluded region effectively is supported only by a closure test on the adjacent bands [115,120] and [140,145] GeV, before the classification-NN requirement; the text should explicitly note that the control-region and bootstrap uncertainties described in Sec. 7 are intended to cover any residual extrapolation bias in the signal region, including the high-score tail.","section":"Sec. 5"},{"comment":"The notation for the final-state invariant mass is inconsistent (m_ℓℓj, mllj, and mℓℓj appear in the text and figures); please unify to a single symbol.","section":"Throughout"},{"comment":"The phrase 'exclusion limit is ~10% for the lower masses' is vague; specify the exact mass range (e.g., 0.5–1 GeV) for which this holds.","section":"Abstract and Sec. 8"},{"comment":"The lower panel label 'Significance' could be confused with the significance of the fitted signal; consider renaming it to 'Per-bin residual significance' for clarity.","section":"Fig. 3(b)"},{"comment":"There is a typo: 'non hard collision jets' should read 'non-hard-collision jets'.","section":"Sec. 4"}],"recommendation":"accept","confidential_remarks":"The stress-test concern about the sideband-to-signal-region extrapolation is real but adequately mitigated by the dedicated control-region and bootstrap systematic uncertainties, which are constructed to cover exactly this extrapolation. The closure test on adjacent bands is a reasonable validation given the blinding strategy, and the paper's claims are consistent with the level of detail expected in an experimental HEP publication. The manuscript is already published in Phys. Lett. B; the present review is confirmatory."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a well-executed, honest ATLAS search that roughly halves the previous constraints on H→Za at low resonance mass and adds new ALP coupling limits. The genuinely new piece is the data-driven NN background reweighting, which improves the background model enough to make the limits meaningfully stronger. The reweighting is trained on m_llj sidebands and interpolates into the blinded signal region; the closure test shown only covers adjacent bands and is done before the classification-NN cut. That leaves a real, but minor-to-moderate, soft spot: the high-score tail of the classification output, where the SR lives, is validated only indirectly by a control-region shape systematic and a bootstrap. The CR and bootstrap are sensible, but they rely on the assumption that the data/MC ratio in the score band 0.883–0.93 matches the SR. I don't see a better way to do it without unblinding, and the paper is transparent about it. The null result is robust; the limits are what they are with the stated uncertainties.\n\nThe statistical model is standard: profile likelihood with Gaussian constraints, free background normalization. Systematics are dominated by parton-shower modeling, with a 60% uncertainty on the gg signal efficiency; that's large but honestly estimated. The ALP interpretation is a nice addition, with the pi+pi-pi0 assumption clearly stated. The 4 GeV qq point is dropped with a reasonable justification. No circularity: the reweighting is trained on data before the SR fit, and the previous result is only used for comparison. Citations look appropriate.\n\nWho is this for: people doing exotic Higgs decays and ALP searches, plus anyone interested in data-driven NN background estimation in LHC analyses. It deserves a serious referee. I'd recommend acceptance with minor revisions, mainly asking for a bit more discussion of the extrapolation limitation and possibly showing the closure test after the classification cut or in the SR if blinding permits. Not a takedown; this is good work.","headline":"Solid, honest ATLAS search with improved limits and a novel NN background-reweighting; the main caveat is that the extrapolation into the blinded signal region is only indirectly validated.","tokens_in":51009,"tokens_out":1998,"would_cite":false,"duration_ms":22691,"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 searches for Higgs boson decays into a Z boson and a light resonance using the full ATLAS Run 2 dataset and finds no significant signal, setting new branching-fraction limits.","keywords":["Higgs boson","Z boson","light resonance","branching fraction","axion-like particle","neural network reweighting","LHC","ATLAS"],"falsifier":"A future LHC dataset with more integrated luminosity should show whether the mild 1.5$\\sigma$ excess near 135 GeV grows into a 3$\\sigma$ or larger deviation, which would overturn the paper's no-excess conclusion, or fades away, which would confirm it.","tokens_in":50015,"feed_emoji":"⚛️","tokens_out":6293,"duration_ms":57422,"temperature":0.7,"pith_summary":"This paper tries to establish whether the Higgs boson can decay into a Z boson and a light, short-lived resonance of mass 0.5 to 3.5 GeV, such as a charmonium state, an axion-like particle, or a light pseudoscalar from a two-Higgs-doublet model. Analyzing the full 140 fb$^{-1}$ of 13 TeV proton-proton collisions recorded by ATLAS, it finds no significant excess over the Standard Model background. It therefore sets upper limits at 95% confidence: the branching fraction $B(H\\to Za)$ is below about 10% for the lightest resonance masses when the light particle decays to gluons, and the effective axion coupling to the Higgs and Z bosons is constrained to 0.9 to 2 TeV$^{-1}$. The result matters because a measurable $H\\to Za$ rate would be a clear sign of new particles coupling to the Higgs boson.","feed_headline":"No excess found in Higgs to Z plus light resonance search","feed_subtitle":"Full Run 2 dataset restricts Higgs-to-Z-plus-light-resonance branching to about 10 percent at low mass.","key_machinery":"The analysis is carried by three neural networks plus a profile-likelihood fit. A reweighting neural network, trained on the $m_{\\ell\\ell j}$ sidebands of 100 to 120 GeV and 140 to 180 GeV, estimates a log-likelihood ratio between Monte Carlo background and data and reweights the simulation so that its kinematic and jet-substructure distributions match the data, while the 120 to 140 GeV signal region is excluded from training. A regression neural network maps seven track-based jet-substructure variables into an estimate of the light-resonance mass, and a classification neural network then separates signal from background using those variables plus the mass estimate. Finally, a binned profile-likelihood fit to the $m_{\\ell\\ell j}$ distribution in the range 100 to 178 GeV, with systematic uncertainties treated as nuisance parameters, extracts the signal strength and sets limits using the $CL_s$ asymptotic formulae.","core_discovery":"The paper's central claim is that, in the dataset examined, Higgs boson decays into a Z boson and a light hadronically decaying resonance are not present at a statistically significant level. For each assumed resonance mass between 0.5 and 3.5 GeV and for each decay mode ($a\\to gg$ or $a\\to q\\bar{q}$), the observed invariant-mass distribution of the dilepton-plus-jet system is compatible with the background-only prediction. The most notable feature is a mild excess near 135 GeV in the reconstructed mass that reaches a local significance of about 1.5$\\sigma$ for the $m_a=0.5$ GeV gluon-decay hypothesis, which is not enough to claim a discovery. The paper converts this null result into 95% CL upper limits on $B(H\\to Za)$ and on the effective ALP coupling $C^\\text{eff}_{ZH}/\\Lambda$, reporting limits roughly two to three times stronger than the earlier version of this search.","pith_inferences":["The same sideband-reweighting technique could be ported to other searches for narrow resonances where Monte Carlo does not model jet substructure accurately, provided a signal-free sideband exists.","Because the sensitivity drops for higher resonance masses mostly because their decays look more like ordinary QCD jets, improvements in low-energy jet substructure or in reconstructing softer tracks could push the reach below the current 10% branching-fraction barrier.","The ALP limit only covers the charged three-pion decay mode; an analysis with sensitivity to the neutral $3\\pi^0$ mode would be needed to close the gap for ALPs in the 0.5 to 1 GeV range.","The mild 1.5$\\sigma$ excess near 135 GeV is a candidate fluctuation to watch in future data: if it grows with more statistics, it would most naturally appear for the lightest resonances with gluon decays and would shift the limits downward."],"forward_implications":["If the limits are correct, any new particle that makes $H\\to Za$ occur with a branching fraction above about 10% for $m_a\\approx 0.5$ GeV (gluon decays) is ruled out, with the excluded range extending to higher masses at weaker limits.","The effective ALP coupling $C^\\text{eff}_{ZH}/\\Lambda$ is excluded above 0.9 to 2 TeV$^{-1}$ at 95% CL, narrowing the parameter space for a Higgs-coupled axion-like particle.","For charmonium states, the search cannot exclude physical branching fractions: the limits are $B(H\\to Z\\eta_c)>1.2$ and $B(H\\to ZJ/\\psi)>1.4$.","The factor-of-two-to-three improvement over the previous ATLAS search indicates that data-driven neural-network background reweighting substantially reduces the dominant background-modeling uncertainty."],"supporting_citations":[{"why":"The earlier version of this search that this paper supersedes, providing the baseline limits and the main background-modeling limitation that motivates the neural-network reweighting.","marker":"[30]"},{"why":"Supplies the ALP decay rates and the effective coupling assumptions used to translate the branching-fraction limits into bounds on $C^\\text{eff}_{ZH}/\\Lambda$.","marker":"[18]"},{"why":"Establishes that a binary classifier approximates a density ratio, which is the theoretical basis for the background-reweighting neural network.","marker":"[70]"},{"why":"Provides the log-likelihood-ratio estimation method and exponential loss function used to train the reweighting neural network.","marker":"[71]"},{"why":"Motivates the gluon and quark decay modes by showing that in the 2HDM+S scenario, $a\\to gg$ dominates for masses up to about 3 GeV.","marker":"[28]"},{"why":"Gives the asymptotic formulae for the profile-likelihood test statistic used to compute significances and fit results.","marker":"[79]"},{"why":"Defines the $CL_s$ technique used to set the reported 95% confidence-level upper limits.","marker":"[80]"}],"fun_headline_variants":["ATLAS finds no Higgs decay to Z and light hadrons","No new physics in Higgs to Z plus light resonance search","ATLAS sets tight limits on Higgs to Z plus light resonance","No excess in Higgs to Z plus light resonance search","Higgs to Z plus light resonance: no signal, limits set"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole result rests on the assumption that the background-reweighting neural network, trained on data from mass sidebands on either side of the signal region, predicts the background inside the 120 to 140 GeV signal region correctly, including for jet-substructure variables after the classification selection; if that extrapolation is biased, the background shape and all quoted limits would shift.","fun_headline_variants_meta":{"raw":{"variants":["ATLAS finds no Higgs decay to Z and light hadrons","No new physics in Higgs to Z plus light resonance search","ATLAS sets tight limits on Higgs to Z plus light resonance","No excess in Higgs to Z plus light resonance search","Higgs to Z plus light resonance: no signal, limits set"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000509,"raw_usage":{"total_tokens":2545,"prompt_tokens":1078,"completion_tokens":1467,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":694,"completion_tokens_details":{"reasoning_tokens":1383}},"tokens_in":694,"tokens_out":1467,"duration_ms":12448,"temperature":1.0,"reasoning_tokens":1383,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:12:46.162260+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A future LHC dataset with more integrated luminosity should show whether the mild 1.5$\\sigma$ excess near 135 GeV grows into a 3$\\sigma$ or larger deviation, which would overturn the paper's no-excess conclusion, or fades away, which would confirm it.","supporting_citations":[{"cited_title":"Search for Higgs boson decays into a $Z$ boson and a light hadronically decaying resonance using 13 TeV $pp$ collision data from the ATLAS detector","cited_arxiv_id":"2004.01678","evidence_quote":"The earlier version of this search that this paper supersedes, providing the baseline limits and the main background-modeling limitation that motivates the neural-network reweighting."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the $CL_s$ technique used to set the reported 95% confidence-level upper limits."}],"review_version":1}