{"id":"a8ed4cb1-bf6a-45a4-a33e-09f966053477","arxiv_id":"2411.09357","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A low-pT collimated hadronic di-tau tagger is built from a large-radius jet with two subjets and a BDT, with data-to-simulation scale factors of 1.00 and 1.01 measured in Z-gamma events.","lead":"This ATLAS paper presents a new algorithm that reconstructs pairs of highly collimated hadronically decaying tau leptons as a single object, using a large-radius jet, two subjets, and a boosted decision tree. The identification efficiency is measured in Z-gamma events and matches simulation within 26-37% uncertainties, which enables low-mass new-physics searches.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Inclusive Zgamma scale factor may not transfer to light-resonance phase space; no extrapolation uncertainty is quantified.","rationale":"The reader's weakest assumption identifies exactly the load-bearing concern: the inclusive Zgamma scale factor may not transfer to the light-resonance phase space, and no extrapolation uncertainty is quantified. This is the main limitation for the tagger's stated physics use. The paper itself discloses the phase-space mismatch in Section 6 and the Conclusions, so the concern is not a hidden flaw but a self-identified gap. The internal measurement appears sound: the SF from Eq. (1) is consistent with the yields in Table 5, the ABCD background estimation is standard and validated in the VR, and the reweighting is derived in a signal-depleted control region. The reconstruction efficiency is not measured in data, but the quoted SF is a combined reconstruction+identification correction for truth-matched events, so it is still applicable as an event weight. The most decisive check is a kinematic-binned SF measurement; the proposed ΔR binning directly probes the variable that separates mX=20 from the Z mass. Therefore the reader's conditional verdict is appropriate and no verdict change is needed.","tokens_in":58331,"tokens_out":15747,"duration_ms":151008,"concrete_test":"Recompute SF(BDT>0.35) in two bins of ΔR(lead, sublead), e.g., ΔR<0.5 and ΔR≥0.5, using the existing Zgamma tag-and-probe data with the same ABCD background estimation and reweighting. The mX=20 signal predominantly populates ΔR<0.5. If the per-bin scale factors are compatible with each other and with unity within the (larger) per-bin statistical uncertainties, the inclusive SF is plausibly transferable; if they differ by more than the combined uncertainty, the inclusive SF is not flat in the kinematic variable most correlated with parent mass, and a quantitative extrapolation uncertainty must be assigned before applying the tagger to light-resonance searches.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central result is SF(BDT>0.35)=1.00±0.37 and SF(BDT>0.5)=1.01±0.27, measured inclusively in Z→ττ+γ events. The tagger's intended application is light resonances with mX=20–60 GeV, which populate a more collimated region (smaller ΔR(lead, sublead)) and a softer seed-jet pT spectrum than the Zgamma sample. Figures 7(c) and 7(d) show that the identification efficiency in simulation varies strongly with ΔR(lead, sublead) and seed pT, and Section 6.4 explicitly states that the efficiency depends on the kinematic phase space. Because the SF is a single inclusive number, any data/simulation difference in BDT response that varies with ΔR or pT would not be captured. The paper acknowledges this—'the phase space probed in the tag-and-probe measurement is not expected to match that of ttbar X processes' and 'additional uncertainties may be required'—but provides no quantitative extrapolation uncertainty. If the true SF for mX=20 differs from the Zgamma value by more than the quoted 26–37%, the tagger efficiency in light-resonance searches would be incorrect by an unquantified amount.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper describes a new ATLAS algorithm for reconstructing and identifying highly collimated hadronically decaying tau-lepton pairs at low transverse momentum. The reconstruction uses a large-radius jet (R=1.0) and its two leading subjets, and the identification uses a boosted decision tree (BDT) trained on ttbar+X events with X masses of 20 and 60 GeV. The identification efficiency is measured in Zgamma events using 139 fb^-1 of sqrt(s)=13 TeV data. The resulting data-to-simulation scale factors are SF(BDT>0.35)=1.00 +/- 0.37 and SF(BDT>0.5)=1.01 +/- 0.27, with statistical and systematic components given in Table 4. The measurement is statistically limited, with signal purity below 1% in the signal region, and relies on an ABCD method for background normalisation plus a reweighting procedure validated in a validation region.","tokens_in":58514,"tokens_out":8623,"duration_ms":89427,"significance":"If correct, this provides the first measured identification scale factors for a low-pT di-tau tagger in ATLAS, enabling future searches for light resonances decaying to tau pairs. The analysis is careful: the ABCD background estimate, the CR/VR-validated reweighting, and the full propagation of experimental and theoretical systematics are all standard and appropriate. The authors are transparent about the fact that the Zgamma phase space does not match the ttbar+X target phase space and that additional uncertainties may be required for other analyses. The main limitation is that no quantitative extrapolation uncertainty is provided; this is acknowledged but left to future work.","major_comments":[],"minor_comments":[{"comment":"The paper states that 'additional uncertainties may be required' for analyses targeting a different phase space, but it does not provide any quantitative estimate of the extrapolation uncertainty. Given the strong dependence of the identification efficiency on Delta R(lead, sublead) and seed pT shown in Figures 7(c) and 7(d), and the fact that the measurement uses 90 < pT(seed) < 360 GeV, future analyses of light resonances (mX = 20-60 GeV) would benefit from an explicit statement that the inclusive SF is only valid in the measured phase space, or from a simulation-based closure test comparing the SF in the Zgamma and ttbar+X phase spaces.","section":"Section 6.4 and Conclusions"},{"comment":"Consider overlaying the kinematic distributions of the reconstructed di-tau objects in the Zgamma signal region on the efficiency curves in Figures 7(c) and 7(d); this would help the reader assess how much of the efficiency variation is actually probed by the measurement and quantify the extrapolation to the light-resonance topology.","section":"Figure 7"},{"comment":"The claim of 'no evidence of classifier overtraining' is based on visual inspection of Figure 6(a); adding a quantitative test statistic (e.g., a Kolmogorov-Smirnov p-value) for the training and testing BDT score distributions would strengthen this statement.","section":"Section 5.3"},{"comment":"In the sentence 'used for a heavy, narrow, scalar resonance search in the high mass regime of 1-3TeV', the unit should read '1-3 TeV' with a space before TeV.","section":"Section 1"},{"comment":"In the 'Statistical' uncertainty category, the meaning of 'MC' is not explicitly defined; the text should clarify whether this refers to the finite size of the simulated signal sample only or to all simulated samples entering the measurement.","section":"Table 4"}],"recommendation":"minor_revision","confidential_remarks":"The manuscript is a standard ATLAS performance paper with a very large author list. The science is sound and the presentation is generally clear. The only substantive limitation is the acknowledged phase-space extrapolation issue, which I do not consider blocking for publication in this journal; a more explicit caveat in the conclusions would suffice."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a genuine new tool—a tagger for collimated hadronic tau pairs with seed jet pT down to 50 GeV—plus a careful, statistically limited measurement of its identification efficiency in Z→ττ+γ events. The scale factors are 1.00±0.37 (Medium) and 1.01±0.27 (Tight), consistent with unity, with uncertainties dominated by statistics. The authors are transparent that reconstruction efficiency is not measured in data and that the SF is measured in a kinematic regime that does not match the intended light-resonance phase space.\n\nWhat is new and good: the adaptation of the existing high-pT di-tau tagger down to low pT is a real extension, not a trivial retread. The dedicated energy-scale calibration for subjets, with pile-up subtraction and response functions, is well motivated. The BDT is trained on mX = 20 and 60 GeV samples and the working points are reasonable. The ABCD background estimate, the reweighting derived in a control region and validated in a validation region, and the full systematic propagation through the chain are all standard and executed with care. The signal purity in the signal region is below 1%, and the authors correctly note that the inclusive SF is therefore a direct statement about data/MC agreement for the di-tau object, not a tuned parameter.\n\nThe soft spot is the one the authors flag themselves: phase-space transferability. The SF is measured inclusively in Zgamma events (Z mass ~91 GeV), while the tagger is intended for mX = 20–60 GeV resonances, which populate smaller ΔR and softer pT. Figures 7(c) and 7(d) show the identification efficiency varies with ΔR and seed pT in simulation, so a single inclusive SF could miss a phase-space-dependent data/MC difference. The text says 'additional uncertainties may be required' but gives no quantitative extrapolation uncertainty. That is a real gap, but it is an acknowledged limitation of an intentionally statistically limited measurement, not a hidden flaw. Future analyses using the SF will have to assign their own extrapolation uncertainty, and the community will need this number measured in a closer phase space or derived with a closure test.\n\nWho this is for: experimentalists in ATLAS (and CMS, by analogy) working on low-mass BSM searches, especially exotic Higgs decays to light scalars with di-tau final states. It deserves a serious referee: the method is sound, the result is useful, and the limitations are honestly stated. The referee should push for a quantified extrapolation uncertainty or a clear prescription for deriving one, but the paper should not be desk-rejected. I would send it to peer review.","headline":"A solid, honestly limited ATLAS method paper: new low-pT di-tau tagger with a data-driven scale factor, but the inclusive Zgamma SF needs a quantified extrapolation uncertainty before use in light-resonance searches.","tokens_in":59143,"tokens_out":1430,"would_cite":true,"duration_ms":18525,"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":"Pairs of collimated, hadronically decaying tau leptons can be reconstructed and identified as a single di-tau object down to seed-jet $p_{\\mathrm{T}}$ of 50 GeV, with identification scale factors of $1.00\\pm0.37$ and $1.01\\pm0.27$…","keywords":["tau lepton pairs","hadronic tau decays","boosted di-tau tagger","large-radius jets","boosted decision tree","data-to-simulation scale factor","Z gamma tag-and-probe","light resonance searches"],"falsifier":"Compute the scale factor in the same $Z\\gamma$ sample in separate bins of seed-jet $p_{\\mathrm{T}}$ (for example below and above 150 GeV) or of the subjet angular separation $\\Delta R(\\mathrm{lead},\\mathrm{sublead})$; if the BDT response depends on parent-mass kinematics, the per-bin scale factors will deviate from unity by more than the quoted 26-37% total uncertainties, which the inclusive measurement cannot reveal. The corresponding simulation test is to compare truth-matched di-tau identification efficiencies at fixed reconstructed kinematics between $m_X=20$ GeV events and $Z$-mass events.","tokens_in":58024,"feed_emoji":"⚛️","tokens_out":14630,"duration_ms":135659,"temperature":0.7,"pith_summary":"When two tau leptons decay hadronically and their parent particle is light but boosted, the visible decay products can overlap so heavily that the standard reconstruction of each individual tau fails. This paper claims that such a collimated pair can instead be recovered as a single 'di-tau' object, built from a large-radius seed jet with two leading subjets, a dedicated energy-scale calibration, and a boosted-decision-tree identifier that separates genuine tau pairs from QCD jets. The validation in data is the central new step: the identification efficiency is measured in $Z(\\to\\tau\\tau)+\\gamma$ events from 139 fb$^{-1}$ of 13 TeV proton-proton collisions, yielding data-to-simulation scale factors of $1.00\\pm0.37$ at the Medium working point and $1.01\\pm0.27$ at the Tight working point. The scale factors being close to unity means the tagger's modelling is adequate for use, provided the efficiency measured at Z-boson kinematics transfers to the lighter resonances the tagger targets, such as Higgs decays to pairs of light pseudoscalars.","feed_headline":"Di-tau tagger scale factors 1.00 and 1.01 match data within 37%","feed_subtitle":"Verified in Zγ events at 139 fb−1, the tagger lets low-mass tau-pair searches reach seed jets down to 50 GeV.","key_machinery":"The central object is the di-tau object: an anti-$k_t$ seed jet of radius $R=1.0$ whose constituents are reclustered into $R=0.2$ subjets, with the two leading subjets assumed to carry the two tau signatures, each with a core cone of $R=0.1$, and the remaining seed-jet area acting as an isolation region. A two-step energy-scale calibration subtracts a pile-up term linear in the number of primary vertices, then applies a response function binned in subjet pseudorapidity and track multiplicity (1- or 3-prong). The identifier is a boosted decision tree on 16 variables drawn from the subjet cores, the subjet areas, the isolation region, and the whole seed jet, trained on truth-matched di-tau objects from $t\\bar{t}X$ events with $m_X = 20$ and 60 GeV against fake di-tau objects from fully hadronic $t\\bar{t}$ events. The efficiency measurement machinery is the $Z\\gamma$ tag-and-probe: a Tight photon selects the events, the BDT score is the discriminant, the $\\gamma$+jets and multijet backgrounds are normalised with an ABCD sideband method in the photon isolation-identification plane and their BDT shape reweighted to data, and the scale factor is the ratio of background-subtracted data to truth-matched signal in the signal region.","core_discovery":"The paper's central claim is that collimated, hadronically decaying tau pairs with seed-jet $p_{\\mathrm{T}}$ as low as 50 GeV can be reconstructed from the two leading $R=0.2$ subjets of an $R=1.0$ seed jet and identified against QCD jets by a boosted decision tree, and that the identification efficiency of this di-tau tagger agrees with data. The efficiency is measured with a tag-and-probe in $Z(\\to\\tau\\tau)+\\gamma$ events selected by a photon trigger with $E_{\\mathrm{T}}>140$ GeV, with the dominant $\\gamma$+jets and multijet backgrounds normalised by an ABCD sideband method and their BDT score shapes reweighted to data. In the signal region the measured efficiencies are about 70% (Medium) and 34% (Tight) at background rejections of about 240 and 3600, giving SF(BDT$>0.35$) $=1.00\\pm0.35\\,\\mathrm{(stat.)}\\pm0.13\\,\\mathrm{(syst.)}$ and SF(BDT$>0.5$) $=1.01\\pm0.24\\,\\mathrm{(stat.)}\\pm0.12\\,\\mathrm{(syst.)}$. The paper interprets these scale factors, both within 26-37% of unity, as demonstrating good data-to-simulation agreement for the di-tau object modelling, and presents the measurement as the step that allows the tagger to be used in future searches where resolved tau reconstruction is insufficient.","pith_inferences":["The inclusive scale factor is measured at Z-boson kinematics, and the size of the extrapolation to $m_X\\approx20$ GeV targets is left unquantified; binning the scale factor in seed-jet $p_{\\mathrm{T}}$ or in the subjet angular separation $\\Delta R(\\mathrm{lead},\\mathrm{sublead})$ within the existing $Z\\gamma$ sample would test whether the efficiency is flat where the tagger will actually be used.","Because the classifier is trained on an equal mixture of $m_X=20$ and 60 GeV events while the data probe sits at higher mass and boost, a simulation-level comparison of BDT score distributions at fixed reconstructed kinematics would separate mass-driven efficiency loss from phase-space-driven loss, informing how much of the Medium-to-Tight gap is intrinsic to the classifier.","The same seed-jet-plus-subjets structure and the pile-up and response calibration chain could be reused for other highly collimated two-body hadronic decays, such as a boosted $W\\to q\\bar{q}'$ or a heavy-flavour resonance, retraining the classifier while sharing the calibration."],"forward_implications":["Light-resonance searches, for example Higgs decays to pairs of light pseudoscalars with $m_X$ around 20-60 GeV, gain a validated identification efficiency for collimated tau pairs whose decay products overlap and defeat resolved tau reconstruction.","The Medium and Tight working points provide measured signal efficiencies of about 70% and 34% at background rejections of about 240 and 3600, so analyses can trade efficiency against rejection at a data-anchored rate.","Future studies using the tagger will apply the measured scale factors as event weights to simulated truth-matched di-tau objects, correcting residual data-to-simulation differences in the identification.","Analyses may combine the di-tau tagger with the standard resolved tau reconstruction to reject fake di-tau background, improving search sensitivities at the $p_{\\mathrm{T}}$ values relevant to light resonance searches.","The dominant uncertainty on both scale factors is statistical (35% and 24%), so additional collision data will directly sharpen the calibration."],"supporting_citations":[{"why":"Supplies the original seed-jet-plus-subjets reconstruction topology that this paper re-adapts to low seed-jet $p_{\\mathrm{T}}$; the prior tagger targeted $p_{\\mathrm{T}}>300$ GeV and lacked the dedicated energy-scale calibration.","marker":"[7]"},{"why":"Provides the track selection, track-to-vertex matching, and two-step energy-scale calibration procedure that are applied to each subjet of the di-tau object.","marker":"[68]"},{"why":"Defines the photon Tight identification and isolation working points used to select the $Z\\gamma$ events for the tag-and-probe measurement.","marker":"[56]"},{"why":"Specifies the lowest unprescaled single-photon trigger with $E_{\\mathrm{T}}>140$ GeV that seeds the event selection.","marker":"[72]"},{"why":"Is the source of the ABCD sideband-counting method used to normalise the $\\gamma$+jets and multijet backgrounds in the scale-factor extraction.","marker":"[75]"},{"why":"Supplies the 5% theoretical normalisation uncertainty assigned to the simulated signal and background event yields.","marker":"[74]"},{"why":"Motivates the $R=0.2$ subjet and $R=0.1$ core-cone definitions through the core-cone and isolation-annulus conventions for single taus.","marker":"[67]"}],"fun_headline_variants":["Di-tau tagger reconstructs collimated pairs down to 50 GeV","Scale factors 1.00 and 1.01 for di-tau tagger in Zγ data","Low-pT collimated tau pairs tagged with BDT and verified","Tau-pair tagger matches data within 37% at seed jets of 50 GeV","BDT di-tau tagger reaches 50 GeV with near-unity scale factors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The measurement assumes that the identification scale factor measured in $Z(\\to\\tau\\tau)+\\gamma$ events, where the tau pairs come from a Z boson of roughly 91 GeV, also holds for the much lighter resonances ($m_X$ near 20-60 GeV) that the tagger was built for, and the paper only notes that additional uncertainties may be required in other phase spaces without quantifying this extrapolation.","fun_headline_variants_meta":{"raw":{"variants":["Di-tau tagger reconstructs collimated pairs down to 50 GeV","Scale factors 1.00 and 1.01 for di-tau tagger in Zγ data","Low-pT collimated tau pairs tagged with BDT and verified","Tau-pair tagger matches data within 37% at seed jets of 50 GeV","BDT di-tau tagger reaches 50 GeV with near-unity scale factors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001996,"raw_usage":{"total_tokens":7865,"prompt_tokens":1095,"completion_tokens":6770,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":711,"completion_tokens_details":{"reasoning_tokens":6659}},"tokens_in":711,"tokens_out":6770,"duration_ms":46252,"temperature":1.0,"reasoning_tokens":6659,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:44:15.898727+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the scale factor in the same $Z\\gamma$ sample in separate bins of seed-jet $p_{\\mathrm{T}}$ (for example below and above 150 GeV) or of the subjet angular separation $\\Delta R(\\mathrm{lead},\\mathrm{sublead})$; if the BDT response depends on parent-mass kinematics, the per-bin scale factors will deviate from unity by more than the quoted 26-37% total uncertainties, which the inclusive measurement cannot reveal. The corresponding simulation test is to compare truth-matched di-tau identification efficiencies at fixed reconstructed kinematics between $m_X=20$ GeV events and $Z$-mass events.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the 5% theoretical normalisation uncertainty assigned to the simulated signal and background event yields."}],"review_version":1}