{"id":"558ca7e3-8408-43f3-9ae5-742d5e33f5e9","arxiv_id":"2602.20475","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":11,"one_line_summary":"PhyGHT, a graph-transformer hybrid with a learned pileup-suppression gate, reports state-of-the-art jet energy and mass correction on simulated HL-LHC pileup.","lead":"PhyGHT is a machine-learning model that combines graph attention, transformers, and a learned pileup-suppression gate to strip away the roughly 200 extra proton collisions expected at the HL-LHC and recover jet energy and mass. On a new simulated top-quark dataset, it reports higher accuracy than existing pileup-mitigation baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-set hyperparameter selection and single-seed evaluation leave the headline R2 gain over PUMINet unvalidated.","rationale":"The reader's chosen weakest assumption is that the particle-level simulation is a faithful proxy for HL-LHC data. That is a genuine external-validity limitation, and Section 7 acknowledges it. However, it is a scope limitation rather than a flaw in the internal benchmark: the paper's strongest quantitative claim is about performance on the released simulated dataset, not on real collision data. A more load-bearing concern is whether the comparison itself is trustworthy. The paper selects PhyGHT's two key hyperparameters using Tables 4-5, which are not labeled as validation results, and uses a single seed with no uncertainty quantification. The closest baseline gap at <mu>=200 is 0.006 in energy R2, which could easily be within run-to-run variation. If the hyperparameters were chosen on the test split, the reported superiority is partially an artifact of test-set selection. This is concrete, testable from the released code, and directly bears on the headline 'consistently outperforms.' I therefore keep the reader's CONDITIONAL verdict unchanged rather than moving to ACCEPT or REJECT: the architecture is plausible and the ablations are internally consistent, but the central comparison needs validation-only model selection and multi-seed error bars before the superiority claim is established.","tokens_in":16972,"tokens_out":8096,"duration_ms":86045,"concrete_test":"Check the released code/training logs to determine whether Tables 4-5 were produced on the validation split or the test split. If they used test, re-run the PhyGHT vs PUMINet comparison with the test set held out: select k and lambda_aux on validation only, train with at least 5 seeds, and report mean±std R2 on test. If the <mu>=200 energy gap (0.006) exceeds the combined seed std, the headline gain is supported; if not, the advantage is within selection noise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that PhyGHT consistently outperforms all baselines on the released simulated benchmark. The most load-bearing condition for that claim is a fair, statistically stable comparison. The paper does not yet establish this. Section 4.1.3 states that k=8 and lambda_aux=0.1 were chosen 'based on the hyperparameter analysis in Section 4.3.2.' Tables 4 and 5 report R2 curves for these hyperparameters but never state whether they are computed on the validation split or the same 10% test split used in Table 1. If they are test-set curves, then the model's hyperparameters were selected using the test set, so the Table 1 R2 values are optimistic and the comparison to baselines (which are not shown to have tuned hyperparameters) is not a level field. Appendix B reports a single random seed (42), so there are no error bars. At <mu>=200 the energy R2 gain over PUMINet is only 0.932 vs 0.926 (0.006), and the mass gain is 0.836 vs 0.805; without seed-to-seed variance one cannot tell whether either difference is real. The detector-realism limitation is important but secondary; even granting the particle-level simulation as the target environment, the evidence for 'consistently outperforms' is not yet secured.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PhyGHT, a hybrid architecture for pileup mitigation at the HL-LHC. The model combines distance-aware graph attention for local geometry, a transformer encoder for global event context, a learnable pileup suppression gate (PSG) that soft-masks tracks, and a hypergraph attention block that aggregates purified tracks to per-jet predictions of energy and mass correction factors. The authors introduce a public simulated t-tbar dataset at <mu>=60 and <mu>=200, benchmark against several baselines including PUPPI, ParticleNet and PUMINet, report improved R2 values, present ablations, and demonstrate a top-quark mass reconstruction as a downstream physics task. The core technical derivation is coherent, and the released dataset/code are a concrete contribution.","tokens_in":17215,"tokens_out":5690,"duration_ms":51998,"significance":"If the reported results are statistically robust, the architecture and the public particle-level dataset would be a useful stepping stone for ML-based pileup mitigation at the HL-LHC. The design idea of making the pileup filter explicit and interpretable via the PSG, while keeping a hierarchical track-to-jet aggregation, is well motivated and goes beyond generic graph/transformer baselines. The paper also contains a useful ablation analysis showing the relative contribution of each block. However, the evaluation currently rests on single-seed runs, unclear train/validation/test discipline for hyperparameter selection, and several oracle-in-the-loop elements in the physics demonstrations. These issues need to be resolved before the central claim of consistent superiority over baselines can be accepted.","major_comments":[{"comment":"Section 4.1.3 states that k=8 and lambda_aux=0.1 were chosen 'based on the hyperparameter analysis in Section 4.3.2.' Section 4.3.2 (Tables 4 and 5) reports R2 values but does not state which split is used. If these tables are evaluated on the same 10% test split that produces Table 1, then the test set has been used for model selection, and the Table 1 R2 values are optimistically biased. Moreover, the baselines (PUMINet, ParticleNet, etc.) are not reported to have been tuned on the same basis, so the comparison is not on a level field. Please clarify the split used for hyperparameter selection and, if it is the test split, re-run selection on the validation split or use nested cross-validation.","section":"Section 4.1.3, Tables 4-5"},{"comment":"Appendix B reports 'a random seed of 42 for reproducibility' and all experiments appear to be single-seed runs. Table 1 shows PhyGHT's advantage over PUMINet at <mu>=200 is 0.932-0.926=0.006 in energy R2 and 0.836-0.805=0.031 in mass R2. Without seed-to-seed variance (or any uncertainty estimate), it is not possible to assess whether this difference is significant. This directly affects the claim of 'consistently outperforms all baselines' (Section 4.2.1). Please run at least 5 seeds for PhyGHT and the closest baselines and report mean +/- std (or confidence intervals).","section":"Appendix B and Table 1"},{"comment":"Appendix D: the top-quark mass reconstruction selects W1, W2, and B jets by tracing the parton shower history and using truth labels. Appendix A.8/E: the PUPPI baseline constructs its chi2 metric 'using truth labels to identify pileup contributions.' These oracle uses of generator-level information mean that (i) the downstream physics study does not demonstrate how PhyGHT would perform with a realistic candidate-selection procedure, and (ii) the PUPPI comparison is not a comparison to standard PUPPI as deployed in ATLAS/CMS. Please either replace truth-based selection with reconstructed-object selection, or explicitly frame these as oracle-assisted feasibility checks and adjust the wording in Sections 4.4.1 and 5.","section":"Appendix D and Appendix E"},{"comment":"Section 2.1 describes a particle-level simulation: neutral particles and charged tracks with pT<400 MeV are removed, vertices are Gaussian-smeared, and no calorimeter response, reconstruction efficiency, fake tracks, or other detector effects are included. Section 7 states 'Our next step is to integrate the model into the ATLAS software framework to process real collision data with full detector effects.' This contradicts the claim in Section 2 that the dataset 'closely mimic[s] the real-world detector data,' and the claim in Section 4.4.1 that the model 'can be used in real world physics analysis.' Please temper these claims to a particle-level benchmark and state the simulation limitations more prominently.","section":"Section 2.1 and Section 7"}],"minor_comments":[{"comment":"Typo: 'auxillary' should be 'auxiliary'.","section":"Section 3.7.2"},{"comment":"PUMINet and ParticleNet are described as 'state-of-the-art baselines from the ATLAS and CMS experiments,' but PUMINet is not an ATLAS/CMS production algorithm; rephrase to 'baselines proposed in the literature.'","section":"Section 4.1.2 and Table 1"},{"comment":"Minor language issues: 'Heavyside' should be 'Heaviside'; 'derived an R2 score' is acceptable but consider 'computed an R2 score.'","section":"Appendix E"},{"comment":"Inconsistent spacing in '⟨𝜇⟩=60(left)' and '⟨𝜇⟩=200(right)'; add spaces for readability.","section":"Figure 2 caption"},{"comment":"The manuscript ends with 'Received 20 February 2007; revised 12 March 2009; accepted 5 June 2009' — this is a template artifact and should be removed.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for KDD and the dataset release is a genuine asset. My main concern is statistical rigor: the hyperparameter/test-set ambiguity and single-seed results need to be fixed. I see no indication of intentional misreporting; the issues appear to be fixable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the thing to know: this is a solid, incremental HEP-ML paper with a genuinely useful public dataset and a clean architecture write-up, but the headline claim is not yet statistically supported. The main empirical result rests on a single seed and on hyperparameters apparently selected on the test set. What is new: PhyGHT combines k-NN distance-aware graph attention, a global transformer, a learned PUPPI-inspired gating module, and hypergraph aggregation. Each ingredient exists in the literature, but the combination is new, and releasing a public top-pair simulation at <mu>=60 and 200 is a real service. The ablations are informative: the global context block matters most, the local block matters for mass, and the gate gives a clear boost. The latency numbers are also credible and the paper is well written. The architecture math is coherent and I found no internal error. Where it falls short: the evidence for the central comparative claim is fragile. Section 4.1.3 picks k=8 and lambda_aux=0.1 based on the analysis in Section 4.3.2, but Tables 4 and 5 never say whether those R2 curves are from the validation split or the same 10% test split used in Table 1. If they are test-set curves, the gains are optimistically biased. Second, Appendix B reports a single seed (42) with no error bars; the energy R2 margin over PUMINet at <mu>=200 is 0.006, which could easily be seed noise. Third, the dataset is particle-level with no full detector effects, and the paper's own Limitations section says full detector integration is the next step—so the abstract's HL-LHC discovery language overreaches. Fourth, the PUPPI baseline and top-mass reconstruction use oracle truth labels (Appendices D and E), which weakens those comparisons. None of these is fatal to the architecture, but together they mean 'consistently outperforms all baselines' is not yet demonstrated. Who is this for: HEP-ML practitioners and anyone wanting an open benchmark for pileup mitigation. It deserves a serious referee, but the revision needs validation-set model selection, multiple seeds with error bars, a non-oracle PUPPI comparison, and an honest statement that particle-level results are a proxy until full-detector work is done. I would engage with it.","headline":"A useful public benchmark and a well-described architecture, but the 'consistently outperforms' claim needs validation-set tuning and error bars before it is credible.","tokens_in":17827,"tokens_out":2389,"would_cite":false,"duration_ms":23915,"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 claims that a physics-guided hypergraph transformer with a learnable pileup-suppression gate recovers jet energy and mass under extreme HL-LHC pileup better than all eight tested baselines, and reconstructs the top-quark mass reso","keywords":["pileup mitigation","hypergraph transformer","graph attention","physics-guided machine learning","top quark mass reconstruction","HL-LHC","jet energy correction","open simulated dataset"],"falsifier":"A concrete check: evaluate PhyGHT and PUMINet on a full detector simulation (including calorimeter response, fake tracks, and reconstruction) of the same ttbar events; if PhyGHT's R^2 advantage over PUMINet shrinks below the reported margin or reverses, the central claim as stated for HL-LHC conditions would fail. A second check: rerun PUPPI with generator-level filtering on the hard-scatter events, as the original method did; if its mass R^2 at <mu>=200 rises substantially above the reported 0.114, the claim that PhyGHT outperforms the physics baseline is weakened.","tokens_in":16758,"feed_emoji":"⚛️","tokens_out":10285,"duration_ms":87945,"temperature":0.7,"pith_summary":"The paper seeks to establish that a single neural architecture, PhyGHT, can remove contamination from up to two hundred simultaneous proton-proton collisions and recover the energy and mass of signal jets. It does this by predicting, for each jet, two bounded correction factors—the fraction of the raw jet energy and mass that originated from the hard-scatter collision—using a hybrid of local distance-aware graph attention, global transformer context, an explicit learnable soft-mask called the Pileup Suppression Gate, and hypergraph attention for aggregation. On a newly released simulated dataset of top-quark pair production, PhyGHT beats all eight baselines, including physics-based production algorithms and prior ML methods, with the largest gains in mass reconstruction; it also reconstructs the top-quark mass resonance in close agreement with truth. If the claims hold, pileup mitigation at the HL-LHC could become a two-number correction per jet, computed faster than existing approaches.","feed_headline":"Gated graph-transformer beats all eight pileup baselines","feed_subtitle":"Restores the top-quark mass peak and runs faster than prior ML methods even at 200 pileup collisions.","key_machinery":"The Pileup Suppression Gate (PSG) is the central object: a learnable, differentiable soft-mask that maps a track's fused local-and-global embedding to a scalar signal probability via a small MLP, then multiplies the embedding by that probability. It is trained with auxiliary track-level labels (hard-scatter vs pileup) so the mask tracks physical vertex association, and it is applied before the hypergraph aggregation stage. Around it, PhyGHT stacks a Distance-Aware Graph Attention layer that biases neighbor weights by a learnable spatial-decay parameter (a learned 'soft cone' in eta-phi space), a Transformer encoder for event-wide pileup density, and a bipartite/hypergraph attention that per","core_discovery":"The central discovery, on the paper's own benchmark, is that PhyGHT predicts per-jet energy and mass correction factors with test-set R^2 of 0.943 and 0.869 at a mean of 60 pileup collisions per event, and 0.932 and 0.836 at 200—consistently above all baselines. The decisive component, per the ablations, is the Pileup Suppression Gate: a differentiable soft-mask that estimates the probability each track comes from the hard-scatter vertex and multiplies that track's features by the probability before aggregation. Removing this gate costs several points of R^2 depending on the target, while removing the global transformer context costs more than any other single component. After applying the p","pith_inferences":["The paper does not compare the PSG's learned track weights to PUPPI's physics-derived weights on the same events; a direct calibration study could show whether the learnable gate recovers the same signal/pileup discrimination or something qualitatively different.","Because the model predicts only two scalar fractions per jet, it discards information about how the hard-scatter energy is distributed within the jet; extending the same architecture to per-subjet or per-momentum-component corrections could further improve mass resolution and is a natural, untested follow-up.","The strong degradation of PUPPI at <mu>=200 (mass R^2 0.114) is flagged by the authors as possibly due to missing generator-level filtering in their dataset; if that is the cause, the gap between physics-based and learned mitigation at extreme pileup may be smaller than reported. This is the kind of detail a skeptical reader should check.","The architecture is described as generalizable to any domain with dense local signal clusters in global noise (point clouds, anomalous communities), but the paper states this without testing it; a cheap transfer experiment on a non-physics benchmark would indicate whether the claim is specific to collider geometry or general."],"forward_implications":["If the central claim is correct, pileup mitigation becomes a per-jet two-number correction: multiply the raw jet energy and mass by the predicted fractions, and downstream physics analyses proceed.","Because the gains hold as pileup rises from 60 to 200 (mass R^2 falls only 0.033), the method is positioned for the HL-LHC's extreme environment rather than just current conditions.","The top-quark mass reconstruction, a standard physics benchmark, shows the corrections are physically usable, not just good on a regression metric.","PhyGHT does this at lower inference latency than the strongest baselines (about 1.9x faster at <mu>=200), which matters for offline reprocessing of the enormous HL-LHC dataset.","The released simulated dataset and code allow independent reproduction, lowering the barrier for other groups to develop and compare pileup-mitigation methods."],"fun_headline_variants":["Hybrid graph-transformer tames 200-pileup chaos at HL-LHC","Physics-gated transformer beats ATLAS, CMS in mass reconstruction","New AI model restores top-quark mass amid extreme pileup","Pileup Suppression Gate key to transformer's physics win","Graph-transformer with physics gate outperforms all baselines"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the simulated, particle-level dataset—built by overlaying Poisson-distributed min-bias collisions, smearing vertices with Gaussians, and discarding neutral particles and low-pT tracks—captures the pileup contamination a real detector will see; if detector-level effects such as fake tracks, calorimeter response, and reconstruction inefficiencies break this proxy, the reported gains may not transfer to actual collision data.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid graph-transformer tames 200-pileup chaos at HL-LHC","Physics-gated transformer beats ATLAS, CMS in mass reconstruction","New AI model restores top-quark mass amid extreme pileup","Pileup Suppression Gate key to transformer's physics win","Graph-transformer with physics gate outperforms all baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001465,"raw_usage":{"total_tokens":5758,"prompt_tokens":800,"completion_tokens":4958,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":4867}},"tokens_in":544,"tokens_out":4958,"duration_ms":31225,"temperature":1.0,"reasoning_tokens":4867,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T21:19:20.070696+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete check: evaluate PhyGHT and PUMINet on a full detector simulation (including calorimeter response, fake tracks, and reconstruction) of the same ttbar events; if PhyGHT's R^2 advantage over PUMINet shrinks below the reported margin or reverses, the central claim as stated for HL-LHC conditions would fail. A second check: rerun PUPPI with generator-level filtering on the hard-scatter events, as the original method did; if its mass R^2 at <mu>=200 rises substantially above the reported 0.114, the claim that PhyGHT outperforms the physics baseline is weakened.","supporting_citations":[],"review_version":1}