{"id":"e02d7ead-047b-44bb-98ba-a8d4b161de5a","arxiv_id":"2507.16686","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Ring-based boosted decision trees with an in situ pileup correction improve real-time jet trigger calibration, doubling simulated HH to bbbb signal efficiency at fixed background acceptance.","lead":"The paper trains boosted decision trees on calorimeter rings to estimate jet energy and reject pileup in real-time LHC triggers. In simulation at 200 simultaneous collisions, the method doubles the trigger efficiency for Higgs pair production decaying to four b-quarks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported factor-of-two gain may be inflated because the Fig. 6 baseline 'primitive jets' are compared without the per-bin calibration that the paper itself introduces in Fig. 4.","rationale":"The reader's CONDITIONAL verdict and the mention of a missing 'strong baseline' are in the right direction, but the specific load-bearing weakness is more concrete: the paper's own Fig. 4 establishes a simple calibration for primitive jets, yet the trigger comparison in Fig. 6 appears to use raw E_prim^T. This is an internal fairness issue, not a claim about external simulation realism. If the primitive baseline is uncalibrated, the 13.5-14.5 GeV shift is not cleanly attributable to the ML method; part of the gain may come from comparing against a control arm that omits an obvious calibration already demonstrated in the paper. This concern is directly testable with the authors' existing simulated samples and code, and it would not be resolved by switching to a more realistic detector simulation. Conversely, if the primitive-jet curves in Fig. 6 are in fact calibrated, the concern is resolved. The Delphes-fidelity issue raised by the reader is real and broader, but this baseline-calibration ambiguity is more tightly coupled to the central numerical claim and should be settled first. For that reason I would keep the reader's CONDITIONAL verdict unchanged: the work is promising and internally reproducible, but the headline factor-of-two needs a clarified and fair control comparison before it can be considered established.","tokens_in":11539,"tokens_out":16981,"duration_ms":201534,"concrete_test":"Reproduce Fig. 6 with the primitive-jet baseline calibrated exactly as in Fig. 4: replace E_prim^T by E_prim^T / R_prim(bin) before applying the fourth-jet trigger threshold, then recompute the 1% and 5% dijet-acceptance thresholds and the HH4b efficiency curves. If the 1%-acceptance threshold shift drops below about 10 GeV, or if the 80%-efficiency signal acceptance ratio falls below about 1.5, the headline factor-of-two is partly an artifact of comparing ML jets to uncalibrated jets.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central claim (factor of two in HH4b signal efficiency) is derived from the equal-background-rate turn-on curves in Fig. 6. The baseline is described as the 'E_prim^T-based trigger' and labeled 'Prim. jets', but the Results section never states that the per-bin calibration developed for primitive jets in Fig. 4 is applied to those trigger curves. In Fig. 4, raw primitive jet transverse energies are divided by the mode R_prim of the E_prim^T/E_offline^T distribution in each E_prim^T bin, with R_prim as large as about 3.5 in the lowest bin and about 1.3 in the 90-120 GeV region. If such a calibration is omitted for the primitive-jet baseline, that control arm is artificially weak: pileup-dominated low-ET jets contribute excess background rate, forcing a higher trigger threshold. The ML arm, by contrast, receives both an explicit regression calibration (BDT_ET) and a pileup-rejection weight (p_HS). Comparing ML jets against uncalibrated primitive jets therefore conflates the new method with the absence of a standard calibration step. The claimed 13.5-14.5 GeV threshold shift and factor-2.1 acceptance gain may be partly a property of this weak baseline rather than of the ML itself.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an FPGA-oriented ML jet-energy calibration for real-time L1 triggers at the HL-LHC. Calorimeter-tower energies in four concentric rings around a sliding-window primitive jet seed are fed to two boosted decision trees: a regressor (BDT_ET) that estimates the jet transverse energy on the offline anti-kT scale, and a classifier (BDT_HS) whose score is converted into a hard-scatter probability p_HS. The final trigger quantity is E_ML^T = p_HS * BDT_ET (Eq. 1). Using Delphes ATLAS-card simulations of dijet and HH->bbbb events with <mu>=200 pileup, the authors report improved energy resolution, a 13.5-14.5 GeV reduction in the fourth-jet threshold at fixed dijet background acceptance, and a factor-2.1 increase in HH4b signal efficiency at 80% efficiency relative to primitive cone jets.","tokens_in":11810,"tokens_out":5474,"duration_ms":66376,"significance":"If the headline result holds, the paper is a useful proof-of-principle for real-time ML jet calibration with in situ pileup correction. The ring-based feature design is simple and plausibly FPGA-friendly, the training/evaluation workflow is clearly documented, the simulated samples are made public (Ref. [45]), and the authors cite prior work showing that similar BDTs fit within FPGA resource and latency budgets. The quantitative claims, however, rest on a comparison baseline whose definition is ambiguous, and the results are presented without statistical or systematic uncertainties. The core methodological idea is worth publishing, but the central factor-of-two claim needs to be re-established against a properly calibrated primitive-jet baseline.","major_comments":[{"comment":"It must be stated explicitly whether the \"Prim. jets\" curves in Fig. 6 use the raw E_prim^T or the per-bin calibrated E_prim^T/R_hat introduced in Fig. 4. The paper says in the Machine Learning section that \"to compare the performance, we calibrate each incoming ET by the R_hat derived above,\" but the Results section never says this calibration is applied to the primitive-jet baseline. If Fig. 6 uses uncalibrated primitive jets, the comparison is not like-for-like: the ML arm receives both a regression calibration and a p_HS pileup correction, while the baseline is denied the simple calibration step that Fig. 4 itself demonstrates. In that case the claimed 13.5-14.5 GeV threshold shift and factor-2.1 acceptance gain may be partly an artifact of an artificially weak control arm. The authors should add curves for the calibrated primitive jets (E_prim^T divided by R_hat in each bin) and report the threshold shift and acceptance gain relative to that baseline.","section":"Results, Fig. 6"},{"comment":"The quantitative claims (factor-2.1 acceptance gain, 13.5-14.5 GeV threshold shifts, 33% and 57% resolution gains in Fig. 4) are presented without statistical uncertainties or a systematic-error budget. With 100k simulated events, the 80% efficiency points carry statistical errors that may be small, but the simulation uses a single Delphes ATLAS card, a fixed <mu>=200, and specific choices for the matching radius (0.3), ring boundaries, BDT hyperparameters, and p_HS calibration binning. A sensitivity scan over these choices, or at least a table of variations around the nominal values, is needed to establish that the advertised factor-of-two improvement is robust rather than a point estimate from one simulation setup. This is load-bearing because the abstract and conclusions quote the factor of two as the main result.","section":"Results, Fig. 6"},{"comment":"The real-time applicability of the exact proposed algorithm is asserted but not demonstrated. The paper cites previous FPGA implementations of BDTs with \"parameters similar to those in our paper,\" but the present design includes two BDTs (30 trees, depth 8), a ring-feature summation stage, a p_HS probability lookup/normalization, and the final multiplication in Eq. (1). No resource estimate, latency budget, or throughput analysis is given for this specific combination, nor for the surrounding sliding-window primitive-jet finding. The conclusion that the method \"can lower the fourth leading jet ET by 10-15 GeV\" in a real L1 trigger therefore goes beyond what the paper establishes. The authors should either provide a concrete FPGA resource/latency estimate for this model or soften the hardware claims to say the approach is compatible with previously demonstrated FPGA BDT performance.","section":"Machine learning, last paragraph; Conclusions"}],"minor_comments":[{"comment":"The caption line \"5%, ML jets\" and \"1%, Prim. jets\" is clear, but the embedded text \"dijet accept, alg, gain\" is cryptic; please spell out what \"alg\" and \"gain\" refer to.","section":"Fig. 6 caption"},{"comment":"The horizontal axis label \"R/R_hat\" may be confused with the jet radius R=0.4 used elsewhere; consider renaming the resolution variable (e.g., \"E_T^est/E_T^offline\" after calibration) to avoid ambiguity.","section":"Fig. 4"},{"comment":"The primitive-jet seed threshold of E_seed^T > 20 GeV in the sliding-window algorithm could influence the low-energy turn-on behavior; please state how this threshold was chosen and whether the reported trigger curves are sensitive to it.","section":"Setup, footnote 5"},{"comment":"The phrase \"efficiency of saving Higgs pair production\" is awkward; consider \"efficiency for selecting\" or \"retaining\" Higgs pair production events.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid proof-of-principle, but the headline factor-of-two gain is not yet established because the baseline primitive-jet arm may be uncalibrated. This is fixable within the manuscript's scope by adding calibrated primitive-jet curves and a quantitative uncertainty statement. I would not recommend rejection, since the ring-based ML method and its in situ pileup correction are novel and potentially valuable for L1 trigger design."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper has a real new idea—the ring-based BDT regression plus a pileup-probability weighting—and it is clearly written. The headline factor-of-two gain is not yet established, because the baseline in Fig. 6 is ambiguous: it may be raw, uncalibrated primitive jets, which would inflate the gain.\n\nWhat's actually new: the concentric 0.1x0.1 tower ring features around a fixed-radius cone seed, the combination of a regression (BDT_ET) and a classifier-weight (p_HS), and the target application to HH4b at <mu>=200. Figures 1-5 are informative; the BDT_ET response is flat across E_prim bins and the resolution gain in the 69-93 GeV bin is 33%. That part looks sound.\n\nThe soft spots are in proportion. First, the trigger comparison is under-specified. In Fig. 4 the authors calibrate primitive jets by the per-bin mode R_prim and explicitly say the uncalibrated comparison 'does not reflect the performance.' But the Results/Fig. 6 never says whether that calibration was applied to the Prim. jet turn-on curves. 'Eprim_T-based trigger' reads like raw Eprim. If raw, the 13.5-14.5 GeV threshold shift and factor 2.1 partly come from comparing against an uncalibrated baseline, and the abstract's 'factor of two over the traditional cone jet algorithm' is not supported. If it was calibrated, the paper needs to say so. This is the one issue that really matters.\n\nSecond, there are no statistical or systematic uncertainties, and no full detector simulation or actual FPGA implementation; feasibility is inherited from self-cited prior work. That is a limitation, not a fatal flaw, but it should be stated more carefully.\n\nThe circularity worry is minor: regressing to the offline anti-kt scale and evaluating against that same scale is standard for trigger calibration, and a held-out test split is used. The bigger issue is simply that the baseline comparison must be defined.\n\nWho should read it: anyone working on L1 jet triggers or ML inference in FPGAs. It deserves a serious referee, but the revision needs to clarify the baseline and add uncertainties before the central claim can be taken at face value.","headline":"Plausible and well-written ML jet calibration, but the factor-of-two HH4b gain hinges on an ambiguous baseline that may be uncalibrated primitive jets.","tokens_in":12335,"tokens_out":6031,"would_cite":true,"duration_ms":65280,"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":"Ring-based ML jet calibration in the first-level trigger doubles the retained HH→bbbb signal by lowering the fourth-jet threshold from about 90 GeV to 75 GeV.","keywords":["jet energy calibration","boosted decision trees","FPGA trigger","pileup mitigation","HL-LHC","Higgs pair production","HH4b","real-time trigger"],"falsifier":"Recompute the fourth-jet turn-on curve using a full, detailed detector simulation of the HL-LHC calorimeter trigger (or early real data at comparable pileup) fed through the trained two-BDT model; if the 80% HH4b efficiency threshold does not move from about 90 GeV down to about 75 GeV at 1% dijet acceptance, the claimed factor-of-two signal gain does not hold. A supporting check is whether the exact 30-tree, depth-8 BDT pair fits in the target FPGA at the required latency.","tokens_in":11328,"feed_emoji":"⚛️","tokens_out":14134,"duration_ms":133455,"temperature":0.7,"pith_summary":"The paper claims that a jet-energy calibration built from boosted decision trees can run in the first-level trigger of the HL-LHC and roughly double the amount of retained Higgs-boson-pair events decaying to four b-quarks. The method starts from the simple fixed-radius cone jets a hardware trigger can form, reads the electromagnetic and hadronic energy in four concentric rings around each jet, and combines a regressor that estimates the jet's true transverse energy with a classifier that scores how likely the jet comes from the hard scatter rather than from pileup. The classifier score is used as a weight on the regression output, so pileup-dominated jets are pushed down in the trigger ranking. If the calibration performs as demonstrated, it lowers the required fourth-jet threshold from about 90 GeV to about 75 GeV at fixed background rate, which is the factor-of-two gain in signal.","feed_headline":"ML jet calibration doubles kept Higgs-pair events","feed_subtitle":"First-level jet threshold drops from 90 to 75 GeV, doubling HH->bbbb acceptance at fixed background rate.","key_machinery":"The central object is the ring-based jet representation: each primitive jet (a fixed-radius cone from a sliding-window seed) is described by the electromagnetic, hadronic, and total tower energy summed in four concentric annuli of 0.1×0.1 cells, with $\\Delta R < 0.1$, $0.1 \\le \\Delta R < 0.2$, $0.2 \\le \\Delta R < 0.4$, and $0.4 \\le \\Delta R < 0.6$, plus the inner three rings combined as the 'ring jet' (the primitive jet itself). The outer rings are designed to measure the local pileup density in situ, while the inner rings capture the hard-scatter core. These inputs feed two boosted decision trees: a regressor (BDT$_\\mathrm{ET}$) trained to reproduce the offline anti-$k_t$ jet $E_T$, and a classifier (BDT$_\\mathrm{HS}$) trained to separate matched hard-scatter jets from unmatched pileup jets. The two outputs are combined as $E^T_\\mathrm{ML} = p_\\mathrm{HS}\\cdot\\mathrm{BDT}_\\mathrm{ET}$, which is the quantity a trigger would threshold on.","core_discovery":"On its own terms, the paper establishes that a pair of boosted decision trees trained on ring sums of calorimeter tower energy can correct the energy scale of primitive cone jets and simultaneously suppress pileup jets, and that the product $E^T_\\mathrm{ML} = p_\\mathrm{HS}\\cdot\\mathrm{BDT}_\\mathrm{ET}$ provides a trigger quantity with a sharper turn-on than the raw primitive jet $E_T$. The regression BDT targets the offline anti-$k_t$ jet $E_T$; the classifier BDT outputs a hard-scatter probability $p_\\mathrm{HS}$ calibrated by binning the score; multiplying the two yields the final jet quantity. In the HH4b benchmark, the 80% signal-efficiency threshold on the fourth leading jet moves from about 90 GeV to about 75 GeV at 1% dijet acceptance, a shift of 14.5 GeV, and the signal acceptance grows by a factor of 2.1. The paper also reports that the resolution gain is largest in the calibrated-$E_T$ bin of about 41–69 GeV, where the rms improves by 57% relative to primitive jets.","pith_inferences":["A single combined model that outputs both the energy estimate and the hard-scatter probability might achieve the same gain with a smaller firmware footprint; the paper's two-BDT split is a hardware-motivated design choice, not a proven necessity.","The ring features could also serve as a standalone online pileup-jet veto, a use beyond the $E^T_\\mathrm{ML}$ threshold quantity.","Applying the calibration to all jets in the event rather than only the fourth could produce an even steeper combined trigger turn-on, at the cost of more processing per event.","The method's sensitivity to the tower granularity is untested; a coarser trigger tower geometry would change the ring definitions and could raise or remove the factor-of-two gain."],"forward_implications":["A first-level trigger using this calibration can keep events whose fourth b-jet has $E_T$ near 75 GeV instead of requiring near 90 GeV, directly enlarging the dataset for measuring the Higgs self-coupling.","The same ring-based calibration can be applied to other multi-jet trigger signatures, such as vector-boson-fusion Higgs production or four-top final states, where low jet thresholds are the limiting factor.","Because the pileup correction is local to the jet, the method may relax the requirement on a global pileup-density measurement in the hardware trigger.","The reported 57% resolution gain at $E_T \\sim 41{-}69$ GeV implies the method becomes more valuable if trigger systems can accept lower jet thresholds in the future.","The two-BDT structure is modular: the energy regressor and the hard-scatter tagger can be retrained independently if the pileup environment or detector geometry changes."],"supporting_citations":[{"why":"Provides the fast detector simulation that produces the calorimeter tower energies used for all samples.","marker":"[51]"},{"why":"Defines the sliding-window seed algorithm that creates the primitive jets used as base objects.","marker":"[36]"},{"why":"Introduces the pileup-density estimate that motivates the outer rings as local pileup probes.","marker":"[22]"},{"why":"Supplies the reference trigger system and the 90 GeV fourth-jet threshold benchmark.","marker":"[12]"},{"why":"Demonstrates that boosted decision tree classifiers can run on FPGAs at nanosecond latency with small resource usage.","marker":"[41]"},{"why":"Demonstrates the same feasibility for boosted decision tree regression.","marker":"[42]"},{"why":"Provides the training toolkit used to build the two boosted decision trees.","marker":"[56]"},{"why":"Generates the hard-scatter dijet and Higgs-pair samples used for training and testing.","marker":"[49]"},{"why":"Generates the pileup overlays and parton shower that define the high-pileup environment.","marker":"[50]"}],"fun_headline_variants":["FPGA jet ML doubles HH->bbbb trigger yield","Jet ML trigger keeps twice the HH events","ML trigger calibration cuts jet threshold to 75 GeV","FPGA jet ML: double HH acceptance at same background","HL-LHC jet trigger ML: two-fold HH gain"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire performance evaluation assumes that the fast detector simulation with its simplified sliding-window primitive jet algorithm faithfully represents the HL-LHC first-level calorimeter trigger environment at 200 simultaneous proton-proton collisions per bunch crossing.","fun_headline_variants_meta":{"raw":{"variants":["FPGA jet ML doubles HH->bbbb trigger yield","Jet ML trigger keeps twice the HH events","ML trigger calibration cuts jet threshold to 75 GeV","FPGA jet ML: double HH acceptance at same background","HL-LHC jet trigger ML: two-fold HH gain"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00065,"raw_usage":{"total_tokens":2980,"prompt_tokens":941,"completion_tokens":2039,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":1962}},"tokens_in":557,"tokens_out":2039,"duration_ms":15122,"temperature":1.0,"reasoning_tokens":1962,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:04:19.507812+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the fourth-jet turn-on curve using a full, detailed detector simulation of the HL-LHC calorimeter trigger (or early real data at comparable pileup) fed through the trained two-BDT model; if the 80% HH4b efficiency threshold does not move from about 90 GeV down to about 75 GeV at 1% dijet acceptance, the claimed factor-of-two signal gain does not hold. A supporting check is whether the exact 30-tree, depth-8 BDT pair fits in the target FPGA at the required latency.","supporting_citations":[{"cited_title":"High Energy Phys","cited_arxiv_id":null,"evidence_quote":"Provides the fast detector simulation that produces the calorimeter tower energies used for all samples."},{"cited_title":"Mehdiyev, Z","cited_arxiv_id":null,"evidence_quote":"Defines the sliding-window seed algorithm that creates the primitive jets used as base objects."},{"cited_title":"Cacciari and G","cited_arxiv_id":null,"evidence_quote":"Introduces the pileup-density estimate that motivates the outer rings as local pileup probes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the reference trigger system and the 90 GeV fourth-jet threshold benchmark."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Demonstrates that boosted decision tree classifiers can run on FPGAs at nanosecond latency with small resource usage."},{"cited_title":"Carlson, Q","cited_arxiv_id":null,"evidence_quote":"Demonstrates the same feasibility for boosted decision tree regression."},{"cited_title":"Alwall, M","cited_arxiv_id":null,"evidence_quote":"Generates the hard-scatter dijet and Higgs-pair samples used for training and testing."},{"cited_title":"Sj¨ ostrandet al","cited_arxiv_id":null,"evidence_quote":"Generates the pileup overlays and parton shower that define the high-pileup environment."}],"review_version":1}