{"id":"a60b8f38-1819-4bc3-8ebc-f465073097f9","arxiv_id":"2509.02708","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Generative unfolding is extended to handle backgrounds, acceptance, and efficiency effects in an unbinned, iterative pipeline, demonstrated at percent-level accuracy on Gaussian and Z+jets simulations.","lead":"Two new algorithms, GenFoldC and GenFoldG, extend generative machine learning unfolding to also remove backgrounds and correct for acceptance and efficiency losses, not just detector resolution. The authors show on Gaussian and simulated LHC jets that both reach percent-level agreement with known truth, a step toward using generative unfolding in real measurements.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Acceptance correction is prior-dependent and never iterated; the main benchmark sidesteps this by reweighting pseudo-data, so the analysis-ready claim is established only when the simulation prior matches data.","rationale":"The reader's weakest assumption correctly points to the transfer of simulation-learned corrections to data, including acceptance. I narrow the focus to one specific, load-bearing mechanism: Eq. (2.5) is prior-dependent through p_MC,s(x)_g and p_MC,s(x)_\\bar g, and the acceptance classifier is never updated during iteration. This is mathematically concrete, it is explicitly conceded in App. A, and Sec. 4 deliberately removes the offending prior shift by reweighting Pythia to Herwig. Because the paper's headline claim is about accommodating the full suite of effects for real measurements, the fact that the only physics demonstration avoids the acceptance prior-mismatch is a genuine gap. The concern does not invalidate the method when priors are matched or when the sideband prescription is used; it limits the generality of the analysis-ready claim. The reader already issued a conditional verdict, and my analysis supports that verdict rather than moving it: the method is promising and transparent about its limitations, but the central claim is not yet demonstrated for realistic prior shifts. I mark agreement as partial because the reader lists several conditions (reliable backgrounds, universal detector response, code availability, uncertainties), whereas I identify acceptance prior-dependence as the single most decisive unresolved issue.","tokens_in":15111,"tokens_out":7010,"duration_ms":67268,"concrete_test":"Repeat the Sec. 4 benchmark with the original Herwig pseudo-data (no Pythia reweighting), train the acceptance classifier once on Pythia MC, and unfold with GenFoldC for eight iterations. Compare the final jet mass, jet width, and multiplicity distributions to Herwig truth within the fiducial region; if percent-level closure still holds despite the prior mismatch, the concern is resolved. Alternatively, in the Gaussian setup of App. A, take p_MC,s(y)=N(0,1) and p_d,s(y)=N(0.2,0.8) with a fiducial cut at y>1.2, and quantify the edge bias obtained with the fixed acceptance classifier versus an acceptance classifier retrained at each unfolding iteration.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The acceptance correction in Eq. (2.5) replaces the data-level quantity p_d,s(g|x) with the simulation-level ratio p(g)p_MC,s(x)_g / [p(g)p_MC,s(x)_g + p(\\bar g)p_MC,s(x)_\\bar g]. These densities depend on the gen-level prior p_MC,s(y), so the learned correction approximates p_MC,s(g|x), not p_d,s(g|x), whenever the data prior differs from the simulation prior. Neither algorithm updates this classifier: Alg. 5 uses a fixed acceptance classifier alongside iterated unfolding weights, and Alg. 6 draws Bernoulli masks from the same fixed classifier. The paper itself demonstrates the failure in App. A, where a fiducial cut on the unfolded observable produces edge mismodeling when the acceptance correction is applied naively. In Sec. 4, the issue is avoided by replacing Herwig pseudo-data with reweighted Pythia, forcing the gen-level distributions of simulation and data to agree before unfolding. The central claim that GenFoldC and GenFoldG 'are able to accommodate all effects' is therefore not established for the realistic case of a genuine prior shift: acceptance is handled correctly only when the simulation prior coincides with the data prior.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper extends iterative generative ML unfolding to include background subtraction, acceptance correction, efficiency correction, and prior-simulation dependence removal. It proposes two algorithms: GenFoldC, which combines a generative unfolding model with classifier-based iterative reweighting, and GenFoldG, which uses generative models throughout with an empty-event mechanism. The methods are demonstrated on a Gaussian toy example and on a six-dimensional Z+jets benchmark built from the OmniFold dataset, with percent-level agreement reported after iteration. The paper also includes appendices on prior-dependent acceptance effects, different forward mappings, and hyperparameters.","tokens_in":15265,"tokens_out":6511,"duration_ms":61156,"significance":"If the central claim holds, this is a timely and useful contribution to unbinned, high-dimensional unfolding: it addresses effects that most generative unfolding papers omit and provides two complementary iterative schemes. The strengths of the paper include explicit pseudo-code for every step, closure tests on a Gaussian example with a genuinely shifted prior, a 6D physics benchmark with background, and an honest Appendix A that demonstrates a failure mode of the acceptance correction. The main gap is that the abstract and conclusion claim the methods 'are able to accommodate all effects,' while the validation deliberately avoids the realistic case where acceptance or detector response depends on a prior that differs between simulation and data.","major_comments":[{"comment":"The acceptance correction in Eq. (2.5) replaces the data-level quantity p_d,s(g|x) with the simulation-level ratio p(g)p_MC,s(x)_g / [p(g)p_MC,s(x)_g + p(\\bar g)p_MC,s(x)_\\bar g], so it approximates p_MC,s(g|x), not p_d,s(g|x), whenever the gen-level prior differs between data and simulation. Neither Algorithm 5 nor Algorithm 6 retrains the acceptance classifier during the iterative prior-removal procedure. The paper itself demonstrates the resulting failure in Appendix A, where a cut on the unfolded observable causes edge mismodeling. In the physics benchmark, the issue is avoided by replacing the Herwig pseudo-data with reweighted Pythia, forcing the gen-level distributions of simulation and data to agree. The abstract and conclusion therefore overstate the claim that GenFoldC and GenFoldG 'are able to accommodate all effects'; this is established only when the simulation prior coincides with the data prior. I recommend either retraining \\delta(x) with the iterated unfolded weights, adopting the sideband solution from Appendix A as part of the main algorithms, or explicitly scoping the claim.","section":"Sec. 2.2, App. A, Sec. 4"},{"comment":"As printed, Algorithm 1 assigns label 1 to both {x_d} and {x_MC,b} and label 0 to {x_d}, so the two training classes overlap and the stated likelihood-ratio relation \\nu(x) = C(x)/(1-C(x)) does not follow. This makes the central background-subtraction step ill-defined. Please correct the algorithm statement (presumably labels 1 for x_d and 0 for x_MC,b, or the reverse) and reconcile Eq. (2.1) with the actual class definitions and balanced-training procedure.","section":"Algorithm 1 and Eq. (2.1)"},{"comment":"The main physics benchmark removes differences in the forward mapping by replacing the Herwig pseudo-data with reweighted Pythia 'to more easily facilitate comparisons between methods' (p. 11). As the authors note, even with the same detector simulation the detector response is not universal when higher-level observables are considered. The benchmark therefore does not exercise the pipeline under a genuinely mismatched forward model or a prior shift, which is a load-bearing limitation for the 'analysis-ready' claim. I suggest adding a stress test with an un-reweighted Herwig pseudo-data set, or explicitly listing non-universal detector response as an unvalidated systematic.","section":"Sec. 4 and App. B"},{"comment":"No statistical uncertainties are shown on any unfolded distribution, so the 'percent-level agreement' is not quantified. Because the iterative procedures involve stochastic training and finite Monte Carlo samples, bootstrap or seed-level bands are needed to support the central claim and to determine whether the residual differences at extreme values of \\tau_{21} and \\log\\rho are significant.","section":"Figures 4, 5, 9, 11"}],"minor_comments":[{"comment":"There is a typo in the required data line: 'background samples samples x_{MC,b}' should read 'background samples x_{MC,b}'.","section":"Algorithm 2"},{"comment":"The right panel legend is hard to parse because 'p_{d,s}(x)' and 'p_{d,s}(x)_g' appear similar; please make the notation for the gen-level-selected sample explicit.","section":"Fig. 3"},{"comment":"Appendix A demonstrates the sideband fix only for GenFoldC; stating whether GenFoldG with empty events behaves the same under the fiducial cut would strengthen the discussion.","section":"App. A"},{"comment":"The sentence 'After training we can compute p_{d,unfold}(y)_r by ... p_{d,unfold}(y)_r = \\int dx p_{MC,s}(y|x_{d,s}) p_{d,s}(x)_g' should specify that p_{d,s}(x)_g is the background-subtracted, acceptance-corrected data density, since that is how it is used later.","section":"Sec. 2.4"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid incremental contribution with honest self-assessment in the appendices. The main issue is the mismatch between the abstract's strong claim and the evidence: the acceptance correction is not iterated and the physics benchmark removes prior differences by construction. This is fixable within the manuscript's scope by revising the claims and adding a stress test. The Algorithm 1 inconsistency must be corrected before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuinely useful methods paper. It assembles background subtraction, acceptance, and efficiency corrections into iterative generative unfolding, and the closure tests on the Gaussian toy and the 6-dim Z+jets benchmark are clean at the percent level. The generator-based iteration with empty events (Alg. 6) is a real new piece. The authors also deserve credit for being upfront about failure modes: App. A shows the acceptance classifier goes wrong on a fiducial edge, and App. B explains why the Herwig pseudo-data are replaced by reweighted Pythia.\n\nThe soft spot is in the central claim. The acceptance correction in Eq. (2.5) is learned from simulation and estimates p_MC,s(g|x). That is the right quantity only when the gen-level prior in the simulation matches the data. The iterative loop updates the unfolding prior but never retrains the acceptance classifier, so a genuine prior shift that changes acceptance is not corrected. The paper demonstrates the problem in App. A, then proposes a sideband workaround, but the main algorithms are never validated in that regime. In Sec. 4, the prior shift is deliberately removed by reweighting Pythia to Herwig, so the physics benchmark tests the pipeline under the very condition where the acceptance correction is valid. That makes 'analysis-ready' an overstatement; it is analysis-ready when the simulation prior already matches data.\n\nWeaker but relevant: there are no uncertainty bands anywhere, and no code release, so the percent-level claims have no error bars and are not directly reproducible. The math is otherwise sound, and the citations to the component methods are fair.\n\nBottom line: this is worth a serious referee. The method is useful and the exposition is honest. For publication, I'd want to see a test with a true prior shift and a non-trivial acceptance region, and some estimate of statistical/systematic uncertainty. If those are added, the paper becomes a solid contribution to the unbinned unfolding toolkit. I'd cite it and probably bring it to the reading group.","headline":"Useful integration of background, acceptance, and efficiency into generative unfolding, but the analysis-ready claim is undercut by a prior-dependent acceptance correction that the benchmark sidesteps.","tokens_in":15872,"tokens_out":3718,"would_cite":true,"duration_ms":34300,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that generative unfolding can be made analysis-ready by adding background, acceptance, and efficiency corrections, reaching percent-level agreement with truth on unbinned, high-dimensional examples.","keywords":["unfolding","generative machine learning","unbinned differential cross sections","background subtraction","acceptance correction","efficiency correction","iterative unfolding","flow matching"],"falsifier":"A concrete falsifier: generate pseudo-data with a detector response or shower model different from the one used to train the pipeline, without reweighting the generator-level spectra, and check whether the unfolded distribution still matches truth at the percent level; the paper itself shows the failure mode in Appendix A, where a fiducial cut on the unfolded observable makes the acceptance classifier mismodel the edge.","tokens_in":14845,"feed_emoji":"🎯","tokens_out":9981,"duration_ms":86690,"temperature":0.7,"pith_summary":"Unfolding is the statistical removal of detector effects from particle-collision data, and machine-learning versions do it unbinned and in many dimensions. Until now, most of those methods corrected detector resolution only, leaving background contamination, events that appear at detector level but fall outside the measured fiducial region (acceptance/fakes), and events that belong in the region but are not reconstructed (efficiency/misses) untreated. This paper claims to close that gap with two iterative pipelines, GenFoldC and GenFoldG, that fold all of these effects into one generative unfolding procedure. In a Gaussian toy model and in a six-dimensional simulated Z+jets sample with about 8% background, both methods recover the true particle-level distribution to percent-level agreement after five to ten iterations. If correct, this makes generative unfolding usable for realistic, unbinned differential cross-section measurements rather than a resolution-only proof of concept.","feed_headline":"Unfolding now covers every detector effect, not just resolution","feed_subtitle":"Two generative pipelines take simulated Z+jets data to percent-level truth across six observables.","key_machinery":"The load-bearing mechanism is the iterative reweighting identity that turns a classifier into a density-ratio update. After one unfolding pass, a classifier separating the current unfolded distribution from the simulation prior yields weights $w(y)=\\nu(y)\\,\\delta(y)\\,p_{d,s}(y)/p_{MC,s}(y)_r=E(y)/(1-E(y))$, which define the joint training samples for the next iteration. GenFoldC finishes with the efficiency-corrected product above; GenFoldG instead replaces weighted updates with generative draws, injecting 'empty events' to mimic events lost to acceptance or efficiency while keeping the detector response $p(x|y)$ fixed and updating only the particle-level density. Acceptance and efficiency corrections are themselves classifier-estimated ratios: $\\delta(x)$ for the fraction of reconstructed events inside the fiducial region, and $\\epsilon(y)$ for the fraction of fiducial events that get reconstructed. The generative backbone is conditional flow matching, and the same correction steps can be implemented with classifiers or generators independently.","core_discovery":"On its own terms, the paper's discovery is that the standard generative-unfolding loop—learn the conditional density $p_{MC,s}(y|x)$ from paired simulation, apply it to data, then iterate to remove the simulation prior—can be extended into a complete measurement pipeline without leaving the unbinned setting. The five steps are: subtract background by reweighting or generating the signal density; correct acceptance with a classifier that estimates the conditional probability that a reconstructed event lies in the fiducial region; unfold the resolution effects; iterate to remove prior dependence, either by classifier reweighting (GenFoldC) or by regenerating weighted pairs (GenFoldG); and correct efficiency with a reciprocal classifier weight or with empty events. The final GenFoldC distribution is $p_{\\mathrm{GenFoldC}}(y)=\\epsilon(y)\\,\\delta(y)\\,p^n_{d,\\mathrm{unfold}}(y)_r$, while GenFoldG produces unweighted events approximating $p_{d,s}(y)$. In the Gaussian example and the six-observable Z+jets benchmark with 8% background and small acceptance and efficiency losses, both variants match the truth at the percent level, with residual distortions concentrated at extreme values of the N-subjettiness ratio and groomed mass and attributed to residual prior dependence.","pith_inferences":["Beyond the paper's demonstrations, the sideband prescription suggests a general rule for unbinned analyses: apply a learned acceptance correction only to selections that are genuinely detector-mediated, and enforce any selection that commutes with the detector response as a final cut on unfolded events.","The empty-event construction used by GenFoldG may transfer to full-event, variable-length unfolding, where lost particles could be represented as empty slots; this would let acceptance and efficiency corrections reach event-level observables without binning.","Because the acceptance classifier output is a smooth function of the reconstructed features, one could extend the method to report fiducial cross sections for a continuous family of selection thresholds from a single unfolding run, avoiding retraining for each cut.","The observed instability of the background classifier in high-dimensional, clustered phase space suggests that for dense backgrounds, generator-based subtraction or positive-weight refinement may be more robust; a head-to-head stability comparison on the same sample would test that."],"forward_implications":["Differential cross sections can be reported unbinned and in many dimensions with background, acceptance, and efficiency handled inside the unfolding itself, so observables and fiducial selections can be chosen after the measurement.","Iteration reduces the dependence on the starting simulation: both variants converge toward the data distribution rather than the simulation prior, with closure confirmed when unfolding the reference simulation itself.","The choice of classifier-based or generator-based corrections can be made independently for each pipeline step, so the same framework accommodates different generative models and classifier backends.","Residual differences remain in extreme phase-space regions, such as the high-mass tail and extreme values of the N-subjettiness ratio and groomed mass; in practice these become a systematic uncertainty on the method.","When a fiducial cut acts directly on the unfolded observable, the learned acceptance correction mismodels the boundary, so the paper's recommended recipe is to unfold with sidebands and apply the gen-level selection as the last step."],"supporting_citations":[{"why":"Supplies the simulated Z+jets benchmark and the particle-level/reco-level event pairs used for training.","marker":"[5]"},{"why":"Provides the classifier-based background-subtraction recipe that the paper adapts as its first pipeline step.","marker":"[6]"},{"why":"Introduces the iterative conditional-generative unfolding loop that GenFoldC and GenFoldG extend to background, acceptance, and efficiency effects.","marker":"[24]"},{"why":"Supplies the likelihood-ratio reweighting relation used to convert classifier output into background-subtraction weights.","marker":"[38]"},{"why":"Introduces the classifier-based efficiency correction that GenFoldC applies after the iterative unfolding.","marker":"[42, 43]"},{"why":"Provides the conditional flow-matching model used as the generative backbone for unfolding and generator-based corrections.","marker":"[44]"},{"why":"Defines the acceptance and efficiency selection (leading-jet transverse momentum above 150 GeV) used in the simulated physics benchmark.","marker":"[65]"}],"fun_headline_variants":["Generative unfolding now handles all detector effects","Unfolding goes beyond resolution to full pipeline","GenFoldG and GenFoldC: complete unfolding toolkit","From resolution to full detector correction in one pipeline","Analysis-ready unfolding: no more ignoring acceptance or efficiency"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the simulation used for training describes the detector for the real data: the estimated background is correct, the detector treats data and simulation events the same way, and the acceptance and efficiency rates learned from simulation apply to data unchanged.","fun_headline_variants_meta":{"raw":{"variants":["Generative unfolding now handles all detector effects","Unfolding goes beyond resolution to full pipeline","GenFoldG and GenFoldC: complete unfolding toolkit","From resolution to full detector correction in one pipeline","Analysis-ready unfolding: no more ignoring acceptance or efficiency"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000183,"raw_usage":{"total_tokens":1303,"prompt_tokens":921,"completion_tokens":382,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":537,"completion_tokens_details":{"reasoning_tokens":309}},"tokens_in":537,"tokens_out":382,"duration_ms":3920,"temperature":1.0,"reasoning_tokens":309,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:35:33.802773+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete falsifier: generate pseudo-data with a detector response or shower model different from the one used to train the pipeline, without reweighting the generator-level spectra, and check whether the unfolded distribution still matches truth at the percent level; the paper itself shows the failure mode in Appendix A, where a fiducial cut on the unfolded observable makes the acceptance classifier mismodel the edge.","supporting_citations":[],"review_version":2}