{"id":"2833ae46-7a4d-41a3-8948-a8a51feb788b","arxiv_id":"2509.06370","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A normalizing-flow-based Bayesian unfolding method (NPU) plus a modern Python implementation of Fully Bayesian Unfolding (FBU) are introduced and validated on Gaussian and simulated LHC jet data.","lead":"This paper introduces Neural Posterior Unfolding (NPU), a normalizing-flow-based Bayesian method that learns the mapping from detector-level to particle-level histograms, plus a modern Python version of the classical FBU method, with tests on Gaussian and simulated LHC jet data. Why read it: unfolding is the standard correction step for nearly every cross-section measurement, so a fast, uncertainty-aware alternative to existing unfolding tools could simplify many LHC analyses","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"NPU's learned posterior is only validated under the training prior; the training prior itself is unspecified, and the physics test is a mismatched-prior scenario with no coverage check.","rationale":"The reader's weakest-assumption diagnosis is correct and, if anything, understated. The paper never specifies the training prior over truth histograms in Sec. III, and the only mention in Sec. IV is vague. The Gaussian validation is entirely on-prior: both training and test truths are Gaussian with the same parameters, so the pull tests in Fig. 4 cannot detect prior mismatch. The physics example in Sec. V.C is the only mismatched-prior test, but it is presented without quantitative coverage or bias metrics, making it insufficient to support the central claim of general validity. This is load-bearing because the method's headline advantage—correct uncertainty in degenerate regions—is exactly where the likelihood is weak and the prior dominates. If the training prior does not match the data-generating prior, the 'full posterior' will be the posterior under the wrong prior, potentially returning confident but incorrect uncertainties. The proposed shifted-Gaussian test is a concrete, computationally feasible check that would settle the concern. No other issue appeared more critical: the implementation seems sound, the pull tests are appropriate in the limited setting, and the code is available. The reader's CONDITIONAL verdict is appropriate and should stand pending this test.","tokens_in":11523,"tokens_out":5675,"duration_ms":66950,"concrete_test":"Use the Gaussian setup of Sec. V.B but train NPU on pseudo-experiments with truth drawn from a nominal prior (e.g., mu=0, sigma=1), then evaluate on pseudo-experiments whose truth is drawn from a shifted distribution (e.g., mu=0.5, sigma=1.2). Compute pull distributions as in Eq. 7 and 68%/95% credible-interval coverage per bin. If pulls deviate from N(0,1) or coverage differs from nominal, the learned posterior is prior-dependent and the central claim must be qualified. This directly tests the transfer scenario that is missing from the paper.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that NPU returns a faithful posterior over unfolded cross sections requires that the learned p(t|m) be the correct posterior for the observed data. This in turn requires either (a) an explicitly specified training prior over truth histograms that matches the actual prior, or (b) evidence that the likelihood dominates so prior mismatch is irrelevant. Neither is provided. Section III describes the flow architecture but never states how the (truth, measured) training pairs are generated or what prior over t is used. The only mention, in Sec. IV, that a uniform prior is 'chosen based on the distribution of parameter points in the training sample' is ambiguous and does not specify the sampling distribution. In the Gaussian validation (Sec. V.B), the test truth is drawn from the same distribution as the training samples (mu=0, sigma=1), so the pull tests in Fig. 4 only establish calibration under the training prior. In the physics example (Sec. V.C), the setup actually uses a mismatched prior: Pythia provides the response and presumably the NPU training pairs, while Herwig serves as the data and truth. The agreement shown in Fig. 5 is qualitative; no pulls or coverage tests are reported for this domain-shift scenario. Because the paper's advertised advantage over IBU is precisely in degenerate regions where the likelihood is weak and the prior dominates, an unspecified or mismatched training prior directly threatens the claimed uncertainty quantification. Thus the general validity of NPU is assumed rather than demonstrated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Neural Posterior Unfolding (NPU), a normalizing-flow-based method that learns the posterior p(t|m) over binned particle-level histograms t conditional on measured detector-level histograms m, using simulated (truth, measured) pairs. The unfolded result is obtained by maximizing the learned posterior over t. The paper also provides a modern Python implementation of Fully Bayesian Unfolding (FBU). The methods are first tested on a 2-bin degenerate-response example, where NPU returns a broad posterior while IBU incorrectly reports zero uncertainty on the bin ratio; then on an 8-bin Gaussian example with pull-based calibration checks; and finally on LHC-like jet substructure observables, using Herwig as pseudo-data and Pythia to build the response. The central claims are that NPU provides accurate unfolded central values, full posterior uncertainties that remain well-calibrated in weakly constrained regions, and fast amortized inference.","tokens_in":11928,"tokens_out":2623,"duration_ms":29681,"significance":"If the central claims hold, NPU would be a useful contribution to the unfolding literature: it offers amortized inference, direct posterior access, and a principled treatment of degenerate response regions where classical methods such as IBU are known to underestimate uncertainty. The paper also provides a modern, maintained FBU implementation and releases code, which are practical assets for the community. The conceptual demonstration on the 2-bin degenerate example is clear and valuable. However, the paper's load-bearing claims about posterior fidelity and calibration are not yet fully supported: the training prior for NPU is never specified, the Gaussian validation is same-prior, the acknowledged undercoverage at low smearing is not reconciled with the calibration conclusion, and the LHC example lacks quantitative agreement tests. These gaps are fixable, but they are central to the advertised advantages of the method.","major_comments":[{"comment":"The training prior for NPU is never specified. The method learns p(t|m) from (truth, measured) pairs, but the text never states how the truth histograms t are sampled in training. The only statement, in Sec. IV, that 'A uniform prior is also used for NPU (chosen based on the distribution of parameter points in the training sample)' is ambiguous and does not define a sampling distribution or range. Since the posterior in weakly constrained regions is dominated by the prior, an unspecified prior makes the uncertainty quantification claims unreproducible. Please specify the training pair generation procedure, including the prior over t, explicitly.","section":"Sec. III and Sec. IV"},{"comment":"The Gaussian validation tests NPU only under the training prior. The simulation and natural datasets both have mu=0, sigma=1, so the test truth is drawn from the same distribution as the training samples. The pull tests in Fig. 4 therefore establish calibration under the training prior, not the ability to unfold a genuinely different true distribution. Since real data are not drawn from the simulation prior, this is a load-bearing gap. Please add a validation where the test truth is drawn from a different distribution (e.g., shifted mu or sigma) and report pulls/coverage, or otherwise demonstrate that the likelihood dominates the prior.","section":"Sec. V.B, Table II"},{"comment":"The paper acknowledges that 'NPU appears to underestimate the variance at low smearing' (Fig. 4b), yet the conclusions state that both methods 'consistently produced pulls with mu=0 and sigma=1 across the tested smearing parameters.' These statements are contradictory. Undercoverage is a calibration failure that directly affects the central claim of reliable uncertainty quantification. Please quantify the pull standard deviations for each smearing value, report their statistical uncertainties, and either correct the conclusions or explain why this behavior is acceptable.","section":"Sec. V.B, Fig. 4b and Sec. VI"},{"comment":"The LHC example provides only a qualitative comparison between the unfolded result and the truth. The ratio-to-truth panels show agreement within bands, but no pulls, chi-square statistics, or coverage tests are reported for this mismatched-prior scenario. Given that this is the only external grounding of the method on a data distribution different from the training prior, a quantitative agreement test is necessary to support the claim that NPU 'accurately recovered their central values' and to validate the reported posterior widths.","section":"Sec. V.C, Fig. 5"}],"minor_comments":[{"comment":"The IBU update is written as t_j^{(n)} = sum_i Pr^{n-1}(truth_j|measure_i) Pr(measure_i). As written, this appears to be missing the normalization by total counts; the subsequent Eq. (2) is clearer. Please correct or clarify the notation.","section":"Sec. II, Eq. (1)"},{"comment":"The response matrix R is defined as R_ij = Pr(measure_i | truth_j), but the displayed matrix is labeled Pr(Truth | Reco). Please clarify whether the plotted matrix is the transpose of R or an element-wise conditional probability, and ensure the axes and colorbar labels are consistent.","section":"Fig. 2(b) caption"},{"comment":"Typo: 'including including jet width' appears in the conclusions. Also, the references to 'T able' and 'Fig. 4' formatting should be checked for journal style.","section":"Sec. VI"}],"recommendation":"major_revision","confidential_remarks":"The paper is promising and the conceptual advantage of NPU in degenerate regions is well illustrated. However, the missing training prior specification and the same-prior-only validation are central, not peripheral, issues for a method whose main selling point is posterior uncertainty quantification. The authors should be asked to specify the training prior, add an out-of-prior validation, and provide quantitative calibration checks for the LHC example. The code release should make these additions straightforward."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper introduces Neural Posterior Unfolding, a normalizing-flow neural posterior estimator for binned unfolding, and it's worth a serious look despite some unfinished validation. What's new is the application of NPE to histogram unfolding; the cited ML-unfolding works are mostly unbinned or use different frameworks. The two-bin degenerate example nails the motivation: IBU gives zero uncertainty where the posterior should be broad, and NPU returns something sensible. The modern PyMC rewrite of FBU is also a useful gift to the community.\n\nThe soft spots line up with the stress test. The training prior over truth histograms is never actually specified. Section IV's phrase 'uniform prior ... chosen based on the distribution of parameter points in the training sample' doesn't tell you how the truth samples were drawn. That's not a pedantic point: NPU's selling point is uncertainty in degenerate regions, which is exactly where the prior dominates. If the prior is arbitrary, the posterior is arbitrary there.\n\nThe Gaussian validation is same-prior: D_natural has the same mu and sigma as D_sim, so the pull tests only check the flow's consistency under the training distribution. It's good that the authors report NPU's under-coverage at low smearing (Fig 4b) rather than hiding it, but it undercuts the blanket claim of calibrated uncertainty.\n\nThe LHC example does give an external, mismatched-prior test: Pythia builds the response, Herwig supplies data and truth. That's a real domain shift and the results look plausible. But the agreement is visual only; no pulls or coverage for that scenario. So we learn the method doesn't blow up under shift, not that it's calibrated there.\n\nThis is a coherent paper with honest limitations and working code. It deserves peer review. A referee should ask for an explicit prior specification, a coverage test with a shifted truth distribution, and quantitative agreement metrics for the physics case. Those are addressable. I'd take it to a reading group as a good example of where flow-based unfolding stands, but I wouldn't cite it as an analysis-ready method yet.","headline":"NPU is a credible proof-of-concept for flow-based binned unfolding, but the training prior is unspecified and the validation is mostly same-prior, so the UQ claims outrun the evidence.","tokens_in":12329,"tokens_out":2948,"would_cite":true,"duration_ms":33371,"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":"Neural Posterior Unfolding claims that a normalizing flow trained on simulated truth-versus-measured histogram pairs yields a full, calibrated posterior over unfolded cross sections, with accurate maximum-likelihood readout and correct unce","keywords":["unfolding","deconvolution","normalizing flows","Bayesian inference","differential cross sections","uncertainty quantification","jet substructure","LHC"],"falsifier":"Take a trained NPU model and unfold pseudo-experiments whose true spectrum is deliberately far from the training prior, e.g. a Gaussian with a different mean and width under the same detector response; if the maximum-likelihood readout is visibly pulled toward the training distribution, or the posterior's credible intervals do not contain the true spectrum at the claimed rate, the accuracy claim fails outside the training regime.","tokens_in":11516,"feed_emoji":"⚛️","tokens_out":9426,"duration_ms":103239,"temperature":0.7,"pith_summary":"Neural Posterior Unfolding (NPU) is proposed as a way to correct measured histograms for detector distortions while keeping the uncertainty faithful. The paper's central claim is that a normalizing flow trained on simulated pairs of particle-level and detector-level histograms can learn the conditional density of the truth given the measured histogram, and that reading out this learned density with maximum likelihood recovers the true unfolded cross section. This matters because standard iterative Bayesian unfolding returns only a point estimate, and in degenerate bins—where two different truth spectra would produce identical detector data—it can report zero uncertainty when the true ambiguity is maximal. The paper shows, on a two-bin toy, an eight-bin Gaussian problem, and LHC-like jet substructure simulations, that NPU returns accurate central values and a full posterior, and that amortized training makes later uses much faster than MCMC-based Fully Bayesian Unfolding.","feed_headline":"Neural unfolding maps smeared data to full cross-section posteriors","feed_subtitle":"One trained network does the unfolding, and it keeps uncertainty wide exactly where detectors cannot tell bins apart.","key_machinery":"The central object is a conditional normalizing flow—an invertible neural network that maps a simple base distribution to a flexible probability density—trained with a negative log-likelihood loss on (truth, measured) histogram pairs to approximate p(t|m). The measured histogram serves as the conditional input, and the output is a density over truth histograms. After training, a maximum-likelihood estimation step optimizes the learned log-density with respect to the truth vector to extract the unfolded spectrum from new data. This flow does the work that MCMC does in Fully Bayesian Unfolding, but amortized: the expensive density learning happens once, and each new dataset only needs the fast","core_discovery":"The paper's central claim is that binned unfolding can be cast as an amortized neural posterior estimation problem. A normalizing flow is trained on pairs of particle-level and detector-level histograms to approximate p(t|m), the density of the true histogram t given the measured histogram m; after training, unfolding a new measurement is a maximum-likelihood optimization over t under the learned density. The claim is that this gives a full, properly calibrated posterior with correlations, that it reproduces the true spectrum in Gaussian and LHC-like tests, and that it handles degenerate response regions correctly by returning a broad posterior where iterative Bayesian unfolding would report","pith_inferences":["Inference: The two-bin degenerate argument implies that any regularized matrix-inversion method whose regularization acts on the truth vector—not just IBU—will understate uncertainty where response rows nearly coincide; NPU-style posterior estimation is a generic fix for that failure mode.","Inference: Since the paper does not specify the training prior over truth histograms in detail, the method's reliability on data whose true spectrum is far outside that prior is untested; an obvious stress test is to train on one prior and unfold pseudo-data drawn from a different one.","Inference: The flow's conditional density could be extended to include detector-systematic parameters such as smear widths and efficiencies by conditioning on them, turning the method into a joint statistical-plus-systematic unfolding.","Inference: Because the flow provides a density over histograms, the same machinery is naturally suited to unbinned unfolding if the conditional density is defined over individual event-level observables instead of bin counts; the paper names this as a future direction."],"forward_implications":["Every unfolded cross section comes with a full posterior over bins, including correlations, rather than a single point estimate, so downstream physics statements can propagate the complete statistical uncertainty.","In detector-degenerate regions, NPU reports wide credible intervals reflecting the true ambiguity, correcting the zero-uncertainty failure of iterative Bayesian unfolding.","The amortized training makes re-unfolding with bootstrapped or newly arriving data cheap after one training pass: the paper reports about five minutes of training versus roughly 67 minutes for 100 pseudo-experiments with MCMC-based unfolding.","NPU and FBU both pass pull-based calibration tests with means near zero and widths near one across tested smearing parameters, indicating the posterior is statistically well calibrated in the tested cases.","The method reproduces truth for LHC-like jet substructure observables in simulation, making it a plausible candidate for analysis-ready unfolding at the LHC and beyond."],"supporting_citations":[{"why":"Defines Fully Bayesian Unfolding, the MCMC baseline whose posterior NPU approximates with an amortized flow.","marker":"[8]"},{"why":"Introduces normalizing flows, the flexible density-estimation machinery NPU trains to model p(t|m).","marker":"[22]"},{"why":"Defines iterative Bayesian, Lucy-Richardson deconvolution, the frequentist baseline contrasted with NPU in degenerate bins.","marker":"[24–26]"},{"why":"Supplies the masked-autoregressive flow architecture used as the invertible network in NPU.","marker":"[31, 32]"},{"why":"Supplies the Bayesian sampling toolkit used to implement the FBU baseline.","marker":"[34]"},{"why":"Provides the Hamiltonian Monte Carlo sampler used for efficient posterior sampling in the FBU comparison.","marker":"[35]"},{"why":"Supplies the simulated LHC-like jet datasets used for the proof-of-concept validation of NPU and FBU.","marker":"[36, 37]"}],"fun_headline_variants":["Bayesian unfolding with neural flows cuts retraining and yields full uncertainty","Normalizing flows for unfolding: full posterior without the inversion","Amortized neural posterior for cross-section unfolding","Unfolding detector smears via learned posterior density"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that the simulated (truth, detector-level) pairs used to train the flow describe the same relationship as the real data, and that the distribution of truth histograms in training is close enough to the real spectrum that the likelihood dominates; otherwise the learned p(t|m) can be wrong and the unfolded result silently biased.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian unfolding with neural flows cuts retraining and yields full uncertainty","Normalizing flows for unfolding: full posterior without the inversion","Amortized neural posterior for cross-section unfolding","Unfolding detector smears via learned posterior density"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000548,"raw_usage":{"total_tokens":2445,"prompt_tokens":726,"completion_tokens":1719,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":470,"completion_tokens_details":{"reasoning_tokens":1653}},"tokens_in":470,"tokens_out":1719,"duration_ms":12154,"temperature":1.0,"reasoning_tokens":1653,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T23:42:27.275623+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a trained NPU model and unfold pseudo-experiments whose true spectrum is deliberately far from the training prior, e.g. a Gaussian with a different mean and width under the same detector response; if the maximum-likelihood readout is visibly pulled toward the training distribution, or the posterior's credible intervals do not contain the true spectrum at the claimed rate, the accuracy claim fails outside the training regime.","supporting_citations":[{"cited_title":"Salvatier, T","cited_arxiv_id":null,"evidence_quote":"Supplies the Bayesian sampling toolkit used to implement the FBU baseline."}],"review_version":1}