{"id":"7fa8363f-4d9b-455c-99d2-80ddf2600c7b","arxiv_id":"2608.10961","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A conditional normalizing flow reconstructs tau polarimetric vectors from simulated LHC events, improving spin-observable resolution by about 40% over a regression baseline and projecting >4.3 sigma entanglement separation at the HL-LHC.","lead":"TauPolaris uses a type of AI called a conditional normalizing flow to estimate the invisible neutrinos in tau lepton decays, recovering the tau spin direction more accurately than earlier methods. If it performs in real LHC data as it does in simulation, it could sharpen CP violation, entanglement, and background suppression measurements.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 4.3 sigma entanglement projection rests on an unvalidated S/B-preservation assumption in the highest-BDT-score window; a moderate S/B degradation could move the headline significance below the claimed value.","rationale":"The reader's weakest_assumption identifies exactly the same issue I consider most load-bearing: the 4.3 sigma entanglement projection is obtained by scaling the highest-BDT-score window yields from 62.4/fb to 3/ab while assuming the signal-to-background ratio is preserved. This is an explicit, testable assumption, and the paper is transparent about it, but the abstract headlines the resulting significance as if it were the expected HL-LHC sensitivity. My proposed check would settle whether the concern actually lands by quantifying how much the p-value degrades under plausible S/B reductions. The paper is otherwise internally consistent and well-executed, with public code and detailed simulation; I do not see an internal inconsistency that would justify rejecting the work. The secondary null-hypothesis caveat is worth noting, but the chosen separable benchmark is natural for H to tau tau and does not by itself change the verdict.","tokens_in":20908,"tokens_out":21841,"duration_ms":236662,"concrete_test":"Recompute the Section V pseudoexperiments with the highest-window signal and background yields scaled to 3/ab under alternative S/B assumptions (e.g., S/B multiplied by 0.7, 0.5, and 0.3), keeping all other fit choices fixed. If the one-sided p-value for C remains below 1.35e-3 (3 sigma) for S/B reduced by a factor of two, the 4.3 sigma projection is robust to the flagged assumption; if it rises above that threshold, the headline should be reported as conditional on the S/B-preservation assumption.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section V derives the headline HL-LHC entanglement significance by taking the yields of the highest-BDT-score window of the CMS CP analysis [4] and scaling them from 62.4/fb to 3/ab under the explicit assumption that the signal-to-background ratio in that window is preserved. This assumption is load-bearing because the quoted p-values are produced by pseudoexperiments whose null distribution is set by the background yield and shape; at HL-LHC pileup, tau reconstruction efficiency, trigger thresholds, and the BDT score distribution will change, so the S/B in a fixed high-score window is unlikely to be exactly preserved. If the S/B is lower than assumed, the no-entanglement pseudoexperiment distribution broadens and the separation (currently p<1e-5 for C) shrinks. The 18% CP improvement is less exposed because it is a normalized signal-only asymmetry, but the abstract's 'at least 4.3 sigma' claim is directly conditional on this extrapolation. A secondary caveat is that the no-entanglement null is one specific separable state (Eq. 15); a fully robust 'absence of entanglement' claim would profile over the separable-state boundary, though this null is a reasonable benchmark.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces TauPolaris, a conditional normalizing flow for reconstructing the undetected neutrino momenta in tau decays from reconstructed event-level features, and uses the resulting momenta to build tau polarimetric vectors. The flow is trained on Pythia+Delphes samples of H->tautau and Z/gamma*->tautau events; a transformer conditioning network summarizes the inputs, and a MAP estimate is obtained by gradient ascent in the latent space. The authors compare the flow with a stand-alone transformer regressor on independent test samples, reporting improved resolution of spin observables, and demonstrate three applications: a HL-LHC projection for distinguishing entangled from non-entangled H->tautau spin states, an 18% improvement in sensitivity to CP violation in H->tautau, and new spin-correlation variables for suppressing Z->tautau background.","tokens_in":21210,"tokens_out":6809,"duration_ms":74261,"significance":"The methodological core is sound and well presented: the full conditional density approach is a natural improvement over point regression for the intrinsically ambiguous neutrino reconstruction problem, and the paper provides concrete evidence for this through independent test samples, physically sensible reconstructed tau masses, and a direct comparison with a same-input transformer baseline. The public code release is a strength, and the 300k pseudoexperiments give a solid statistical basis for the entanglement projection. The headline entanglement significance, however, is conditional on an extrapolation assumption that is explicitly stated but not stress-tested, and the no-entanglement null hypothesis is one specific separable benchmark rather than a general test for the absence of entanglement. The CP and background-suppression sections are plausible demonstrations, but the 18% improvement statistic would benefit from a more explicit derivation.","major_comments":[{"comment":"The quoted 'at least 4.3 sigma' entanglement separation is obtained by scaling the yields of the highest-BDT-score window of Ref. [4] from 62.4 fb^-1 to 3 ab^-1 under the stated assumption that the signal-to-background ratio in that window is preserved. This assumption is load-bearing because the no-entanglement pseudoexperiment distribution is set by the background yield and shape, and HL-LHC pileup, trigger thresholds, tau identification, and the BDT score distribution will not remain exactly as at 62.4 fb^-1. Please add a robustness study in which the S/B ratio is degraded by, say, 20% and 50%, or rephrase the abstract's claim as conditional on this preservation assumption.","section":"Section V (yield extrapolation)"},{"comment":"The pseudoexperiment p-value is computed against a single separable state with C_kk=-1 and all other C_ij=0. This is a reasonable and physically motivated benchmark, but it does not by itself support the abstract's statement that 'the presence of quantum entanglement can be distinguished from its absence.' A different separable state could produce different reconstructed C and m_12 distributions. Please either profile over a family of separable states, or soften the claim to 'distinguished from the classically-correlated benchmark considered here.'","section":"Section V (Eq. 15 and Fig. 8)"},{"comment":"The 18% improvement in CP sensitivity is based on the asymmetry A defined in Eq. (16) and the summary statistic A_total in Eq. (17). Since A is computed from normalized CP-even and CP-odd histograms, its proportionality to the sensitivity of a binned likelihood fit is not automatic; the text asserts this but does not derive it or specify precisely what N_c represents in Eq. (17). Please provide a derivation or a direct likelihood-based validation of the weighting, and define N_c explicitly, so that the 18% claim can be checked.","section":"Section VI (Eqs. 16 and 17)"}],"minor_comments":[{"comment":"The axis label in the upper-left panel of Fig. 4 appears to read '1 Energy (GeV)'; please check that it correctly identifies the neutrino energy variable.","section":"Section IV.D / Fig. 4"},{"comment":"The hyperparameters selected by Optuna are described only in the text; for reproducibility, please include the final hyperparameter values or a configuration file alongside the code release.","section":"Section IV.B"},{"comment":"The abstract's significance statement does not mention that the quoted 4.3 sigma is a statistical-only projection with no nuisance parameters. Please add a qualifier such as 'statistical-only' or 'under the simplifying assumptions described in Section V'.","section":"Section V"},{"comment":"The background-suppression variables are demonstrated through normalized distributions only; a quantitative measure such as signal efficiency versus background rejection would make the claimed utility concrete.","section":"Section VII"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a solid methods paper that does what it claims on simulated data. The flow-based reconstruction resolves spin observables about 40% better (IQR) than an MSE-trained transformer, the full-density modeling fixes the known bias of per-component regression toward the visible tau direction, and the authors ship code and validate on independent samples. The physics projections are plausible, but the headline 4.3 sigma entanglement claim is softer than it looks because it inherits an untested assumption: the signal-to-background ratio in the highest-BDT-score bin of the CMS CP analysis survives the jump from 62.4/fb to 3/ab at much higher pileup. The authors state this assumption in Section V, so they are not hiding it, but it is load-bearing. The no-entanglement null distribution in their pseudoexperiments is set by background shape and yield, and a modest S/B degradation would move the separation down. The CP improvement (18% in A_total) is a normalized signal-only asymmetry, so it is much less exposed; the Z-discrimination variables are new and look useful.\n\nWhat is genuinely new: applying normalizing flows to tau polarimetric-vector reconstruction with tau-specific touches (lifetime-sensitive inputs, visible-tau-aligned coordinates, latent-space MAP, transformer conditioning). The comparison against a transformer baseline is the right control and it shows a real gain. I do not see a circularity problem: the flow is trained on a reweighted Higgs sample, but the density-matrix elements are not fit to the final observables, and the test samples are independent.\n\nWeaknesses, in proportion: the HL-LHC projection is a feasibility study, not a full sensitivity estimate. It uses fixed yields and shapes, no nuisance parameters, one specific separable-state null (Eq. 15), and the S/B-preservation extrapolation. For a paper whose abstract leads with 'at least 4.3 sigma', that is a meaningful gap. I would want the authors to stress-test it with a degraded S/B assumption and perhaps profile over the separable boundary. That said, the caveat is stated, and the physics conclusion may well survive; it just is not established at the level the abstract implies.\n\nThis paper deserves a serious referee. It is clearly argued, technically sound on the reconstruction side, and the tool is public. I would accept it with requests to either soften the headline or add a sensitivity scan. Worth bringing to the reading group if anyone works on tau spin/CP or ML-based neutrino reconstruction.","headline":"Solid ML reconstruction paper with a real resolution gain, but the headline HL-LHC entanglement significance leans on an explicitly stated yet untested yield-extrapolation assumption.","tokens_in":21651,"tokens_out":2537,"would_cite":true,"duration_ms":67620,"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":"A conditional normalizing flow reconstructs tau polarimetric vectors from missing neutrino momenta, improving spin observables and HL-LHC sensitivities.","keywords":["tau lepton","polarimetric vector","normalizing flows","neutrino reconstruction","spin correlations","quantum entanglement","CP violation","HL-LHC"],"falsifier":"A reader could settle the projection by recomputing the pseudoexperiment significance with the signal-to-background ratio actually measured in that bin under HL-LHC pileup conditions; if the ratio falls below the scaled value, the at-least-4.3-sigma separation would not be reproduced.","tokens_in":20698,"feed_emoji":"⚛️","tokens_out":7481,"duration_ms":69795,"temperature":0.7,"pith_summary":"This paper claims that the optimal tau spin observable, the polarimetric vector, can be reconstructed accurately by learning the full conditional density of the undetected neutrino momenta rather than by point regression. The method, TauPolaris, uses a conditional normalizing flow conditioned on reconstructed event quantities, and the paper argues that this density-based approach preserves correlations between neutrino momentum components that a mean-squared-error regressor washes out. In simulated LHC collisions with detector effects, the reconstructed spin observables achieve better resolution, and the improved observables translate into three physics results: distinguishing quantum entanglement from its absence at at least 4.3 sigma at the HL-LHC, an 18% improvement in sensitivity to CP violation in H to tau tau decays, and new variables for suppressing Z to tau tau background. If these claims hold, the paper provides a recipe for extracting maximal spin information from tau final states at hadron colliders.","feed_headline":"Neural flow recovers tau spin with 4.3-sigma entanglement reach","feed_subtitle":"TauPolaris models the full density of missing neutrinos, sharpening spin resolution and CP sensitivity at the HL-LHC.","key_machinery":"The central object is the conditional normalizing flow: an invertible neural transformation f mapping neutrino kinematics nu to a Gaussian latent variable z, conditioned on a context vector c learned by a transformer encoder. The density is given by the change-of-variables formula log p(nu|c) = log N(z; 0, I) + log |det J_f(nu|c)|, which makes the full joint density tractable and lets the model capture correlations and multimodality in the neutrino posterior. The most probable neutrino configuration (MAP) is found by gradient ascent in the latent space, exploiting the isotropic Gaussian base to stabilize optimization; sampling from the latent distribution reproduces the generator-level kinematic distributions. The neutrino momenta are parametrized in a per-tau orthonormal basis aligned with the visible tau momentum, which closely matches the basis in which the spin correlations are expressed, and displacement variables sensitive to the tau lifetime are included as conditioning inputs.","core_discovery":"The paper's central claim is that the polarimetric vectors of tau leptons can be reconstructed well enough to make spin-correlation measurements at the LHC significantly more powerful, by using a conditional normalizing flow to estimate the momenta of the invisible neutrinos. The flow models the joint probability density of the neutrino kinematics given the reconstructed event, so sampling from it reproduces the generator-level distributions and its mode provides an accurate per-event estimate. Compared with a transformer regressor trained on a mean-squared-error loss, the flow's MAP estimates improve the interquartile-range resolution of the reconstructed cos theta spin variables by about 40%, and its sampled density resolves the multimodality that biases point regression toward the visible tau direction. From the fitted spin density matrix, the paper projects that HL-LHC data can separate the entangled standard-model H to tau tau state from a separable classical-correlation state with significance at least 4.3 sigma, that the sensitivity to the CP mixing angle improves by 18% overall relative to previous approximate methods, and that polarimetric-vector variables can separate Higgs from Z events independently of the Higgs CP state.","pith_inferences":["Inference: Because the flow provides a per-event uncertainty on phi_CP, an optimized event selection or likelihood reweighting using that uncertainty could push the CP sensitivity beyond the 18% reported, which uses only one fixed split.","Inference: The same conditional-flow density approach should transfer to other missing-energy final states with sparse visible objects, such as W to tau nu or heavy-resonance decays, wherever a simulator can provide training pairs.","Inference: A direct test of the method's probabilistic calibration, such as coverage of the true neutrino momentum by the predicted posterior, would clarify whether the per-event uncertainties are trustworthy enough for future precision analyses."],"forward_implications":["The MAP estimates from TauPolaris improve the resolution (IQR) of reconstructed spin observables by about 40% relative to an MSE-trained transformer and about 20% relative to sampling, so spin-correlation fits become more precise for the same event sample.","At the HL-LHC, the fitted spin density matrix can separate the standard-model entangled H to tau tau state from a separable classical-correlation state with at least 4.3 sigma significance using concurrence and 4.0 sigma using the Bell-variable m12.","The sensitivity to the CP mixing angle in H to tau tau improves by 18% overall compared with previous approximate methods, with up to 88% improvement in the 3pi 0pi0-3pi 0pi0 channel and first-time reconstruction of the phi_CP angle in 3pi 1pi0 decays.","Variables constructed from the reconstructed polarimetric vectors (cos theta+k cos theta-k, cos theta+n cos theta-n - cos theta+r cos theta-r, and the tau polarizations) separate Higgs from Z to tau tau events and are invariant under the Higgs CP mixing angle, providing new background-suppression handles."],"supporting_citations":[{"why":"Supplies the analysis categories, classification score windows, and yields at 62.4 fb^-1 that are scaled to project HL-LHC sensitivities.","marker":"[4]"},{"why":"Earlier diffusion-model approach for neutrino reconstruction in tau pairs, providing the density-modeling comparison point.","marker":"[8]"},{"why":"Established conditional normalizing flows for neutrino regression in top-quark decays, which this work adapts to tau decays.","marker":"[10, 11]"},{"why":"Provides the spin-reweighting formalism used to generate samples with arbitrary spin density matrix elements.","marker":"[15]"},{"why":"Provides the polarimetric-vector formalism used for reweighting and for the 3pi 1pi0 decay mode.","marker":"[16–19]"},{"why":"Defines concurrence, the entanglement measure used in the significance projection.","marker":"[12]"},{"why":"Supplies the rational-quadratic spline coupling layers that make the flow invertible with tractable Jacobians.","marker":"[28]"},{"why":"Provides the transformer encoder architecture used as the conditioning network.","marker":"[31]"},{"why":"Hadron-plus-strips reconstruction of hadronic tau decays that defines the input objects for the flow.","marker":"[22]"},{"why":"HEPData record for the experimental CP analysis yields used in the HL-LHC extrapolation.","marker":"[39]"}],"fun_headline_variants":["Flow sharpens tau spin: 4.3σ entanglement at HL-LHC","Normalizing flow maps neutrinos to reveal tau spin correlations","Flow recovers tau spin: 4.3σ entanglement, 18% CP gain","Flow turns missing neutrinos into sharp tau spin observables","TauPolaris flow: better tau spin means 4.3σ for entanglement"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the signal-to-background ratio in the most signal-rich bin of the experimental CP analysis remains unchanged when the dataset is scaled from 62.4 $fb^{-1}$ to 3 $ab^{-1}$ at the HL-LHC, since the quoted entanglement significance is projected under that scaling.","fun_headline_variants_meta":{"raw":{"variants":["Flow sharpens tau spin: 4.3σ entanglement at HL-LHC","Normalizing flow maps neutrinos to reveal tau spin correlations","Flow recovers tau spin: 4.3σ entanglement, 18% CP gain","Flow turns missing neutrinos into sharp tau spin observables","TauPolaris flow: better tau spin means 4.3σ for entanglement"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001119,"raw_usage":{"total_tokens":4702,"prompt_tokens":1032,"completion_tokens":3670,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":648,"completion_tokens_details":{"reasoning_tokens":3570}},"tokens_in":648,"tokens_out":3670,"duration_ms":24561,"temperature":1.0,"reasoning_tokens":3570,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:20:32.102622+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could settle the projection by recomputing the pseudoexperiment significance with the signal-to-background ratio actually measured in that bin under HL-LHC pileup conditions; if the ratio falls below the scaled value, the at-least-4.3-sigma separation would not be reproduced.","supporting_citations":[{"cited_title":"Analysis of the $C\\!P$ structure of the Yukawa coupling between the Higgs boson and tau leptons in proton-proton collisions at $\\sqrt{s}$ = 13.6 TeV","cited_arxiv_id":"2606.03510","evidence_quote":"HEPData record for the experimental CP analysis yields used in the HL-LHC extrapolation."}],"review_version":1}