{"id":"58176876-82bc-40a7-b422-ba1854241c2d","arxiv_id":"2501.03981","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Machine learning on raw spectral data can separate the many-body critical phase from ergodic and localized phases in a quasiperiodic chain and reproduce known critical exponents from scaling collapse.","lead":"The authors use neural networks and principal component analysis on raw energy-level spacings and eigenvector probabilities of a quasiperiodic quantum chain, reporting that the tools can separate the ergodic, many-body critical, and many-body localized phases and even suggest finer substructure. If correct, the approach offers a relatively cheap way to map phase diagrams and extract critical exponents without designing bespoke order parameters.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The MBC and subphase claims rest on training labels from the phase diagram being tested plus an unquantified broadness heuristic; a label-free clustering check is needed to show the network is not just reproducing Ref [28].","rationale":"The reader's conditional verdict is the right calibration. The supervised three-class classification is a consistency check, not an independent discovery: the labels encode the very phase diagram the paper opens by questioning, so the high accuracy cannot settle the existence question. The binary-confusion signal is the only non-circular piece, and it is under-controlled. The subphase split at lambda=2 and mu=1 is arbitrary, and the broad-probability criterion is applied post hoc without error bars or a null distribution, so the MBC-I/II and MBL-I/II claims are the least secure part of the paper. I do not see the single-particle exponent extraction as a serious flaw: since the classifier is trained on IPR components and the same scaling cost function is applied, reproducing nu is a reasonable sanity check, but the abstract should say 'nu agrees' rather than 'identical critical exponents', since Table III shows w differs. The paper earns credit for honest limitations: it explicitly states that the many-body scaling collapse does not reliably extract exponents, and the transfer test in Appendix C, where an EAAH-trained network is applied to the ABF diamond chain, provides genuine out-of-model evidence that raw eigenvalue-spacing features carry phase information. The proposed k-means control would settle whether the MBC and substructure boundaries can be found without the Ref [28] labels; until that is shown, the central many-body claim remains conditional rather than established.","tokens_in":22962,"tokens_out":12636,"duration_ms":134213,"concrete_test":"Run a label-free clustering control on the same N=14 eigenvalue-spacing features used in Section IIIC: compute PCA on the unlabelled training data, keep components capturing 90% of the variance, apply k-means with k=3, and map each cluster to the closest Ref [28] phase label. If the three clusters correspond to ME, MBC and MBL and the cluster boundaries along the Fig. 2 paths fall at the reported transition points, the MBC detection survives without the contested labels. If the k-means solution instead merges MBC with MBL or ME, the supervised detection in Section IIIC is label-driven rather than an independent confirmation of the MBC phase. To test the substructure claim, repeat the same clustering protocol with k=4 and check whether a cluster boundary appears at lambda about 2 independent of the arbitrary cut.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is that the many-body MBC detection is not independent of the assumption it is meant to test. In Section IIIC the three-class network is trained with MBC labels taken directly from the phase diagram of Ref [28] (mu=2, lambda=0.3, 1, 2.5, 3), so the reported 99%/95% accuracy and the MBC boundaries in Fig. 2 merely confirm that those labelled regions are distinguishable in raw eigenvalue spacings and PDs; they do not establish that a distinct MBC phase exists. The binary classifier in Section IIIB provides the only label-free signal (P about 0.5 in the MBC region), but a network trained only on ME and MBL clusters will generically output intermediate confidence away from its training manifold; without a matched control the plateau is not diagnostic. The subphase claims in Section IIID and Appendix D are weaker still: the phase is divided at an arbitrary cut (lambda=2 for MBC-I/II, mu=1 for MBL-I/II) and a split is declared physical if the network probability is broad rather than peaked only at training points. This heuristic has no null model, no threshold, and no confidence intervals, and the same argument is applied to both the proposed subphases and the MBL crossover. The single-particle exponent agreement (Table III) is credible for nu, but the abstract's phrase 'identical critical exponents' overstates the result because w differs; the paper itself acknowledges that the many-body scaling collapse fails.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper applies supervised fully connected neural networks and unsupervised PCA to eigenvalue spacings and eigenvector probability densities of the interacting extended Aubry-André-Harper (EAAH) model, aiming to detect the many-body critical (MBC) phase, locate ME/MBC/MBL transitions, and identify subphases (MBC-I/II, MBL-I/II). In the single-particle limit, the authors use PCA entropy to locate delocalized-critical-localized transitions and perform finite-size scaling of a binary classifier trained on inverse participation ratio (IPR) components, comparing the extracted exponents with those from scaling of the IPR itself. The paper also tests a network trained on EAAH level spacings on an interacting flat-band diamond model where level statistics are inconclusive. The central conceptual claim is that ML can \"support the existence of the MBC phase\" and reveal genuine substructure, while the concrete quantitative result is that the single-particle correlation-length exponent ν extracted from classifier outputs matches ν from IPR scaling for three transitions.","tokens_in":23386,"tokens_out":4238,"duration_ms":44759,"significance":"If the label-free aspects of the claim could be established, the paper would provide a useful demonstration that raw eigenvalue spacings and eigenvector probability densities carry enough information for machine-learning phase classification in a quasiperiodic interacting system. The single-particle finite-size scaling results in Fig. 9 and Table III are internally consistent and give a concrete, falsifiable demonstration that a classifier output can reproduce the correlation-length exponent ν obtained from IPR scaling. The PCA-entropy diagnostics in Section IV B locate known single-particle transitions with reasonable accuracy, and the cross-model transfer experiment in Appendix C is intriguing. However, the most novel physical claims—independent detection of the MBC phase and identification of MBC and MBL subphases—are not supported by the present analysis because the supervised labels are taken directly from the phase diagram of Ref. [28] and the subphase criterion is an unquantified broadness heuristic. The paper is therefore best viewed as a promising methodology demonstration whose headline physics conclusions require additional label-free controls.","major_comments":[{"comment":"The three-class classifier is trained on labels taken directly from the phase diagram of Ref. [28] (ME: μ=0.3, λ=1, 2.5; MBC: μ=2, λ=0.3, 1, 2.5, 3; MBL: μ=0.5, 2, λ=4.5), so the reported 99% and 95% accuracies and the transition locations in Fig. 2 measure label consistency, not an independent confirmation of the MBC phase. The binary classifier's P≈0.5 plateau in the MBC region could also arise generically from a network trained only on two separated training clusters, without implying a distinct third phase. To support the Section III C claim that \"this analysis supports the existence of the MBC phase,\" the authors need a label-free control, such as unsupervised clustering of the same feature vectors or a network trained on two artificially split clusters and tested on the same parameter region, with a quantitative comparison.","section":"III C, Fig. 2"},{"comment":"The four-class subphase inference relies on an unquantified heuristic: a \"broad probability distribution across the phase\" is taken to indicate a genuine subphase, while a probability peaked only near training points is taken to indicate an artificial split. No null model, threshold, confidence interval, or alternative explanation (for example, finite-size interpolation near the training points) is provided. Since the phase is divided at arbitrary values (λ=2 for MBC-I/II, μ=1 for MBL-I/II) and the same heuristic is used for both the proposed subphases and the MBL crossover, the subphase and crossover claims are not currently supported. Please add a statistical test, such as label permutation, bootstrap resampling, or a comparison with a synthetic single-phase model, to show that broad probability is not a generic feature of four-class softmax outputs.","section":"III D, Appendix D"},{"comment":"The abstract's phrase \"identical critical exponents\" overstates the results: in Table III the correlation-length exponent ν agrees between IPR scaling and classifier scaling, but the exponent w differs substantially for the delocalized-to-localized transition (0.42 vs 0.03) and for the critical-to-localized transition (0.24 vs 0). Furthermore, the paper itself states that the many-body scaling collapse is not satisfactory, with τc and ν deviating from the values reported in Ref. [28]. The abstract and conclusion should make clear that the demonstrated agreement is for the single-particle correlation-length exponent ν only, that w differs by construction or by measurement, and that the many-body exponent extraction remains an open problem.","section":"V, Table III"},{"comment":"The many-body PCA results are presented as an effective phase indicator, but the transition locations are only described as \"roughly\" μ≈1 and λ≈2.7, no error bars or finite-size extrapolation are given, and the text explicitly acknowledges that the MBC-to-MBL transition is not clearly defined. This is acceptable as a preliminary observation, but the main text should consistently label these PCA-based boundaries as indicative rather than as established transition points, particularly since the preprocessing (rearranging normalized PDs) was chosen after observing the data structure in Fig. 7.","section":"IV C, Fig. 8"}],"minor_comments":[{"comment":"There is a typo in \"criitcal-to-localized\" in the sentence describing Fig. 11(c); it should read \"critical-to-localized.\"","section":"Appendix B"},{"comment":"The caption says \"The first principle component plotted against the second\"; the correct term is \"principal component.\" The same correction should be applied in the running text where the misspelling appears.","section":"Fig. 7 caption"},{"comment":"The energy window ε∈[0.53,0.55] is stated without justification. A sentence explaining why this window was chosen (for example, proximity to infinite-temperature states or avoidance of spectral edges) would improve reproducibility.","section":"II B"},{"comment":"The text reports that the cost function is minimized for specific (ν, w) pairs, but the cost landscape is not shown for all transitions. Showing CQ as a function of the scaling parameters, or at least stating the grid resolution and search range, would allow readers to assess how well the minimum is determined.","section":"V"},{"comment":"No code or data availability statement is provided. The exact diagonalization and TensorFlow-based training procedures described in Sections II and III would be much easier to reproduce if the scripts and generated datasets were released.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The core tension is that the paper's most marketable claims—detection of the MBC phase and of MBC/MBL subphases—are the ones least supported by the present evidence, while the more mundane single-particle scaling results are solid. I recommend major revision rather than rejection because the label-free control needed is well-defined and within the scope of the manuscript: the authors can retrain/classify without using Ref. [28] phase labels (e.g., clustering, permutation tests, or a null model for the broadness criterion). The authors should also clarify the relationship with the companion paper Ref. [31], which appears to already propose MBC-I/MBC-II and a μ≈1 crossover; the present manuscript should state explicitly what is new relative to that work to avoid overlapping-claim ambiguity."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is worth reading for the single-particle scaling section, but the many-body subphase claims do not survive contact with their own training labels. The three-class classifier is trained on the very phase diagram (Ref [28]) it claims to question, so 99%/95% accuracy mostly shows the labels are self-consistent. That part is a demonstration, not a discovery.\n\nWhat is genuinely new and good: using completely unprocessed eigenvalue spacings and eigenvector PDs as inputs works — that is not trivial. The binary classifier shows a clear uncertainty plateau in the MBC region, which is a nice sanity check that the phase is distinguishable from ME and MBL even without seeing MBC labels. Appendix C is the most convincing transfer test: the same network identifies nonergodic, thermal, and MBL-like phases in the ABF diamond chain where conventional level statistics are inconclusive. The single-particle finite-size scaling is the cleanest result: scaling collapse of the classifier probability trained on IPR components reproduces the correlation-length exponent nu (1, 0.4, 0.76) for all three transitions. That is a real, reproducible contribution.\n\nThe soft spots are concentrated in the many-body claims. The MBC detection is circular by construction — the network is trained on MBC labels taken from Ref [28], so the reported boundaries in Fig. 2 are consistency checks at best. The binary classifier plateau is suggestive but not diagnostic; a network trained only on ME and MBL will often return intermediate confidence far from its training manifold. You need a matched control or a label-free clustering method to make that point. The four-class subphase analysis is the weakest link. The split at lambda=2 (or mu=1) is arbitrary, and the criterion that a genuine subphase shows broad probability across parameter space while an artificial one peaks only near training points has no null model, no threshold, and no uncertainty. The authors apply the same heuristic to MBC-I/II and MBL-I/II, so the invented-entity concern in the reader's report is fair. The abstract's 'identical critical exponents' also overstates Table III: nu agrees, but w does not, and w is part of the scaling ansatz. The paper ships no code, no data, and no error bars, which makes it hard to verify the 99%/95% numbers.\n\nBottom line: the single-particle scaling result and the ABF transfer test are worth referee time. The many-body subphase claims need independent observables (multifractal spectra, entanglement entropy), error bars, and ideally a label-free benchmark before they can be endorsed. I would send it to peer review, but with a clear message that the headline claims need major revision. A careful reader gets value from the method and the honest appendices; just don't cite the subphase conclusion.","headline":"Useful single-particle scaling result wrapped around a circular many-body story; the subphase claims need independent validation before they can be taken seriously.","tokens_in":23911,"tokens_out":2305,"would_cite":false,"duration_ms":22223,"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 networks trained on raw eigenvalue spacings can find the many-body critical phase and its subphases, and reproduce known correlation-length exponents in the single-particle limit.","keywords":["many-body critical phase","supervised learning","unsupervised learning","principal component analysis","eigenvalue spacings","eigenvector probability densities","inverse participation ratio","finite-size scaling"],"falsifier":"Train the same three-class network on eigenvalue spacings from the claimed MBC region with the MBC labels replaced by randomly shuffled ME/MBL labels at the same parameter points; if accuracy stays near 99% and the MBC probability remains broad, the network is tracking label structure rather than a physical phase. Alternatively, extrapolate the subthermal volume-law entanglement coefficient to larger N across the MBC window: if the window shrinks with N, the detected phase is a finite-size artifact.","tokens_in":22753,"feed_emoji":"🧠","tokens_out":9522,"duration_ms":80441,"temperature":0.7,"pith_summary":"This paper asks whether a neural network can detect a disputed third phase—the many-body critical (MBC) phase—in a finite interacting quasiperiodic system using nearly unprocessed inputs. It claims that fully connected networks trained only on eigenvalue spacings or eigenvector probability densities separate MBC from the ergodic and many-body-localized phases, and that the network outputs locate phase boundaries consistently with conventional diagnostics. It further claims that the same outputs reveal two MBC subphases split near $\\lambda=2$ and a crossover inside the MBL region near $\\mu=1$. In the non-interacting limit, it claims that PCA entropy of eigenvector probability densities marks single-particle transitions, and that a binary classifier fed inverse-participation-ratio components recovers the correlation-length exponents $\\nu=1$, $0.4$, and $0.76$ found by scaling the IPR itself. A sympathetic reader would care because it suggests that generic, cheap spectral data can stand in for tailored order parameters in poorly understood disordered phases.","feed_headline":"Neural nets spot the many-body critical phase in raw spectra","feed_subtitle":"Trained on plain eigenvalue spacings, the classifiers match known boundaries and reproduce scaling exponents from standard methods.","key_machinery":"The central machinery is a set of fully connected neural networks that learn phase structure directly from raw spectral data: eigenvalue spacings $\\{E_{i+1}-E_i\\}$ and eigenvector probability densities $|\\psi_i|^2$ drawn from a narrow energy window. Binary classifiers expose the untrained MBC phase through output ambiguity near $0.5$, while multiclass classifiers with softmax outputs provide phase probabilities averaged over disorder samples; a confusion heuristic—a class whose probability stays broadly high across parameter space is genuine, while one that peaks only near training points is an artifact—drives the subphase claims. For unsupervised detection, PCA of eigenvector probability densities yields a PCA entropy $S_{PCA}=-\\sum_i p_i\\log_2 p_i$ whose numerical derivative peaks at transition points. For critical exponents, the paper scales the output probability of a binary classifier trained on IPR components using the ansatz $P=f(|\\tau-\\tau_c|N^{1/\\nu})$ and a cost function that measures the scatter of the collapsed curve.","core_discovery":"The paper's central claim is that the many-body critical phase in the interacting extended Aubry-André-Harper model can be separately detected by supervised neural networks fed with unprocessed eigenvalue spacings and eigenvector probability densities, without needing entanglement spectra or other sophisticated inputs. Trained on labels taken from the known phase diagram, binary classifiers give ambiguous output probability near $0.5$ precisely in the MBC region, and three-class classifiers assign probability near $1$ in each of the ME, MBC, and MBL phases. The resulting boundaries for the ME–MBC, ME–MBL, and MBC–MBL transitions are consistent with conventional measures. When the network is asked to split the phase diagram at $\\lambda=2$, it returns broad, training-point-independent probabilities for two MBC subphases but not for two ME subphases, which the authors interpret as a genuine crossover in the multifractal structure of many-body eigenstates. In the single-particle limit, PCA entropy jumps at the known transition points, and finite-size scaling of the classifier probability trained on IPR components gives correlation-length exponents identical to those from direct IPR scaling.","pith_inferences":["One testable extension is to train the multiclass network on labels from one model and apply it to a second model whose phase boundaries are known independently; the paper's Appendix C does one such transfer, but the MBC-training-label dependence on the first model's phase diagram remains untested.","The subphase claims rest on a heuristic about broad versus peak-like output probabilities; comparing the classifier probability surface against a null model trained on randomly split versions of the same phase would formalize that heuristic.","If the MBC-I and MBC-II split near $\\lambda=2$ is real, it may signal a change in the multifractal spectrum of many-body eigenstates rather than a thermodynamic transition, which could be tested by computing the full multifractal spectrum on each side at larger system sizes.","The PCA-entropy route might generalize to Floquet or open systems where exact eigenstate probability densities are replaced by steady-state density matrices, although the paper does not discuss that extension."],"forward_implications":["If the MBC detection is correct, the interacting EAAH model has interior structure near $\\lambda=2$ that level statistics alone cannot resolve.","If network outputs locate phase boundaries consistently, machine-classified boundaries could serve as a cheap first pass over parameter space before expensive entanglement or Fock-space diagnostics are run.","If IPR-component-trained classifiers reproduce $\\nu=1$, $0.4$, and $0.76$ for the three single-particle transitions, classifier probabilities inherit the scaling content of the input data and can act as a proxy observable for scaling collapse.","The transfer test in Appendix C, if reliable, indicates that eigenvalue-spacing patterns carry phase information even in models where gap ratios are inconclusive.","For the many-body system, the same scaling approach does not yet yield standard $\\tau_c$ and $\\nu$ at the available sizes, so critical exponents in the MBC region remain an open problem."],"supporting_citations":[{"why":"Supplies the interacting EAAH phase diagram and the MBC-phase claim whose boundaries the training labels and three-class classifier assume.","marker":"[28]"},{"why":"Predicts the crossover in multifractal properties near λ≈2 and μ≈1 that the four-class classifier results are read as confirming.","marker":"[31]"},{"why":"Motivates supervised networks as complex nonlinear regression that can outperform fixed observables, the logic behind the phase-detection approach.","marker":"[12]"},{"why":"Provides the all-bands-flat diamond-chain model used in Appendix C to test whether an eigenvalue-spacing-trained network transfers to a system with inconclusive level statistics.","marker":"[34]"},{"why":"Together with [46], defines the single-particle phase diagram and transition lines used for the PCA and classifier tests at U=0.","marker":"[39]"},{"why":"Gives the single-particle critical-to-localized transition at λ=3 and μ=1.5 used as a target for PCA-entropy and scaling analyses.","marker":"[46]"},{"why":"Supplies the scaling ansatz IPR=N^{-w}f(|τ-τ_c|N^{1/ν}) used for the finite-size collapse of IPR and classifier outputs.","marker":"[57]"},{"why":"Defines the cost function C_Q whose minimum determines the optimal collapse and hence the quoted critical exponents.","marker":"[58]"},{"why":"Identifies the exponent w as the fractal dimension of eigenstates at localization transitions, explaining why classifier and IPR scaling give different w values.","marker":"[59]"}],"fun_headline_variants":["Neural nets nail many-body critical phase from bare spectra","Raw spectra alone expose the many-body critical phase","PCA entropy flags many-body phases with no training","Scaling collapse of AI outputs reproduces critical exponents"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training labels assume that the phase diagram of a prior finite-size study is correct—that the many-body critical phase exists and sits where that study placed it—so the network's high accuracy is a consistency check on those labels rather than an independent discovery of the phase.","fun_headline_variants_meta":{"raw":{"variants":["Neural nets nail many-body critical phase from bare spectra","Raw spectra alone expose the many-body critical phase","PCA entropy flags many-body phases with no training","Scaling collapse of AI outputs reproduces critical exponents"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00103,"raw_usage":{"total_tokens":4417,"prompt_tokens":1097,"completion_tokens":3320,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":713,"completion_tokens_details":{"reasoning_tokens":3258}},"tokens_in":713,"tokens_out":3320,"duration_ms":23588,"temperature":1.0,"reasoning_tokens":3258,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:43:11.319863+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same three-class network on eigenvalue spacings from the claimed MBC region with the MBC labels replaced by randomly shuffled ME/MBL labels at the same parameter points; if accuracy stays near 99% and the MBC probability remains broad, the network is tracking label structure rather than a physical phase. Alternatively, extrapolate the subthermal volume-law entanglement coefficient to larger N across the MBC window: if the window shrinks with N, the detected phase is a finite-size artifact.","supporting_citations":[{"cited_title":"Ahmed, N","cited_arxiv_id":null,"evidence_quote":"Together with [46], defines the single-particle phase diagram and transition lines used for the PCA and classifier tests at U=0."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the single-particle critical-to-localized transition at λ=3 and μ=1.5 used as a target for PCA-entropy and scaling analyses."},{"cited_title":"Feng and S","cited_arxiv_id":null,"evidence_quote":"Defines the cost function C_Q whose minimum determines the optimal collapse and hence the quoted critical exponents."},{"cited_title":"Goodfellow, Y","cited_arxiv_id":null,"evidence_quote":"Identifies the exponent w as the fractal dimension of eigenstates at localization transitions, explaining why classifier and IPR scaling give different w values."}],"review_version":1}