{"id":"610b6e8c-2827-412c-a2ef-a2d9674479ae","arxiv_id":"2509.06314","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A redundancy index based on energy distance between coupling-matrix off-diagonals and a Gaussian null is proposed and empirically correlated with accuracy and reconstruction quality.","lead":"This paper introduces a redundancy index, ρ(C), computed from the off-diagonal entries of a learned coupling matrix, and reports that low values correlate with better model performance. The authors position it as a tool for evaluating and guiding latent space design, but the theoretical grounding is inconsistent with the implemented algorithm.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1 does not compute the Fisher-transformed correlations required by Section IV.C, so the implemented ρ measures deviation from normality of raw coupling weights, not inter-dimensional redundancy.","rationale":"The reader's weakest assumption identifies exactly the mismatch between the theoretical Fisher-transform null and the robust standardization in Algorithm 1. This is the most load-bearing concern because the entire statistical grounding of rho rests on the claim that off-diagonal entries follow an approximate standard normal distribution under disentanglement. Section IV.C derives that null for sample correlations, but Algorithm 1 applies median/MAD standardization to raw learned weights, which has no such analytic justification. The paper's own text at the end of Section IV.E.a calls this weight-based variant an 'ablation' lacking a Fisher-z null, yet the experiments report results from this variant. This internal inconsistency makes the central claim that rho directly quantifies redundancy unsupported. The empirical correlations with accuracy and reconstruction could still be a useful heuristic, but the paper's stated contribution as a statistically grounded measure fails. The concrete test would settle whether the theoretical estimator and the implemented estimator give the same answers; if not, the paper's central claim must be revised. Since this is the same concern the reader raised, the verdict remains REJECT and no adjustment is needed.","tokens_in":12999,"tokens_out":5989,"duration_ms":52397,"concrete_test":"Recompute Experiments 1 and 5 with the theoretically defined estimator: compute the sample correlation matrix of the hidden-layer activations, apply the Fisher z-transform z_ij = sqrt(n-3) atanh(r_ij) to the off-diagonals, and calculate the energy distance of these z-values to N(0,1). If the accuracy-rho and reconstruction-rho scatter plots and the proposed thresholds (e.g., rho approximately 0.01) are not reproduced, the implemented Algorithm 1 does not measure the quantity the central claim relies on.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV.C establishes the Gaussian null by applying the Fisher z-transform to sample correlations: z_ij = sqrt(n-3) atanh(r_ij) is approximately N(0,1) under H0: rho_ij = 0. The redundancy index rho is then defined as the energy distance between the distribution of these z-values and N(0,1). Section IV.D and the plug-in estimator are written for these Fisher-transformed quantities. However, Algorithm 1 does not compute sample correlations nor apply the Fisher transform; it extracts the raw off-diagonal entries of the learned coupling weight matrix C and robustly standardizes them via median/MAD. Section IV.E.a explicitly labels this weight-based standardization an 'ablation' that 'lacks an analytic null like Fisher z,' yet this is the only estimator used in all experiments. The theoretical foundation therefore does not apply to the measured quantity. If experiments use Algorithm 1, rho measures deviation from normality of arbitrary learned weights, not inter-dimensional redundancy. The empirical accuracy-rho correlations may hold, but the central claim that rho is a statistically grounded redundancy measure is unsupported because the null model and the estimator are inconsistent. This is an internal inconsistency, not a disagreement with consensus, and it directly affects the interpretation of every reported rho value.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a redundancy index rho(C), defined as the energy distance between the distribution of off-diagonal entries of a coupling matrix and a standard normal distribution. The authors motivate this by a Fisher-transform null for sample correlations, then validate rho against classification accuracy and reconstruction loss across MNIST variants, Fashion-MNIST, CIFAR-10, and CIFAR-100, using MLPs and autoencoders with random and TPE-based hyperparameter search. They report that low rho predicts strong performance, that estimator variance decays with latent dimension, that TPE preferentially explores low-rho regions, and that consistent thresholds separate robust from redundant embeddings. The paper also positions rho as a potential training signal for redundancy-aware representation learning.","tokens_in":13253,"tokens_out":5330,"duration_ms":48551,"significance":"If the metric worked as claimed, rho would be a cheap, interpretable, unsupervised proxy for representational quality, with clear practical value for architecture search and model diagnosis. The paper has real strengths: a broad empirical campaign, explicit research questions, a companion code/data availability statement, and an honest distinction between the Fisher-based index and the weight-based \"ablation\" in Section IV.E.a. However, the central statistical claim is not supported by the implementation: the estimator used in all experiments is the ablation that the paper itself says lacks an analytic null, while the theoretical derivation applies to a different quantity. In addition, the predictive thresholds are derived from the same scatter plots on which they are evaluated, and no comparison against existing redundancy measures is provided. The empirical correlations are suggestive but do not currently establish a statistically grounded measurement tool.","major_comments":[{"comment":"The theoretical foundation is inconsistent with the implemented estimator. Section IV.C establishes the Gaussian benchmark for Fisher-transformed sample correlations, z_ij = sqrt(n-3) atanh(r_ij), approximately N(0,1) under H0: rho_ij = 0, and Section IV.D defines the redundancy index as the energy distance between the distribution of these z-values and N(0,1). Algorithm 1, however, extracts raw off-diagonal entries of the learned coupling matrix C and standardizes them with median/MAD; it does not compute sample correlations or apply the Fisher transform. Section IV.E.a explicitly labels this weight-based variant an \"ablation\" that \"lacks an analytic null like Fisher z\", yet this is the only estimator used in the experiments of Section V. Consequently every reported rho value measures deviation from normality of learned weights, not the statistically grounded inter-dimensional redundancy asserted in the abstract and introduction. This is an internal inconsistency, not a matter of presentation, and it directly affects the interpretation of all empirical results.","section":"Section IV.C, Section IV.D, Section IV.E.a, Algorithm 1"},{"comment":"The predictive claim is not validated out-of-sample. Statements such as \"low rho(C) reliably predicts high classification accuracy or low reconstruction error\" and the thresholds rho approximately 0.01 (accuracy) and 0.010-0.012 (reconstruction) are read off the same scatter plots from which the performance-redundancy relationship is inferred. No independent set of hyperparameter configurations, cross-fitting procedure, or out-of-sample prediction test is reported. Without such a test, the relationship is a descriptive summary of the fitting data rather than a demonstrated predictive relationship, so the abstract's causal-sounding claim is not supported.","section":"Section V, Experiment 1 and Experiment 5; Section V.A"},{"comment":"No existing redundancy or dependency measures are used as baselines. The paper motivates rho by the difficulty of mutual information estimation and cites Total Correlation and HSIC as alternatives, but the experiments never compare rho against these or any other metric in the accuracy/reconstruction prediction tasks. The appendix compares divergence-from-normality statistics on synthetic distributions, which does not address whether rho adds predictive value over established redundancy measures for learned representations. Without baselines, the contribution of rho relative to the existing toolkit is not established.","section":"Section II.B and Section V"},{"comment":"The theoretical link between the coupling matrix and representational redundancy is asserted rather than derived. Section IV.A argues that independent coordinates maximize joint entropy, and then states that in the ideal disentangled regime the coupling matrix C reduces to the identity and its off-diagonal entries behave as Gaussian noise. These are assumptions, not consequences of the preceding entropy discussion. The paper should either provide a derivation showing how learned coupling coefficients relate to latent correlations or mutual information, or substantially soften the theoretical claims. As written, the claim that rho is \"statistically grounded\" rests on an unproven modeling assumption.","section":"Section IV.A and Section IV.C"}],"minor_comments":[{"comment":"The violin plots of the redundancy estimator are referred to as \"Figure 2\", but the relevant figure appears to be Figure 4.","section":"Section V, Experiment 2, Findings"},{"comment":"The setup says \"across eight benchmark datasets\" but then lists seven datasets; please correct the count.","section":"Section V, Experiment 4, Setup"},{"comment":"The plug-in estimator in Section IV.D uses m unique off-diagonal entries (i<j), so m = d(d-1)/2, while Algorithm 1 sets n = d(d-1) and uses all off-diagonal entries. The notation should be harmonized, and the variance-scaling discussion should state which definition is used.","section":"Section IV.D and Algorithm 1"},{"comment":"There is a typo: \"contructive\" should be \"constructive\". The surrounding text also oscillates between calling redundancy a constructive mechanism and describing high rho as a \"redundancy explosion\" in Experiment 3; please clarify the intended interpretation.","section":"Section V.A"},{"comment":"Reference [16] is incomplete or inconsistent: the listed title and year do not match the arXiv identifier; please update the bibliographic details.","section":"References"},{"comment":"The table reports a negative energy distance (-0.0003) for the Gaussian baseline. Energy distance is nonnegative by definition; if this is the clipped estimator rho_+, please say so, otherwise report the raw value.","section":"Appendix, Table III"}],"recommendation":"reject","confidential_remarks":"I agree with the reader's assessment. The core problem is not disagreement with consensus but an internal inconsistency between the theoretical null (Fisher-transformed sample correlations) and the implemented estimator (robustly standardized raw weights). This is load-bearing: every reported rho value is computed with the variant the paper itself labels an ablation lacking an analytic null. Re-running the study with a principled estimator or deriving a valid null for the weight-based index would be a substantial rework, and the current manuscript does not support the abstract's claims. The empirical correlation is suggestive and might be salvageable in a major revision, but as submitted the central claim is unsupported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this paper proposes a redundancy index rho(C) for latent representations, based on energy distance between coupling-matrix off-diagonals and a Gaussian null, and shows empirically that low rho correlates with good accuracy/reconstruction across many datasets and architectures. That empirical pattern is worth knowing about. But the paper has a load-bearing internal inconsistency: the theory (Section IV.C) justifies the Gaussian null via the Fisher z-transform of sample correlations, while Algorithm 1—the estimator used in every experiment—standardizes raw coupling weights via median/MAD and does not compute correlations at all. The paper even labels the weight-based version an 'ablation' that 'lacks an analytic null,' yet that ablation is what all the results are based on. So the central claim that rho is a statistically grounded measure of redundancy is unsupported; at best it's a heuristic that measures deviation from normality of arbitrary learned weights.\n\nWhat's genuinely new: combining energy distance with coupling-matrix off-diagonals as a redundancy probe, and the extensive validation across MNIST variants, CIFAR, Fashion-MNIST, with MLPs and autoencoders. The observation that TPE naturally explores low-rho regions is a nice empirical finding, and the plots suggest there is a real signal there.\n\nSoft spots beyond the theory/algorithm mismatch: thresholds like rho<0.01 are read off the scatter plots post hoc, there are no baselines against existing disentanglement metrics (DCI, MIG, etc.), and the link to mutual information in Section IV is hand-wavy. The dimensionality experiments use random Gaussian matrices, not actual coupling weights from trained models, so the claim of a natural lower bound of n>=96 is only about the estimator's variance under null, not about real representations.\n\nIf the authors either compute Fisher-transformed correlations from the data or find a proper theoretical justification for the weight-based test, the paper could be salvageable. As is, the empirical correlation is suggestive but the metric's meaning is unclear. I'd send it to a serious referee—there's enough novel material and careful experimentation to merit a chance for major revision—but I wouldn't cite it yet, and I'd be cautious about accepting any of the quantitative thresholds.\n\nRecommendation: engage with it as a discussion piece, not as a reliable measurement tool.","headline":"Interesting heuristic but the estimator doesn't implement the theory—worth a referee, not yet citable.","tokens_in":13639,"tokens_out":2811,"would_cite":false,"duration_ms":24358,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"One statistic, the energy distance between coupling-matrix off-diagonals and Gaussian noise, flags redundant latent spaces and predicts which models generalize.","keywords":["representation learning","redundancy","disentanglement","coupling matrix","energy distance","latent space efficiency","hyperparameter optimization","autoencoder"],"falsifier":"Use a generative model with known independent latent factors, train an autoencoder on it, and compute $\\rho(C)$ from the learned coupling matrix; if $\\rho(C)$ stays above roughly $0.01$ despite the ground-truth factors being independent, the Gaussian-null assumption fails. A complementary check is to randomly permute latent coordinates before computing the coupling matrix, since a true pairwise-dependency index should be invariant under such permutations.","tokens_in":12842,"feed_emoji":"📉","tokens_out":8162,"duration_ms":70129,"temperature":0.7,"pith_summary":"Representation learning usually judges latent spaces by downstream accuracy or reconstruction loss, which cannot tell whether a space is wasteful because several coordinates carry the same information. The paper claims that this redundancy is directly measurable as $\\rho(C)$, the energy distance between the distribution of off-diagonal entries of a learned coupling matrix and a standard normal distribution. In a disentangled space those off-diagonals should resemble Gaussian noise, so large $\\rho(C)$ signals structured inter-dimension dependency. Across MLPs and autoencoders on MNIST variants, Fashion-MNIST, and the CIFAR datasets, the authors report that models with $\\rho(C) \\lesssim 0.01$ reach high accuracy or low reconstruction error, and that redundancy rises sharply as performance collapses. If the claim holds, $\\rho(C)$ gives a cheap, label-free diagnostic that could guide architecture search and eventually act as a training regularizer.","feed_headline":"One statistic flags redundant latent spaces and predicts accuracy","feed_subtitle":"The coupling-matrix index measures distance from Gaussian noise, forecasting which models generalize.","key_machinery":"The central object is the coupling matrix $C\\in\\mathbb{R}^{d\\times d}$, a learned linear map $\\tilde{h}=Ch$ inserted after a hidden layer, with the promise that off-diagonal entries $C_{ij}$ probe the dependency between latent coordinates $i$ and $j$. The index is the energy distance between the distribution $Z$ of those off-diagonals and the standard normal benchmark, $\\rho(C)=\\mathrm{ED}(Z,\\mathcal{N}(0,1))$, computed by a plug-in estimator that uses the closed forms $\\mathbb{E}|G-G'|=2/\\sqrt{\\pi}$, $\\mathbb{E}|z-G|=2\\phi(z)+z(2\\Phi(z)-1)$, and a U-statistic for $\\mathbb{E}|Z-Z'|$. Algorithm 1 extracts the off-diagonals, standardizes them with median and MAD, and assembles the energy distance, reporting $\\hat{\\rho}_+=\\max\\{0,\\hat{\\rho}\\}$. The work the machinery does is to turn the abstract notion of redundancy into a scalar whose null distribution is known, so that variance shrinks systematically with the number of off-diagonal samples and thresholds around $0.01$ can be compared across models and tasks.","core_discovery":"The paper's central claim is that representational redundancy reduces to a single discrepancy number: $\\rho(C) = \\mathrm{ED}(Z(C), \\mathcal{N}(0,1))$, where $Z(C)$ is the distribution of the standardized off-diagonal entries of a coupling matrix $C$ and $\\mathrm{ED}$ is the energy distance. The null hypothesis is that an ideally disentangled representation has uncorrelated coordinates, so its Fisher-transformed off-diagonals are approximately standard normal; the index measures how far a learned representation departs from that baseline. The empirical backbone is a systematic correlation: low $\\rho(C)$ tracks high classification accuracy and low reconstruction error, elevated $\\rho(C)$ tracks generalization collapse, the estimator's variance shrinks roughly as $1/\\sqrt{m}$ with $m=d(d-1)/2$ off-diagonal samples, and search methods such as Tree-structured Parzen Estimators concentrate their trials in low-$\\rho$ regions. The paper also finds that redundancy is not always harmful, since the best wide networks on CIFAR-100 carry elevated first-layer $\\rho$ while still generalizing well.","pith_inferences":["If the index genuinely isolates redundancy, it should also predict transfer-learning performance and robustness to corruption, and a direct test would compare $\\rho(C)$ with linear-probing transfer scores on the same checkpoints.","The mismatch between the Fisher-z null and Algorithm 1's median-MAD standardization suggests a sharper version of the paper could be built by actually estimating pairwise correlations of latent activations and Fisher-transforming them, either confirming the current implementation or revealing that its success comes from a different statistic.","The observation that controlled redundancy can be beneficial implies $\\rho$ should be interpreted as an optimal band rather than a quantity to minimize; a testable extension is to regularize $\\rho(C)$ only above a threshold while training VAEs or classifiers and see whether accuracy or reconstruction quality improves.","Since the estimator's variance scales with the number of off-diagonal samples, the index becomes most informative for large latent spaces, and applying it to very small bottlenecks may require a different null or a bias correction that the paper does not address."],"forward_implications":["A model selector could rank trained networks without labels: configurations with $\\rho(C)$ near zero can be trusted to generalize, while those with $\\rho(C)$ above the observed threshold can be discarded or retrained.","The variance analysis gives a practical floor: with $m=d(d-1)/2$ off-diagonal samples, redundancy estimates become dependable only once the latent dimension is large enough, and the paper's experiments adopt $d \\geq 96$.","Architecture design gains a target: excessive width in early layers inflates $\\rho(C)$ and creates redundancy explosions without proportional accuracy gains, so width balancing can be guided by $\\rho$ trends.","Hyperparameter optimization can be made redundancy-aware: since TPE naturally samples low-$\\rho$ regions, $\\rho(C)$ could replace or supplement the acquisition function and steer search away from redundant configurations.","The same threshold applies to generative models, so autoencoder monitoring can use $\\rho(C)$ as an early-warning signal for reconstruction collapse."],"supporting_citations":[{"why":"Formulates disentanglement as statistical independence of latent variables, the ideal regime that defines the paper's null hypothesis.","marker":"[1]"},{"why":"Provides the factorized independence criterion that motivates measuring redundancy in latent coordinates.","marker":"[2]"},{"why":"Shows mutual information estimation is expensive and unstable in high dimensions, motivating the lightweight energy-distance alternative.","marker":"[5]"},{"why":"Supplies the prior dependency-score paradigm the paper positions $\\rho(C)$ against as more interpretable.","marker":"[8]"},{"why":"Introduces coupling layers as invertible transformations, the architectural idea repurposed here as a measurement probe.","marker":"[12]"},{"why":"The sequential model-based optimization method whose low-$\\rho$ exploration is a central finding.","marker":"[15]"},{"why":"Introduced the structured MNIST variants used to test the index across noise, rotation, and background conditions.","marker":"[18]"}],"fun_headline_variants":["ρ(C): a single number that flags redundant latent spaces","Latent redundancy index predicts accuracy and guides search","Distance from Gaussian noise reveals latent space efficiency","One metric links latent redundancy to generalization collapse"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire index rests on the claim that a perfectly disentangled representation makes the coupling matrix's off-diagonal entries look like standard Gaussian noise, but the code applies a robust standardization to raw learned weights rather than the Fisher-transformed correlations the theoretical null is proven for, so the Gaussian benchmark may not be the right null.","fun_headline_variants_meta":{"raw":{"variants":["ρ(C): a single number that flags redundant latent spaces","Latent redundancy index predicts accuracy and guides search","Distance from Gaussian noise reveals latent space efficiency","One metric links latent redundancy to generalization collapse"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000829,"raw_usage":{"total_tokens":3662,"prompt_tokens":1025,"completion_tokens":2637,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":641,"completion_tokens_details":{"reasoning_tokens":2578}},"tokens_in":641,"tokens_out":2637,"duration_ms":17398,"temperature":1.0,"reasoning_tokens":2578,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:17:29.930466+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Use a generative model with known independent latent factors, train an autoencoder on it, and compute $\\rho(C)$ from the learned coupling matrix; if $\\rho(C)$ stays above roughly $0.01$ despite the ground-truth factors being independent, the Gaussian-null assumption fails. A complementary check is to randomly permute latent coordinates before computing the coupling matrix, since a true pairwise-dependency index should be invariant under such permutations.","supporting_citations":[{"cited_title":"Disentangling by factorising,","cited_arxiv_id":null,"evidence_quote":"Provides the factorized independence criterion that motivates measuring redundancy in latent coordinates."}],"review_version":1}