{"id":"42abaab3-d7a2-49da-a979-5a1b2497cfee","arxiv_id":"2507.00298","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Aux-VAE splits a VAE latent space into supervised dimensions aligned with known generative factors and residual dimensions, using polynomial correlation penalties to enforce disentanglement.","lead":"This paper introduces Aux-VAE, a variational autoencoder that uses a few known physical parameters, such as galaxy brightness and radius, to shape its internal latent space so each known parameter gets its own dedicated code dimension. The authors test it on simulated galaxy images and standard benchmark datasets, reporting better disentanglement than two competing VAE methods while keeping image reconstruction quality.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"LDS may be computed on encoder means; conditional variance can make sampled latent space entangled despite high reported scores.","rationale":"The most load-bearing soft spot is the gap between what the objective regularizes (correlations of u with encoder means) and what the central claim asserts (disentanglement of the latent space). The SM's covariance derivation is exact for the numerator because q_phi(z|x) is independent of u given x, so the reader's phrasing of a 'within-encoder covariance term' is imprecise; the real issue is the denominator of the correlation: var(z_j) = E[var(z_j|x)] + var(mu_j). Since the regularizer and the LDS metric are both computed on mu_phi (assuming Table 1 is computed as in the SM), the reported scores can be high even when sampled latents are heavily noise-dominated or entangled. This is not merely a missing error bar; it questions whether the empirical evidence measures the claimed object. The paper's own admission that direct regularization of Corr(u_j, z_aux,j) may be needed for higher-complexity datasets reinforces the concern. Other reviewer concerns—hyperparameter selection using the evaluation metric, the unreported polynomial degree K, and lack of repeated seeds—affect the strength and generality of the comparison but do not change the interpretation of the metric if the metric measures the right thing. The proposed test is a single computational check on the released code: compare LDS computed from means versus from actual samples, and report conditional variances. If they match, the central claim stands and the paper is acceptable with minor clarifications; if they diverge, the headline result is an artifact of using means, and the work would need to be revised to optimize and evaluate sample-level disentanglement. Since the reader already flagged this premise and assigned CONDITIONAL, the verdict remains UNCHANGED.","tokens_in":19994,"tokens_out":11700,"duration_ms":124802,"concrete_test":"Retrain or use the released checkpoint for the galaxy Case 2 model. For the test set, compute Corr(u_j, z_l) in two ways: (a) using the encoder means mu_phi(x), and (b) using one Monte Carlo draw z ~ q_phi(z|x) per test point (or 20 draws with averaging), and evaluate LDS (Eq. 10) for both. Also report the mean conditional standard deviation of each z_aux dimension. If LDS_z is materially lower than LDS_mu (e.g., by more than 0.05) or if the conditional standard deviations are non-negligible relative to the scale of u, the claim that the latent space itself is disentangled is unsupported and the paper should be revised to regularize sample correlations and report sample-based metrics.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that Aux-VAE induces a statistically disentangled latent space, not merely aligned encoder means. However, the objective (Eq. 9) and its theoretical justification (SM Eqs. 17-20) regularize correlations between u and mu_phi. The covariance identity is exact (since z is conditionally independent of u given x), but the correlation denominator is not controlled: var(z_j) = E_x[var(z_j|x)] + var(mu_phi,j), so Corr(u_j, z_j) = Cov(u_j, mu_phi,j) / sqrt(var(u_j) * (E[var(z_j|x)] + var(mu_phi,j))). A model can have Corr(u_j, mu_phi,j) near 1 while Corr(u_j, z_j) is substantially lower if the encoder's conditional variance is large. The main text never states whether the LDS scores in Table 1 and Figure 3 are computed on latent samples z or on means mu; Figure 3 juxtaposes both, and the regularizer and SM use mu. If LDS is computed on mu, the reported 0.94 (galaxy Case 2) largely re-states the alignment that the KL term and regularizer directly enforce on means, and does not establish that the stochastic latent variables used by the decoder are disentangled. The authors themselves note (SM, end of Section A) that higher-complexity datasets may require directly regularizing Corr(u_j, z_aux,j). This is load-bearing because the paper's headline evidence for a disentangled latent space is the LDS table and the scatterplot; without sample-based figures, the central claim is not empirically grounded.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Aux-VAE, a variational autoencoder variant that partitions the latent space into auxiliary-informed factors (z_aux) and residual factors (z_recon). It uses a conditional prior p(z|u) = N((u_1,...,u_d,0,...), diag(1/n I_d, I_{dZ-d})) and augments the ELBO with polynomial-correlation regularizers (Eqs. 7-9) that encourage each auxiliary variable u_j to align with a distinct z_aux dimension while leaving z_recon free to encode unknown factors. The authors propose a new Linear Disentanglement Score (LDS, Eq. 10) and report experiments on a GalSim galaxy image dataset and on Cars3D and DSprites, comparing against beta-VAE and IDVAE. The main claims are that Aux-VAE achieves high LDS scores (up to 0.94 in galaxy Case 2) while preserving reconstruction quality on par with a standard VAE.","tokens_in":20303,"tokens_out":10034,"duration_ms":102991,"significance":"If the reported results hold, Aux-VAE is a practical and minimally invasive semi-supervised disentanglement method for scientific datasets with partial auxiliary information; the code is publicly available and the construction is simple to implement on top of a standard VAE. The paper also proposes the LDS metric, which is intuitively appealing though its definition requires correction. The principal technical risk is that the regularizer operates on encoder means rather than on sampled latent variables, and the manuscript does not clarify which quantity the reported LDS scores are computed on. This ambiguity affects the central claim that the stochastic latent space itself is disentangled.","major_comments":[{"comment":"The objective in Eq. (9) regularizes polynomial correlations between u and the encoder means mu_phi, not between u and sampled latent variables z. The covariance identity in SM Eq. (17) is valid for Cov(u, mu_phi) only when the within-encoder covariance term vanishes, and SM Eq. (20) explicitly gives Corr(u_j, z_aux,j) < Corr(u_j, mu_phi,j) unless E[var(z_j|x)] is negligible. The manuscript never states whether the LDS scores in Table 1 and Figure 3 are computed on z or on mu; Figure 3 displays both. If LDS is computed on mu, the reported scores largely restate the alignment directly enforced by the KL term and the regularizer, and do not establish that the stochastic latent variables used by the decoder are disentangled. The authors' own caveat at the end of SM Section A, that higher-complexity datasets may require directly regularizing Corr(u_j, z_aux,j), is load-bearing here; please report LDS on sampled z and, if necessary, regularize the sample-level correlation or provide evidence that the conditional variance E[var(z_j|x)] is negligible.","section":"Section 3, Eq. (9); Section 4.2.2; SM Section A, Eqs. (17)-(20)"},{"comment":"The polynomial degree K in Eqs. (7)-(8) is never specified in the experimental settings (Section 4.1). Because both sums run over k,k'=1..K with k≠k', for K=1 the regularizers R0 and R1 vanish identically, reducing the Aux-VAE loss in Eq. (9) to the plain VAE objective. The paper must report the value of K used for each dataset and justify the exclusion of same-degree (including linear-linear) correlation terms, which otherwise makes the 'explicitness' regularizer not directly target the stated alignment Corr(u_j, z_aux,j).","section":"Eqs. (7)-(9) and Section 4.1"},{"comment":"Table 1 reports single point estimates for LDS and SAP with no error bars, no repeated-seed runs, and no significance tests. Given the well-documented variance of disentanglement metrics across training runs (Locatello et al., 2019), the claimed margins (e.g., 0.94 vs 0.73 for galaxy Case 2) are not statistically grounded. Please report mean and standard deviation over at least five seeds, and use paired evaluations where possible.","section":"Table 1"},{"comment":"The LDS metric in Eq. (10) uses max_l Corr(u_j, z_l) in the numerator without an absolute value, while the denominator sums |Corr(u_j, z_l)|. As written, a factor that is perfectly anti-correlated with a latent dimension yields a negative contribution, contradicting the claimed range LDS ∈ [1/dZ, 1] and the interpretation that scores near 1 indicate optimal separation. The numerator should be max_l |Corr(u_j, z_l)|, or the metric should be redefined accordingly. This correction is essential because LDS is used both for hyperparameter selection (SM Section C.1) and for all reported comparisons.","section":"Eq. (10)"},{"comment":"The hyperparameter selection described in SM Section C.1 uses the product MSE(1 - LDS) on the validation split, where LDS is the same metric reported in Table 1. This makes the reported scores a selected maximum rather than an independent evaluation; the manuscript should acknowledge this and, ideally, evaluate a final model chosen without LDS on a held-out set.","section":"SM Section C.1"}],"minor_comments":[{"comment":"In the Explicitness proof, the limit of the correlation expression is written as '→ ∞' as n → ∞; the correct limit is '→ 1'.","section":"SM Section A"},{"comment":"The caption contains 'Card3D'; this should be 'Cars3D'.","section":"Table 1 caption"},{"comment":"The text says the metrics aggregate 'all possible polynomial combinations up to degree K', but the sums exclude k=k' terms; please rephrase to describe the actual definition.","section":"After Eq. (8)"},{"comment":"The caption does not specify whether the LDS values are computed from the grey dots (Z) or the maroon dots (µ); please clarify.","section":"Figure 3 caption"},{"comment":"The text refers to 'test MSE and test LDS' while describing a search on the validation split; make the terminology consistent.","section":"SM Section C.1"},{"comment":"The prior variance 1/n is presented without motivation; since n is the sample size, the prior becomes degenerate as n grows, so the authors should explain the rationale and the sensitivity to this choice.","section":"Eq. (3)"}],"recommendation":"major_revision","confidential_remarks":"The core idea is attractive and the availability of code is a plus. The main correctness risk is the ambiguity between z and mu in the LDS evaluation; if the high scores are only on encoder means, the paper's central claim is substantially weaker. I would also urge the editors to ensure the authors provide the value of K and a clear statement of whether LDS is computed on samples or means before considering publication. The use of the same metric for model selection and evaluation compounds this concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, quick take on arXiv:2507.00298: the central idea is worth taking seriously, but the evidence as presented is weaker than it looks. The paper proposes Aux-VAE, where the prior for an auxiliary block z_aux is centered on observed factors u with variance 1/n, residual latents z_recon carry the rest, and polynomial-correlation regularizers push the encoder means to align with u. That is a sensible, cheap way to get interpretable latent dimensions on scientific datasets where some ground-truth factors are known. The LDS metric is a useful bounded correlation-based alternative to SAP, and the galaxy simulation experiments nicely show that z_recon compensates when auxiliary factors are missing, with reconstruction SSIM roughly intact.\n\nThe soft spots are real but fixable. The R0/R1 definitions in Eqs (7)-(8) exclude k=k' pairs, so the \"explicitness\" regularizer R1(u_j, mu_aux,j) does not directly penalize low linear correlation between u_j and mu_j; it only uses cross-degree polynomials. Either the equation is a typo or the regularizer is not doing what the text says, and the polynomial degree K is never reported. More importantly, the text never states whether the LDS scores in Table 1 are computed on sampled latents z or on encoder means mu. The metric definition says z, but the regularizers and the SM's covariance identity work with mu, and the authors themselves note in the SM that direct regularization of Corr(u_j, z_aux,j) may be needed for complex data. If the reported 0.94 is on mu, it largely re-states the alignment the prior and regularizer enforce on means, and says little about the stochastic latents the decoder actually sees. That needs to be clarified.\n\nAlso, no error bars or repeated-seed runs anywhere in Table 1, and hyperparameters are selected using the same LDS metric used for evaluation. Minor, but it matters for a method paper. None of this sinks the core idea; it just means the current version overclaims empirical support. A careful revision should fix the regularizer definitions, report K, give seeded error bars, and state plainly whether LDS is mean-based or sample-based.\n\nVerdict: send it to peer review, but expect major revisions. It's a useful contribution for the scientific-ML crowd.","headline":"A plausible semi-supervised disentanglement method for scientific images, but the paper currently overclaims because the regularizer as written and the LDS evaluation targets may both be about encoder means, not the sampled latent space.","tokens_in":20857,"tokens_out":4911,"would_cite":false,"duration_ms":54821,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","62H20"],"pacs":[],"model":"deepseek-v4-flash","headline":"Aux-VAE aligns each known generative factor with a dedicated latent dimension while residual latents absorb unknown factors and preserve reconstruction.","keywords":["disentangled representation learning","variational autoencoder","auxiliary variables","posterior regularization","generative factors","Linear Disentanglement Score","scientific image datasets","galaxy simulations"],"falsifier":"Recompute the reported LDS and latent-traversal results using sampled latent values $z\\sim q_\\phi(z|x)$ instead of the encoder means $\\mu_\\phi(x)$; if the sampled-latent LDS falls substantially below the mean-based LDS on the galaxy dataset, the claim that the regularized objective disentangles the generative distribution rather than only its conditional means is false.","tokens_in":19748,"feed_emoji":"🔭","tokens_out":6573,"duration_ms":68971,"temperature":0.7,"pith_summary":"The paper introduces Aux-VAE, a modification of the variational autoencoder that uses observed auxiliary variables to separate known generative factors from everything else. It claims that by conditioning the latent prior on these auxiliary variables and adding two correlation-based regularizers, each known factor aligns with its own dedicated latent dimension while residual latent dimensions absorb unknown factors and keep reconstruction quality high. The authors report Linear Disentanglement Scores up to 0.94 on simulated galaxy images, well above the 0.73 of the closest competing method, with reconstruction matching a plain VAE. A reader should care because scientific datasets often come with partial side information about the physical parameters that generated the data, and this method promises to use that information to make learned representations interpretable without sacrificing generative accuracy.","feed_headline":"Aux-VAE gives each known generative factor its own latent dimension","feed_subtitle":"Semi-supervised VAE variant scores 0.94 on galaxy disentanglement while matching plain-VAE reconstruction.","key_machinery":"The central object is the conditional Gaussian prior of Eq. (3), which places the auxiliary values $u_j$ on the first $d$ latent coordinates with small variance $1/n$ and a standard normal prior on the remaining coordinates. Working from it, the paper derives inter-independence ($u\\perp z_{\\mathrm{recon}}$), intra-independence ($u_j\\perp z_{\\mathrm{aux},j'}$ for $j\\neq j'$), and explicitness ($\\mathbb{E}[z_{\\mathrm{aux},j}|u]=u_j$) as properties of the ideal posterior. Because the KL divergence to the expected variational posterior is intractable, the method replaces it with regularizers that measure, through polynomial powers of the variables, correlations between $u$ and the encoder mean $\\mu_\\phi$; these become the $R^K_0$ and $R^K_1$ terms in the loss. The Linear Disentanglement Score then evaluates the result by comparing the strongest absolute correlation between each $u_j$ and a latent dimension against the sum over all latent dimensions.","core_discovery":"The paper's core claim is that a VAE whose prior is set to $p(z|u)=\\mathcal{N}((u_1,\\ldots,u_d,0,\\ldots,0),\\operatorname{diag}(n^{-1} I_d, I_{d_Z-d}))$, with the latent vector partitioned into $z_{\\mathrm{aux}}$ and $z_{\\mathrm{recon}}$, and whose loss adds polynomial-correlation regularizers between the auxiliary variables and the encoder means, produces a latent space in which each known generative factor $u_j$ is captured almost exclusively by one latent dimension $z_{\\mathrm{aux},j}$. The residual dimensions are intentionally left entangled so they can collectively represent unknown factors, and the method is scored with a new bounded metric, the Linear Disentanglement Score, that measures how sharply each generative factor correlates with a single latent dimension. On simulated galaxy images the method achieves LDS scores of 0.88, 0.94, and 0.81 in three settings, versus 0.65, 0.73, and 0.59 for IDVAE, while SSIM reconstruction remains comparable to the baseline VAE.","pith_inferences":["Beyond the paper, a cheap strengthening for high-complexity data would regularize correlations with actual sampled latents rather than encoder means; the paper itself notes this alternative for finer-grained relationships.","Beyond the paper, because the method only needs paired observations $(x,u)$, it should transfer to non-image scientific measurements such as spectra or sensor time series with the same loss and latent partition.","Beyond the paper, recording both mean-based and sample-based LDS on every dataset would give a free check of whether the reported disentanglement holds at the level of the generative distribution or only its conditional averages."],"forward_implications":["On galaxy simulations, supplying all or most known physical parameters as auxiliary variables lets each parameter be controlled by turning a single latent dimension, enabling interpretable generative manipulation.","When important factors are missing from the auxiliary set, the residual latent dimensions automatically take over their representation, keeping reconstruction accuracy high.","The LDS metric provides a cheap way to score disentanglement from a single trained model, bounded between $1/d_Z$ and 1, without retraining supervised regressors.","Adversarial robustness experiments reported in the supplement indicate that Aux-VAE's disentangled representations degrade less under FGSM perturbation than entangled VAE representations."],"supporting_citations":[{"why":"Provides the IDVAE baseline that Aux-VAE is compared against and the semi-supervised double-VAE approach the paper extends.","marker":"Mita et al. (2021)"},{"why":"Supplies the beta-VAE baseline and the classical KL-weighting approach for unsupervised disentanglement.","marker":"Higgins et al. (2017)"},{"why":"Provides the expected variational posterior regularization idea and the SAP disentanglement score that LDS builds on.","marker":"Kumar et al. (2018)"},{"why":"Introduces identifiable VAEs with factorized exponential-family priors conditioned on auxiliary variables, the theoretical backdrop for auxiliary-guided disentanglement.","marker":"Khemakhem et al. (2020)"},{"why":"Generates the galaxy image dataset used in the main experiments.","marker":"Rowe et al. (2015)"},{"why":"Supplies the DSprites benchmark used to validate Aux-VAE beyond scientific images.","marker":"Burgess et al. (2018)"},{"why":"Supplies the Cars3D benchmark used to validate Aux-VAE on synthetic object images.","marker":"Reed et al. (2015)"},{"why":"Defines the SSIM metric used to compare reconstruction quality against baselines.","marker":"Wang et al. (2004)"}],"fun_headline_variants":["Aux-VAE: one latent dimension per known generative factor","Statistical prior drives VAE to split latent space by factor","Aux-VAE lets each generative factor claim its own latent axis","Semi-supervised VAE disentangles known factors into separate dimensions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that regularizing correlations with the encoder's mean values is enough to disentangle the actual latent samples; if the encoder's per-input variability is large, the sampled latent distribution can stay entangled even when the means align.","fun_headline_variants_meta":{"raw":{"variants":["Aux-VAE: one latent dimension per known generative factor","Statistical prior drives VAE to split latent space by factor","Aux-VAE lets each generative factor claim its own latent axis","Semi-supervised VAE disentangles known factors into separate dimensions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000199,"raw_usage":{"total_tokens":1345,"prompt_tokens":892,"completion_tokens":453,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":508,"completion_tokens_details":{"reasoning_tokens":381}},"tokens_in":508,"tokens_out":453,"duration_ms":5491,"temperature":1.0,"reasoning_tokens":381,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:20:25.908447+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the reported LDS and latent-traversal results using sampled latent values $z\\sim q_\\phi(z|x)$ instead of the encoder means $\\mu_\\phi(x)$; if the sampled-latent LDS falls substantially below the mean-based LDS on the galaxy dataset, the claim that the regularized objective disentangles the generative distribution rather than only its conditional means is false.","supporting_citations":[{"cited_title":", Filippone , M","cited_arxiv_id":null,"evidence_quote":"Provides the IDVAE baseline that Aux-VAE is compared against and the semi-supervised double-VAE approach the paper extends."},{"cited_title":", Matthey , L","cited_arxiv_id":null,"evidence_quote":"Supplies the beta-VAE baseline and the classical KL-weighting approach for unsupervised disentanglement."},{"cited_title":", Sattigeri , P","cited_arxiv_id":null,"evidence_quote":"Provides the expected variational posterior regularization idea and the SAP disentanglement score that LDS builds on."},{"cited_title":", Kingma , D","cited_arxiv_id":null,"evidence_quote":"Introduces identifiable VAEs with factorized exponential-family priors conditioned on auxiliary variables, the theoretical backdrop for auxiliary-guided disentanglement."},{"cited_title":", Zhang , Y","cited_arxiv_id":null,"evidence_quote":"Supplies the Cars3D benchmark used to validate Aux-VAE on synthetic object images."}],"review_version":1}