{"id":"2febf971-1a9b-4d78-9bd8-1b9adc41ea76","arxiv_id":"2506.10089","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A dataset-specific geometric split of latent dimensions across HVAE layers improves OOD detection over fixed baseline configurations.","lead":"This paper studies how to allocate a fixed total number of latent dimensions across the layers of a hierarchical variational autoencoder (HVAE) used for out-of-distribution (OOD) detection. It tests a geometric allocation rule on several image datasets and finds that the best compression ratio is dataset-specific and generally beats fixed baseline splits.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 1's existence proof for r* is invalid: (0,1] is not compact and Eq. (10) is singular at r=1, so the Extreme Value Theorem cannot yield a maximizer.","rationale":"The reader's verdict is CONDITIONAL, and I agree with that overall disposition. The single most load-bearing defect is narrower than the reader's stated weakest assumption. The one-parameter geometric family and the unknown f are scope limitations; the decisive problem is internal to the proof. Proposition 1 uses EVT on a non-compact domain and on a formula singular at r=1. The explicit counterexample above satisfies the stated regularity conditions yet has no maximizer, so the 'proof of existence' in the abstract is formally false. This matters because the theoretical guarantee is the first listed contribution and the basis for calling the framework 'theoretically grounded.' It does not necessarily sink the empirical findings: the paper reports a systematic r-sweep across multiple datasets, repeated runs with standard deviations, standard OOD metrics, and control allocations, and those results could support a weaker claim that on these datasets some geometric ratios outperform baseline allocations. The recommended revision is to replace the theorem with a well-posed statement—existence on compact subintervals [ε,1] or a grid-search optimum—or add conditions ensuring the supremum is attained. The model-selection protocol should also be clarified so r* is not chosen on the same OOD sets used for evaluation, but that is secondary to the invalid proof. I would keep the verdict CONDITIONAL rather than moving it: the empirical contribution is plausible, but the theoretical claim needs correction or removal before acceptance. Agreement with the reader is partial: their rationale notes the compactness problem, but their stated weakest assumption focuses on the geometric-family parameterization.","tokens_in":22485,"tokens_out":12561,"duration_ms":153345,"concrete_test":"Construct the counterexample in the attack: N=2, b=2, f(l)=(2-l)^2 for l∈[0,2] (constant 0 for l>2), compute F(r)=Σ f(l_i) with Eq. (10). If F is strictly decreasing on (0,1] with supremum only at r→0+, Proposition 1 is false as stated; no maximizer exists on (0,1]. As a second check, re-derive Eq. (10) at r=1: the denominator 1-r^N vanishes, so the proof must supply a separate definition of l_i(1) and show continuity before EVT applies.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing theorem is Proposition 1 (Sec. 3.3) and its supplementary proof (Sec. 8): it claims r* = argmax_{r∈(0,1]} F(r) exists by calling (0,1] compact and applying EVT. But (0,1] is not compact; r=0 is excluded, so continuous functions on it need not attain maxima. This is not pedantry: under the paper's own assumptions, the maximum can fail to be attained. Take N=2, b=2, f(l)=(2-l)^2 (continuous, bounded, lim_{l→0} f=4, f'(l)≤0 for the relevant l∈[0,2]). Then l1=2/(1+r), l2=2r/(1+r), so F(r)=4(1+r^2)/(1+r)^2. On (0,1] this is strictly decreasing; sup F=4 is approached as r→0+ but never attained, so no r* exists. Also, Eq. (10) is undefined at r=1 (denominator 1-r^N=0), so F is not even defined on the claimed domain without a separate limit definition, which the proof does not provide. Thus the existence of r* is not established. The empirical grid search can still find a best grid point, but it cannot substitute for the claimed theoretical guarantee.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript studies how the distribution of a fixed total latent dimensionality across layers of a hierarchical VAE affects out-of-distribution detection. It restricts allocations to a geometric progression with a single compression ratio r, defines a utility F(r)=Σ_i f(l_i) for an unspecified layer utility f, and claims (Proposition 1) that an optimal ratio r* exists on (0,1]. The empirical part sweeps r ∈ {0.1, 0.25, 0.5, 0.75} for FashionMNIST, MNIST, and CIFAR10 in-distribution sets, evaluates four OOD detection metrics with several OOD counterparts, and reports that dataset-specific r* values outperform the tested baseline and control allocations. It also estimates I(X;Z3) to motivate an information-bottleneck interpretation.","tokens_in":22757,"tokens_out":8373,"duration_ms":77928,"significance":"The empirical protocol has notable strengths: fixed latent budgets and depth for most configurations, multiple OOD pairings per in-distribution set, standard metrics with standard deviations, and non-geometric control configurations. If the results are reproducible, the paper would offer simple practical guidance for HVAE architecture design. However, the central theoretical result is not established: the proof of Proposition 1 is invalid (and the proposition is false under the stated assumptions), the claimed cross-pairing generalization of r* is circular given the way r* is selected, and a fixed-budget inconsistency in the CIFAR10 setting undermines the controlled comparison. These are load-bearing issues, so the paper requires substantial revision before the contributions can be accepted.","major_comments":[{"comment":"The proof of Proposition 1 is invalid. The interval (0,1] is not compact, so the Extreme Value Theorem cannot be applied; in addition, the expression for li in Eq. (10) has denominator 1-r^N and is undefined at r=1, so F is not defined on the stated domain without a limit convention. The failure is not a technicality: under the paper's own assumptions the maximum can fail to exist. With N=2, b=2, and f(l)=(2-l)^2, the assumptions of Proposition 1 hold (f is continuous, f'(l)≤0 for l∈[1,2], and lim_{l→0} f(l)=4), but F(r)=4(1+r^2)/(1+r)^2 is strictly decreasing on (0,1], so its supremum 4 is approached as r→0+ and never attained. Proposition 1 as stated is therefore false, and the claimed existence guarantee must either be corrected with additional assumptions or removed.","section":"Section 3.3 and Section 8, Eqs. (17)-(20) and (27)-(36)"},{"comment":"The claim that r* 'generalizes across OOD pairings' is circular. Section 5.2 defines r* for each in-distribution dataset as the ratio yielding the highest average performance across exactly the OOD pairings listed in the same section, and Section 6.1 then reports that this same r* is 'consistently optimal across its corresponding OOD pairings.' That statement is true by construction. A non-circular evaluation would select r* on one set of OOD datasets and test it on held-out OOD datasets, or would clearly separate model selection from evaluation; without such a split, the cross-pairing generalization claim is unsupported.","section":"Section 5.2 and Section 6.1"},{"comment":"The fixed-budget constraint for CIFAR10 is not met by the reported configurations. The text states b=228 for CIFAR10, but every compressed configuration in Figure 5 and Table 3 sums to 224 (e.g., 128+64+32, 170+43+11, 202+20+2, 97+73+54), and the control configuration 56→80→98 sums to 234. Since the paper's central comparison is explicitly 'under a fixed latent budget,' this discrepancy must be resolved: either the actual budget is 224 and the text is wrong, or the displayed configurations violate the stated budget and the corresponding results need to be rerun or re-reported.","section":"Section 4.3, Figure 5, and Table 3"},{"comment":"The mutual information estimator is presented without any error analysis or validation, yet it is used to draw the paper's information-theoretic conclusions. Equation (23) replaces the intractable marginal log-likelihood with the ELBO, and the non-target latents are sampled from their priors; these substitutions introduce an unknown bias that is not bounded or compared against a tractable ground-truth case. Before the RQ3 claims about information loss and attenuation can be accepted, the paper should quantify the estimator's bias (e.g., on synthetic settings where I(X;Zi) is known) or substantially weaken the interpretation.","section":"Section 3.5 and Figure 6"}],"minor_comments":[{"comment":"Several typos interrupt the exposition, including 'exhibitsuboptimal' and 'which limits. the model'; a careful proofread is needed.","section":"Sections 1 and 2.2"},{"comment":"The caption says 'Only the top latent z3 is inferred from the posterior; lower latents are sampled from the prior,' which appears to contradict the hierarchical inference described in Eq. (2) and Figure 1; please clarify which inference scheme is used in the reconstructions.","section":"Figure 2 caption"},{"comment":"The notation 'dVar(I(X;Z3))' is not defined; specify whether this is the variance across runs or across samples.","section":"Section 13.1.2"},{"comment":"Table 2 contains malformed entries (e.g., '1.±0.' and truncated Omniglot columns) that must be repaired before publication.","section":"Table 2"},{"comment":"The paper should state explicitly that Proposition 1 concerns optimality only within the one-parameter geometric family defined by Eq. (10); the current wording 'optimal latent compression ratio' can be misread as a global optimum over all allocations.","section":"Section 3.3, Eq. (13)"}],"recommendation":"major_revision","confidential_remarks":"This manuscript's empirical contribution is potentially salvageable, but the advertised theoretical result is false as stated and the generalization claim is circular. If the authors cannot supply a correct existence theorem, I recommend reframing the paper as an empirical study of geometric latent allocation. The budget mismatch in the CIFAR10 experiments should be checked before any revision is returned."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the empirical part of this paper is worth paying attention to, but the theoretical proof that the paper leans on does not hold. The claim that an optimal geometric compression ratio r* exists is not established; the domain (0,1] is not compact and the objective is undefined at r=1. The stress-test counterexample is right: with N=2, b=2, f(l)=(2-l)^2, F(r)=4(1+r^2)/(1+r)^2, which is strictly decreasing on (0,1] and approaches 4 as r->0+ without attaining it. So no r* exists under the paper's own assumptions. That is a load-bearing flaw, because the abstract promises a proof.\n\nWhat is genuinely new: the systematic sweep of geometric allocation ratios for HVAE OOD detection. Prior work used fixed architectures. The experiments show consistent, if modest, gains: best r is dataset-specific (0.75 for FashionMNIST, 0.5 for MNIST, 0.25 for CIFAR10), and the compressed configurations beat several hand-designed controls. The mutual information plots give a plausible story about attenuation vs. information loss. That is a useful design guideline for practitioners.\n\nWhere it is soft: first, the theorem. I already gave the counterexample; the proof also says (0,1] is compact, which is false. The 'for practical purposes' phrase in the supplementary is not a fix. Second, the optimal r is chosen by grid search on the same OOD pairings that are later reported as 'generalization across pairings.' That is selection on the evaluation metric, not independent prediction. Third, the MI estimator is a heuristic: it substitutes the ELBO for the intractable marginal and ignores the KL at layer i; there is no error analysis. Fourth, no code or data are released; with 48-200 hour training runs per model, that is a real barrier to verification.\n\nThe empirical finding could survive without the theorem. If the authors reframe the contribution as 'the compression ratio is a meaningful and tunable design axis, and here is a consistent empirical pattern,' the paper is credible. If they keep the existence proof, it needs to be fixed or removed. I would send this to a serious referee, but it requires major revision before acceptance. It is the kind of paper where a good referee could help the authors separate the useful empirical claim from the overreach.","headline":"The empirical sweep makes a credible design point, but the existence proof for r* is invalid; this needs major revision.","tokens_in":23283,"tokens_out":2451,"would_cite":false,"duration_ms":26966,"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":"The paper claims that a dataset-specific optimal latent compression ratio r* exists for hierarchical VAEs under a fixed latent budget, and that tuning to it improves out-of-distribution detection.","keywords":["out-of-distribution detection","hierarchical variational autoencoder","latent dimension allocation","geometric compression ratio","information bottleneck","mutual information","posterior collapse"],"falsifier":"Train every integer allocation ($\\ell^1$,$\\ell^2$,l3) with $\\ell^1$ ≥ $\\ell^2$ ≥ l3 ≥ 1 and $\\ell^1$ + $\\ell^2$ + l3 = 32 on FashionMNIST and test on MNIST, KMNIST, and notMNIST; if any non-geometric triple beats the reported r* = 0.75 allocation (14→10→8) on AUROC or FPR95, then r* is not optimal in the full allocation space. A second check: with any explicit smooth concave f, maximize F(r) symbolically and compare its argmax to the empirical r*; disagreement shows that the claimed optimum depends on the unknown f.","tokens_in":22209,"feed_emoji":"🎯","tokens_out":6419,"duration_ms":74146,"temperature":0.7,"pith_summary":"This paper argues that how a hierarchical VAE (a VAE whose latent variables are split into several layers of abstraction) divides a fixed latent-dimension budget among its layers is not a neutral design choice but a tunable quantity with a real optimum. It restricts allocations to a geometric progression governed by a single compression ratio r and proves, under mild regularity assumptions on an unknown per-layer utility function, that an optimal ratio r* exists. Empirically, the paper finds distinct dataset-specific optima—0.75 for FashionMNIST, 0.5 for MNIST, and 0.25 for CIFAR-10—and shows that tuning to r* improves AUROC, AUPRC, FPR80, and FPR95 over baseline allocations without changing the model size. This matters because OOD detection is safety-critical, and the proposed tuning offers a cheap, architecture-level way to improve it without new training objectives.","feed_headline":"Dataset-specific ratio r* sharpens OOD detection in hierarchical VAEs","feed_subtitle":"Tuning how latent dimensions shrink across layers beats fixed allocations in OOD tests.","key_machinery":"The central object is the latent dimension ratio r with the geometric allocation formula li = b(1-r)$r^{{i-1}}$/(1-r^N), which compresses each successive latent layer by the same factor while exactly consuming the fixed budget b. This one-parameter family turns architecture search into a one-dimensional optimization over F(r) = sum_i f(li), and the proof that an optimum exists uses continuity and compactness of the domain (0,1] together with boundedness of f. The same formula also generates the layer sizes used in all experiments, so the theoretical and empirical claims rest on the same mechanism.","core_discovery":"For a hierarchical VAE with N latent layers and a fixed total latent budget b, the paper parameterizes layer dimensionalities as li = b(1-r)$r^{{i-1}}$/(1-r^N), so that the entire architecture is described by one ratio r. Defining total efficacy as F(r) = sum_i f(li) for an unknown utility function f, the paper claims that under continuity, saturation for large li, and a finite limit as li goes to 0, F attains its maximum at some r* in (0,1]. Empirically, the paper reports that r* is dataset-specific and consistently outperforms baselines: FashionMNIST r* = 0.75 (14→10→8), MNIST r* = 0.5 (18→9→5), and CIFAR-10 r* = 0.25 (170→43→11), with gains across four OOD metrics and multiple OOD pairings per in-distribution dataset. Mutual-information estimates I(X; Z3) are low at r = 0.1, matching information-loss failures, and high but increasingly variable as r approaches 1, matching attenuation and overfitting, which the paper takes as evidence for the claimed tradeoff.","pith_inferences":["My inference: the existence proof applies to any continuous bounded f, so the theorem itself is broad but nearly content-free; the empirical r* values are the substantive result, and they are established only within the one-parameter geometric family.","My inference: r* likely tracks the difficulty of the OOD task—for far-OOD pairs almost any allocation succeeds, so the practical gains should concentrate on near-OOD discrimination.","My inference: the mutual-information estimator samples non-target latents from the prior, so the reported I(X; Z3) measures information flow through one specific generative path; a different estimator could change the attenuation picture.","My inference: relaxing the constant-r constraint to per-layer ratios or a learned allocation could improve performance further, which the paper itself leaves open."],"forward_implications":["Under a fixed latent budget, sweeping r and selecting the dataset-specific r* improves OOD detection metrics over baseline allocations without increasing model size.","The optimal r* transfers across different OOD pairings for a given in-distribution dataset, so practitioners can tune once per dataset rather than per OOD task.","Extreme compression (r = 0.1) consistently causes information loss and poor detection, while relaxed compression (r near 1) raises mutual information without improving detection, so both extremes should be avoided.","Non-geometric heuristic allocations such as expand-then-compress, stable, and compress-then-expand underperform optimized geometric allocations, indicating that structured compression is doing real work."],"supporting_citations":[{"why":"Supplies the ladder-style hierarchical latent structure and bottom-up inference that the paper's allocation scheme builds on.","marker":"[7]"},{"why":"Provides the baseline HVAE OOD detection method, the LLR score, and the dataset pairings used for comparison.","marker":"[8]"},{"why":"Prior HVAE-based OOD detection with an adaptive likelihood ratio; the paper positions its allocation framework as an improvement over it.","marker":"[11]"},{"why":"Formalizes the information bottleneck principle that underlies the claimed compression-versus-utility tradeoff.","marker":"[34]"},{"why":"Gives generalization bounds linking latent dimensionality to estimation error, used to argue that large latents cause attenuation and overfitting.","marker":"[35]"},{"why":"Empirical evidence that increasing latent dimensionality improves OOD detection only up to a point of diminishing returns, motivating the existence of an optimum.","marker":"[37]"}],"fun_headline_variants":["Optimal latent ratio r* boosts OOD detection in HVAEs","One ratio to rule latent dimensions: r* for OOD","Tuning latent shrink ratio r* improves OOD flags","Latent allocation ratio r*: key to OOD robustness","Single ratio r* optimizes hierarchical VAE for OOD"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes the best latent allocation is a single compression factor shared by every adjacent pair of layers, and that the unknown per-layer utility function is continuous, saturates for large layers, and has a finite limit as layer size goes to zero.","fun_headline_variants_meta":{"raw":{"variants":["Optimal latent ratio r* boosts OOD detection in HVAEs","One ratio to rule latent dimensions: r* for OOD","Tuning latent shrink ratio r* improves OOD flags","Latent allocation ratio r*: key to OOD robustness","Single ratio r* optimizes hierarchical VAE for OOD"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000784,"raw_usage":{"total_tokens":3473,"prompt_tokens":969,"completion_tokens":2504,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":2427}},"tokens_in":585,"tokens_out":2504,"duration_ms":19567,"temperature":1.0,"reasoning_tokens":2427,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:35:32.690296+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train every integer allocation ($\\ell^1$,$\\ell^2$,l3) with $\\ell^1$ ≥ $\\ell^2$ ≥ l3 ≥ 1 and $\\ell^1$ + $\\ell^2$ + l3 = 32 on FashionMNIST and test on MNIST, KMNIST, and notMNIST; if any non-geometric triple beats the reported r* = 0.75 allocation (14→10→8) on AUROC or FPR95, then r* is not optimal in the full allocation space. A second check: with any explicit smooth concave f, maximize F(r) symbolically and compare its argmax to the empirical r*; disagreement shows that the claimed optimum depends on the unknown f.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior HVAE-based OOD detection with an adaptive likelihood ratio; the paper positions its allocation framework as an improvement over it."}],"review_version":1}