{"id":"5c343867-6272-4698-b998-a8d5ed6949ba","arxiv_id":"2412.01023","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"HypStructure trains image representations whose pairwise distances follow a label tree by adding a hyperbolic CPCC regularizer and a centering loss, reducing hierarchy distortion and modestly improving classification and OOD detection.","lead":"This paper adds a hyperbolic-space regularizer to representation learning so that image features respect a given label hierarchy, improving low-dimensional embeddings and out-of-distribution detection. It matters because hierarchical structure is cheap to obtain, yet most representation learners treat labels as independent.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline distortion claim conflates hyperbolic distance with two other algorithmic changes; without a Euclidean control using internal nodes and the centering loss, the hyperbolic attribution is untested.","rationale":"The reader's verdict is CONDITIONAL, and I agree that the paper needs additional evidence before the full claims are accepted. However, I think the more decisive missing experiment is not hierarchy noise (a scope limitation already acknowledged in Section 7) but the missing Euclidean control that would isolate the hyperbolic distance. The ablation in Table 5 is not conclusive because rows 2–4 change multiple components at once, and the text's claim that removing ωhyp or Tint individually harms performance is not directly supported by clean pairwise comparisons. The eigenvalue analysis in Section 5, Theorem 5.1, also depends on CPCC being maximized and on an idealized block-structured K; it is an interpretive result, not evidence of hyperbolic advantage over Euclidean structured regularization. The empirical distortion and accuracy numbers appear internally consistent and are supported by released code, so this is not a rejection. But the central scientific attribution—that hyperbolic geometry is what makes the method work—is currently untested, and one additional baseline experiment would settle it.","tokens_in":35091,"tokens_out":5876,"duration_ms":55208,"concrete_test":"Train an 'Euclidean HypStructure' baseline on CIFAR10 and CIFAR100 (and ideally ImageNet100) that uses the same full tree with internal nodes, the same centering loss, and the same α, β, architecture, but replaces the Poincaré distance in HypCPCC with ℓ2 distance on the same (clipped) features and replaces hyperbolic averaging with Euclidean averaging. Report δrel, test CPCC, fine/coarse kNN accuracy, and OOD AUROC over the same 3 seeds. If the Euclidean version closes most of the gap to HypStructure (e.g., δrel within 10% relative), the hyperbolic geometry is not the driver; if it does not, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that HypStructure reduces hierarchy distortion and improves downstream tasks relative to ℓ2-CPCC (Table 1, Figure 4). But HypStructure (Section 3.2, Equation 8) differs from ℓ2-CPCC [104] in at least three ways simultaneously: it uses hyperbolic distance in the CPCC term, it embeds internal tree nodes (Tint), and it adds a centering loss (HypCenter). The ablation in Table 5 varies components but never includes the combination 'internal nodes plus centering loss with Euclidean centroids and Euclidean distance,' so the unique contribution of hyperbolic geometry cannot be isolated. The ablation rows suggest that internal nodes and centering alone account for a substantial part of the accuracy gain (e.g., fine accuracy 76.22 to 76.91 on CIFAR100), and it is unknown whether they also account for the distortion reduction. If a Euclidean CPCC with the full tree and the same centering loss achieves comparable δrel and test CPCC, then the paper's contribution is better described as 'tree-complete structured regularization' rather than hyperbolic embedding. This is load-bearing because the title, abstract, and Section 1 motivate the method specifically by the failure of Euclidean ℓ2 distance to embed trees, and the claimed 45–59% distortion reduction is the headline result.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces HypStructure, a structured regularizer that adds a hyperbolic-space CPCC term over the label tree and a hyperbolic centering loss to a standard flat training loss. The method is evaluated on CIFAR10, CIFAR100, and ImageNet100 using SupCon as the main flat loss, with additional results for hyperbolic SupCon losses and hyperbolic backbones. The reported experiments compare HypStructure to Flat and ℓ2-CPCC baselines on Gromov distortion (δrel), CPCC, fine/coarse accuracy, and OOD detection AUROC on nine OOD datasets. The paper also presents an eigenspectrum analysis meant to explain why structured features help OOD detection, with a balanced-tree theorem and a general-tree extension in the appendix.","tokens_in":35381,"tokens_out":11542,"duration_ms":110408,"significance":"The paper addresses a meaningful problem — explicitly injecting label-hierarchy structure into learned representations — and the proposed regularizer is simple, general, and accompanied by a released implementation and detailed training details. The distortion reductions reported in Table 1 are substantial, and the combination of a theoretical eigenvalue analysis with an empirical OOD study is ambitious. The main weakness is that the current experimental design does not isolate the effect of hyperbolic geometry from the simultaneous use of internal tree nodes and the centering loss, and a few empirical claims are not fully supported by the tables. With an additional Euclidean control and more cautious reporting, the contribution could be a useful advance for hierarchy-aware representation learning.","major_comments":[{"comment":"The central claim that hyperbolic distance is responsible for the reduced distortion is not supported because HypStructure changes three things at once: the CPCC distance is changed from Euclidean to Poincaré, internal tree nodes T_int are included (whereas the ℓ2-CPCC baseline in Table 1 is leaf-only), and the centering loss ℓ_center is added. The ablation in Table 5 never evaluates a Euclidean full-tree CPCC with a Euclidean centering loss, and it reports only fine/coarse accuracy, not δrel or CPCC. A Euclidean variant with internal nodes and centering is therefore needed to determine whether the 45.4% distortion reduction over ℓ2-CPCC is due to hyperbolic geometry or to the complete-tree CPCC and centering. Please add this control and report δrel and CPCC for all ablation rows.","section":"Section 3.2 (Eq. 8), Section 4.1 (Table 1), Appendix C.2 (Table 5)"},{"comment":"The statement that HypStructure consistently improves the test CPCC is not accurate as written. On CIFAR100, HypStructure has CPCC 0.766 (0.007), which is lower than ℓ2-CPCC's 0.779 (0.002). Since CPCC is one of the two hierarchy-quality metrics used in the paper, please either correct this statement, report per-dataset numbers clearly, or explain why a lower CPCC is expected on CIFAR100 despite the lower δrel.","section":"Section 4.1, Table 1"},{"comment":"The OOD results are reported without standard deviations or significance tests, and the appendix states that the reported OOD numbers use the Euclidean-centroid computational variant of HypStructure while Section 4.3 does not disclose this. Because OOD AUROC improvements over prior methods are often around 1–2 points, repeated-run variability can matter. Please state which variant is used in the main tables, report variance or confidence intervals, and clarify whether the Euclidean-centroid variant is also the one used for the δrel and CPCC tables.","section":"Section 4.3, Tables 2 and 7a, Appendix C.3"},{"comment":"The theorem assumes an exactly block-constant Gram matrix K with entries r_h determined by the height of the lowest common ancestor. The proof, however, only establishes that the relative order of Poincaré distances matches Euclidean distances for vectors clipped to the boundary of the ball; general interior points of the Poincaré ball do not have this monotonicity, and the empirical CPCC values in Table 1 are below 1. Please state Theorem 5.1 as a conditional characterization of the idealized CPCC-optimal case, quantify the perturbation from the measured K, and soften the claim that the experiments formally characterize the observed structured features.","section":"Appendix A, Corollary A.1 and Theorem 5.1"}],"minor_comments":[{"comment":"The claim of 'up to 2.2% improvements over Flat' is not directly visible in Table 1; on CIFAR100 the fine accuracy gain is 1.72 points and the coarse gain is 1.86 points. Please clarify the origin of the 2.2% figure.","section":"Section 4.2"},{"comment":"The displayed formula for the Poincaré distance appears to have malformed parentheses and line breaks; please typeset it correctly.","section":"Eq. (3)"},{"comment":"The centering loss is written as ∥HypAveB(˜z1,...,˜zB∥) in the main text with a missing closing parenthesis; the pseudocode should be checked.","section":"Algorithm 1, line 11"},{"comment":"The dependence on a correct external hierarchy is acknowledged, but no experiment with noisy or misspecified hierarchies is provided; since both the CIFAR10 and ImageNet100 hierarchies are constructed, a small robustness study would strengthen the claims.","section":"Section 7"}],"recommendation":"major_revision","confidential_remarks":"The paper's core regularizer is sensible and the experiments are extensive. My recommendation is major revision because the hyperbolic attribution needs a missing control and because several reporting issues — the CIFAR100 CPCC discrepancy, missing OOD variance, and the undisclosed variant switch — affect the strength of the claims rather than the feasibility of the method. I would be willing to look at a revision with those points addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Useful paper, with one real confound that needs addressing. The authors extend the Euclidean CPCC regularizer of Zeng et al. (2022) to hyperbolic space, embed internal tree nodes instead of leaves only, and add a centering loss. They report large distortion reductions (up to 59% vs Flat, 45% vs l2-CPCC) and the main classification table includes standard deviations over three seeds. The code is released, the appendix has a genuine ablation table and a worked eigenvalue analysis, and the limitations section honestly flags the dependence on an external hierarchy. That is a solid empirical package.\n\nThe main soft spot is the one the stress-test note flags: the comparison conflates three changes at once. The l2-CPCC baseline is leaf-only Euclidean; HypStructure uses hyperbolic distances, internal nodes, and a centering loss. The ablation varies components but never includes the combination full tree + Euclidean centroids + Euclidean distance + centering. Without that control, the 45% distortion reduction cannot be attributed to hyperbolic geometry. It could come just as well from adding internal nodes and centering to a Euclidean CPCC. That is load-bearing because the title, abstract, and motivation make the hyperbolic claim, and the paper argues for hyperbolicity based on the impossibility of exact tree embeddings in l2.\n\nThe accuracy gains are real but modest, and on CIFAR10/ImageNet100 they are within about one standard deviation of Flat. OOD tables have no error bars; on CIFAR100 the average gains over SSD+ and KNN+ are around 2-3 AUROC points, plausible but not overwhelming. The eigenvalue theorem is a clean conditional statement given a block-structured Gram matrix with level-wise correlations, but it is not derived from the actual training objective, and its link to OOD detection is suggestive rather than proven. The robustness concern about noisy hierarchies is fair but minor, and the paper acknowledges it as future work.\n\nThis paper is for researchers working on hierarchical representation learning or structure-aware regularizers. If the authors add the missing Euclidean full-tree control and some OOD error bars, it becomes a dependable contribution. It deserves peer review, not a desk rejection; I would recommend major revision on the attribution issue.","headline":"Useful extension of Euclidean CPCC to hyperbolic space with a real distortion-reduction result, but the headline hyperbolic attribution is confounded by a missing Euclidean full-tree control; deserves refereeing with revisions.","tokens_in":35871,"tokens_out":3320,"would_cite":true,"duration_ms":30286,"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":"HypStructure moves the CPCC hierarchy-matching loss into a Poincaré ball and claims up to 59.4% lower distortion than flat training and up to 45.4% lower than Euclidean ℓ2-CPCC, with gains in classification and OOD detection.","keywords":["hyperbolic geometry","structured representation learning","label hierarchy","CPCC","out-of-distribution detection","eigenvalue analysis","Poincaré ball","supervised contrastive learning"],"falsifier":"On a fixed benchmark such as CIFAR10, train HypStructure and ℓ2-CPCC with the same SupCon base loss while randomly corrupting a growing fraction of fine-to-coarse assignments (for example 0%, 10%, 25%, and 50%). If HypStructure's test CPCC and fine accuracy do not degrade monotonically with hierarchy noise, or if it degrades more slowly than ℓ2-CPCC, the claim that hyperbolic geometry gives a fidelity advantage in the realistic noisy-hierarchy regime would be contradicted. Alternatively, finding any matched configuration in which ℓ2-CPCC achieves lower δrel than HypStructure would directly falsify the distortion claim.","tokens_in":34901,"feed_emoji":"📉","tokens_out":11265,"duration_ms":99197,"temperature":0.7,"pith_summary":"HypStructure is a regularizer that forces a learned feature space to respect the tree structure of the class labels by computing the CPCC objective—the Cophenetic Correlation Coefficient, a Pearson correlation between tree distances and feature distances—in hyperbolic rather than Euclidean space. The paper argues that Euclidean methods cannot faithfully embed even simple label trees, whereas hyperbolic geometry can, and that this fidelity pays off concretely: on CIFAR10, CIFAR100, and ImageNet100, HypStructure lowers the distortion of hierarchical information by up to 59.4% relative to flat training and up to 45.4% relative to the Euclidean ℓ2-CPCC baseline, while improving test CPCC and, in most settings, classification accuracy and OOD detection AUROC. The contribution is a drop-in regularizer that works with ordinary cross-entropy or supervised contrastive losses and with hyperbolic backbones, plus an eigenvalue analysis that connects the geometry of hierarchy-informed features to out-of-distribution detection.","feed_headline":"Hyperbolic regularizer cuts hierarchy distortion by up to 59.4%","feed_subtitle":"A plug-in loss for label trees also lifts classification and OOD detection on vision benchmarks.","key_machinery":"The load-bearing object is the HypStructure regularizer, -α HypCPCC(d_T, d_B^c) + β ℓ_center, added to any flat task loss. HypCPCC generalizes the Euclidean CPCC by taking Poincaré distances between hyperbolic class prototypes, computed by mapping encoder outputs into the ball and averaging them with the Einstein/Klein midpoint; HypCenter keeps the root node's prototype near the ball's center, following the low-distortion tree-embedding construction of [75]. The theoretical machinery is the block-structured Gram matrix K = ZZ^\top induced by CPCC maximization, whose eigenspectrum is characterized recursively in Theorem 5.1: at each level h of a balanced tree, C_h - C_{h+1} eigenvalues equal λ_h = λ_{h-1} + (r_h - r_{h+1}) C_0 / C_h, with the coarsest level dominating.","core_discovery":"The central claim is that a CPCC-style structured regularizer, previously computed with Euclidean centroids, should be moved into a Poincaré ball to represent label hierarchies. The paper shows that the Euclidean version provably cannot embed a tree without distortion and that the hyperbolic version—with hyperbolic class centroids and a centering loss that keeps the root at the origin—produces more tree-like features, with δrel dropping from 0.232 to 0.094 on CIFAR10 and test CPCC reaching 0.992. It also proves Theorem 5.1, which states that for a balanced label tree the Gram matrix K = ZZ^\top of CPCC-optimized features has eigenvalue groups with a phase transition at each hierarchy level, and argues this coarse-level eigenvalue concentration is what makes hierarchy-informed representations good for OOD detection.","pith_inferences":["Editorial inference: the paper tests only correct, externally given hierarchies; a natural stress test is to corrupt the fine-to-coarse assignments and measure how quickly the distortion, accuracy, and OOD gains erode relative to ℓ2-CPCC.","Editorial inference: the phase-transition theorem suggests a general diagnostic—the size and location of eigenvalue gaps in any representation's Gram matrix could reveal whether and at what granularity a label hierarchy is encoded, independent of hyperbolic training.","Editorial inference: the root-at-origin centering idea may transfer to partial orders that are not trees, such as DAGs or multiple-inheritance hierarchies, where the tree-metric CPCC assumption does not hold but hierarchical containment still matters.","Editorial inference: the OOD explanation relies on coarse-level separation, so the benefit may weaken for fine-grained OOD drawn from the same coarse classes as the in-distribution data; that regime is not tested in the paper."],"forward_implications":["Using HypStructure with supervised contrastive training makes learned features significantly more tree-like: on CIFAR10, δ_rel falls from 0.232 with flat training to 0.094 with HypStructure, while test CPCC rises from 0.573 to 0.992.","Because hyperbolic low-dimensional representations can beat the distortion of 512-dimensional Euclidean ones, hierarchy regularization is a way to recover representative capacity when the feature dimension is small.","CPCC-style structured regularization produces an eigenspectrum whose dominant components align with coarse classes; for CIFAR100 the top 20 eigenvectors carry most of the OOD signal, improving Mahalanobis-score AUROC over flat and ℓ2-CPCC baselines on nine OOD datasets.","HypStructure remains beneficial when the flat loss is replaced by a hyperbolic supervised contrastive loss or the backbone by a clipped hyperbolic neural network, so the regularizer composes with non-Euclidean training as well."],"supporting_citations":[{"why":"It defines the Euclidean ℓ2-CPCC structured regularizer that HypStructure extends, and it serves as the main baseline.","marker":"[104]"},{"why":"It defines the Cophenetic Correlation Coefficient, the objective optimized by both ℓ2-CPCC and HypCPCC.","marker":"[79]"},{"why":"It provides the low-distortion tree-embedding construction whose root-at-origin placement inspires the hyperbolic centering loss.","marker":"[75]"},{"why":"It introduces Poincaré-ball embeddings for hierarchical symbolic data, motivating the use of hyperbolic geometry.","marker":"[66]"},{"why":"It supports the claim that Euclidean distance in bounded dimension distorts tree-like semantic structure.","marker":"[8]"},{"why":"It supplies the block-structured correlation matrix eigenstructure theorem used as the base of the eigenspectrum proof.","marker":"[3]"},{"why":"It provides the eigenvalue perturbation bound used to extend the phase-transition analysis to arbitrary trees.","marker":"[93]"},{"why":"It supplies the Poincaré distance and exponential/logarithm map formulas used to compute hyperbolic prototypes and distances.","marker":"[88]"},{"why":"It provides the Supervised Contrastive loss used as the flat task loss in the main experiments.","marker":"[39]"}],"fun_headline_variants":["Hyperbolic loss cuts label-tree distortion by 59%","Hyperbolic space reduces hierarchy distortion in features","Hyperbolic embeddings keep label trees intact","Hyperbolic regularizer boosts OOD detection and tree fidelity","Moving structured loss to hyperbolic space cuts distortion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the label hierarchy fed into HypStructure is correct and complete; if the supplied or constructed hierarchy is noisy or wrong, the regularizer can push features into a structure that does not match the data, and the reported gains may disappear.","fun_headline_variants_meta":{"raw":{"variants":["Hyperbolic loss cuts label-tree distortion by 59%","Hyperbolic space reduces hierarchy distortion in features","Hyperbolic embeddings keep label trees intact","Hyperbolic regularizer boosts OOD detection and tree fidelity","Moving structured loss to hyperbolic space cuts distortion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000655,"raw_usage":{"total_tokens":2990,"prompt_tokens":922,"completion_tokens":2068,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":538,"completion_tokens_details":{"reasoning_tokens":1997}},"tokens_in":538,"tokens_out":2068,"duration_ms":17031,"temperature":1.0,"reasoning_tokens":1997,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:45:29.129326+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a fixed benchmark such as CIFAR10, train HypStructure and ℓ2-CPCC with the same SupCon base loss while randomly corrupting a growing fraction of fine-to-coarse assignments (for example 0%, 10%, 25%, and 50%). If HypStructure's test CPCC and fine accuracy do not degrade monotonically with hierarchy noise, or if it degrades more slowly than ℓ2-CPCC, the claim that hyperbolic geometry gives a fidelity advantage in the realistic noisy-hierarchy regime would be contradicted. Alternatively, finding any matched configuration in which ℓ2-CPCC achieves lower δrel than HypStructure would directly falsify the distortion claim.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It defines the Euclidean ℓ2-CPCC structured regularizer that HypStructure extends, and it serves as the main baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It defines the Cophenetic Correlation Coefficient, the objective optimized by both ℓ2-CPCC and HypCPCC."},{"cited_title":"Nickel and D","cited_arxiv_id":null,"evidence_quote":"It introduces Poincaré-ball embeddings for hierarchical symbolic data, motivating the use of hyperbolic geometry."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the Poincaré distance and exponential/logarithm map formulas used to compute hyperbolic prototypes and distances."}],"review_version":1}