acceptodds
Under review as a conference paper at ICLR 2027

Hyperbolic Gravitational Cone: A Physically-Inspired Approach for Partial Order Embeddings

Abstract

Partial-order modeling is essential for capturing asymmetric entailment semantics in knowledge graph reasoning and multimodal hierarchical alignment, which necessitates an inductive bias capable of strictly preserving transitivity. While hyperbolic entailment cones provide rigorous geometric guarantees for this property, their efficacy is limited by two constraints: geometric rigidity and optimization stagnation. Specifically, isotropic apertures mask semantic heterogeneity, and binary objectives sever gradient flow for entailed pairs. To address these issues, we propose Hyperbolic Gravitational Cone (HGC), inspired by the physical principle that mass governs semantic influence. The HGC formulates the entailment scope as a dynamic gravitational field, yielding a closed-form cone aperture modulated jointly by a learnable anisotropic mass and radial depth. To ensure hierarchical consistency, this mass function is strictly governed by positivity, radial monotonicity, convexity, and directional anisotropy. Furthermore, a continuous gravitational energy is derived to replace rigid binary constraints. By employing a soft-gated mechanism to evaluate angular deviation and coupling it with parental mass and axial distance, this energy function provides informative gradients even within the cone interior, driving precise structural optimization. Crucially, radial transitivity is mathematically guaranteed by imposing strict monotonicity on mass decay, preserving logical nesting along any given radial direction. Empirical evaluations demonstrate that HGC achieves superior performance on standard partial-order benchmarks, where the learned mass distributions reveal the underlying structure of semantic hierarchies and quantify the intrinsic breadth of concepts.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Partial-Order Representation, Hyperbolic Gravitational Cone, Anisotropic Semantic Scope, Gravitational Potential Energy

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_VWP3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies partial-order (e.g., is-a hierarchy) embeddings based on hyperbolic entailment cones. It identifies two limitations of existing cone methods such as Hyperbolic Entailment Cones (HEC) and Shadow Cones: (1) geometric rigidity, since a cone's aperture is determined solely by the node's radial position and is therefore isotropic; and (2) optimization stagnation, since binary containment objectives assign zero energy, and hence zero gradient, to entailed pairs that already lie inside the cone.

The proposed Hyperbolic Gravitational Cone (HGC) replaces HEC's global aperture constant with a learnable anisotropic "mass" m(r,Ω)=m0 (1+α(Ω) e−βr)m(r,\Omega) = m_0\,(1 + \alpha(\Omega)\,e^{-\beta r}), where the directional factor α(Ω)\alpha(\Omega) is produced by a small Möbius-layer network, and the cone aperture follows in closed form. A finite-difference regularizer enforces radial monotonicity and convexity of the mass, under which the paper proves ray-wise transitivity. Training minimizes a "gravitational energy" within a max-margin loss: a softplus-gated angular term weighted by the parent's (stop-gradient) mass and by a hyperbolic distance, measured to the cone axis for positive pairs and to the cone boundary for negative pairs. The paper also argues that HEC and Shadow Cones are special cases of HGC.

Experiments evaluate hierarchy reconstruction, i.e., predicting held-out transitive-closure edges, on WordNet Mammal, WordNet Noun, and the Microsoft Concept Graph. Embedding dimensions range from 2 to 10, and the fraction of transitive edges included in training is varied. Baselines are box embeddings, HEC, and Shadow Cones variants. HGC reports the best F1 in most settings, e.g., 80.5 vs. 69.4 on Mammal (d = 2, 50%). The appendix adds ablations of the anisotropic mass and the energy function, as well as visualizations of the learned cones and the energy landscape.

Critical strengths

1. A simple and interpretable generalization of entailment cones. The paper clearly articulates two limitations of existing cone embeddings: isotropic apertures, and zero gradients for entailed pairs inside the cone. It addresses them with lightweight, well-motivated changes: a learnable, direction-dependent mass that yields a closed-form aperture, and a softplus-gated energy. HEC is recovered as a special case (α = 0), so the idea is easy to understand and could be dropped into existing cone-based models.

2. Clear gains in low-dimensional and low-supervision settings. HGC performs particularly well when capacity or supervision is limited. For example, it reaches 80.5 vs. 69.4 F1 on Mammal at d = 2 with 50% of transitive edges, and 55.3 vs. 52.6 on Noun at d = 10 when only the transitive reduction is used for training. HGC also trains from a simple uniform initialization without a separate pre-training stage, and uses a single formulation across datasets instead of requiring a choice among Shadow Cones variants.

Critical weaknesses

Significance of the problem. It is difficult to assess how important the two problems addressed in this paper are. The abstract motivates partial-order modeling by its role in knowledge graph reasoning and multimodal hierarchical alignment, yet the experiments contain no results in either setting. Could the authors provide evidence, e.g., statistics or results from prior work, that such hierarchical structure plays a key role in practical NLP tasks?

Experiments. Overall, the experimental evaluation is too thin:

  1. Datasets and applications. The evaluation covers only hierarchy reconstruction on WordNet Mammal, WordNet Noun, and MCG, and the method is not validated in any real application. Since the abstract highlights knowledge graph reasoning and multimodal alignment, I would expect at least one experiment in such a setting.

  2. No direct evidence that the two problems are solved. I agree that, in principle, the anisotropic mass and the gravitational energy address the fixed-aperture and optimization-stagnation problems. However, beyond end-to-end F1 (including the ablations in Appendix I, which are shown only as plots and are not discussed in the main text), the experiments do not measure the effect of either component directly. For the learnable directional factor that controls the cone aperture, what does it actually achieve compared with an isotropic aperture? For example, do broader concepts receive wider cones (e.g., a correlation between aperture and subtree size), and do both false positives and false negatives decrease? Precision and recall are not reported separately. Similarly, I could not find any measurement of optimization stagnation, such as training curves, gradient norms of positive pairs inside the cone, or how far inside the cone the positive pairs lie, for HEC versus HGC. The qualitative evidence in Figs. 7 and 8 relies on a few selected nodes.

  3. Comparison with language models. Given the rapid progress of language models, a comparison with LM-based approaches would be informative, e.g., embeddings from a pretrained encoder such as BERT, or LM-based hypernymy prediction on the same datasets. Such models use concept names, which the current structure-only setting does not, so this would not be a like-for-like baseline. I also do not mean that the proposed method lacks value if a language model performs better. Rather, such a comparison would clarify where geometric embedding methods currently stand.

Presentation. The font size in the figures is too small to read comfortably.

Questions for authors

see weaknesses

The paper is interesting, and the proposed method is fairly novel. However, its main weakness is that the experimental evaluation is too thin. If the authors can provide additional relevant or key experiments, I would consider raising my score.

3 comments

We sincerely thank the reviewer for the thorough, insightful, and constructive feedback. Below, we address each of your concerns in detail.

Q1: Significance of the problem.

R1: Real-world conceptual knowledge is intrinsically structured as hierarchical DAGs, meaning that encoding partial-order constraints establishes the intrinsically correct problem formulation that aligns with the data’s true logic. Crucially, semantic heterogeneity is ubiquitous in practical NLP ontologies: concepts residing at identical hierarchical depths often govern radically disparate sub-branch breadths (e.g., in WordNet, organism spans >14,000 descendants whereas its sibling monotreme spans <10), a fundamental asymmetry that isotropic cones rigidly fail to capture. Extensive literature across fine-grained classification, knowledge graph reasoning, and multimodal alignment has proven that enforcing geometric containment is vital for eliminating symmetric retrieval fallacies and preserving zero-shot generalization across abstraction levels. Following established literature (HEC, Shadow Cones), we evaluated pure hierarchical reconstruction to isolate and benchmark this geometric inductive bias without confounding artifacts from pre-trained text encoders, and we will explicitly integrate these practical task statistics and connections into the revised manuscript.

Q2: No direct evidence that the two problems are solved.

R2: We thank the reviewer for this constructive critique. We completely agree that end-to-end F1 alone is insufficient to isolate the individual contributions. We provide a correlation between aperture and subtree size (whether the learned mass m(u,Ωu) and aperture ψ(u) meaningfully capture intrinsic concept breadth rather than merely memorizing hierarchical depth (Review 2, Q4)), and compared with BERT (with Precision and recall reported in Q3), and explicitly qualify the transitivity violated ratio across directions (Review 1, Q2).

  1. Precision and Recall Breakdown (Eliminating Both FPs and FNs): On Mammal (𝑑=5, 50% split), HGC achieves 82.3% Precision and 97.6% Recall. Rather than trading off false positives against false negatives, the learned adaptive aperture concurrently prevents over-expansion (reducing FPs) and under-coverage (reducing FNs). (data in Q3)
  2. Interior Gradient Norms (Resolving Optimization Stagnation): Across all ten WordNet Mammal settings, HGC produced a nonzero Riemannian gradient with respect to the child embedding for every strictly interior positive pair (Ξ <𝜓), with mean norms ranging from 0.01782 to 0.02951. The angular hinge yielded zero gradients on the same embeddings and pairs, showing that HGC’s positive loss continues to provide an optimization signal after a hyponym enters its hypernym’s cone.
Interior gradient norm on Mammal0%10%25%50%90%
HGC (2D)0.01782 ± 0.008670.02056 ± 0.012450.02224 ± 0.013670.02315 ± 0.012100.02813 ± 0.01459
HGC (5D)0.02175 ± 0.017170.02372 ± 0.016520.02768 ± 0.014000.02531 ± 0.010540.02951 ± 0.01241

Q3: Comparison with language models.

R3: While LMs like BERT leverage rich lexical priors from concept names, in contrast to our pure structure, they operate in unconstrained Euclidean feature spaces that inherently lack mathematical guarantees for asymmetry and transitivity, making them prone to structural inconsistency and hallucinations across deep subtrees. To clarify where geometric methods stand, we will evaluate BERT' s baseline capability on hypernymy prediction. This paradigm directly aligns with recent advances in multi-modal hierarchical learning, such as [2], which proved that pairing foundation models with hyperbolic entailment cones provides the critical asymmetric inductive bias needed to resolve cross-granularity alignment without symmetric confusion. Consequently, geometric methods and LMs are fundamentally complementary rather than mutually exclusive: standalone HGC achieves strict topological fidelity at a tiny fraction of an LM's parameter footprint, while our gravitational formulation simultaneously serves as a plug-and-play geometric regularizer to enforce global hierarchical consistency in modern foundation models.

We added a baseline [3] using frozen BERT-base embeddings and an order-sensitive classifier trained on the same Mammal relation splits. Both 2D and 5D HGC achieved higher test F1 than BERT at all five training percentages. The largest gap occurred when training used only basic edges: BERT achieved 14.02% precision and 55.50% recall (22.39 F1), while 5D HGC achieved 66.7% precision and 74.6% recall (70.4 F1).

VariantMetricMammal
0%10%25%50%90%
Bertprecision14.0225.6044.6548.1175.00
HGC (d=2)precision48.062.267.777.469.6
HGC (d=5)precision66.771.584.482.389.2
Bertrecall55.5060.2845.9348.8053.11
HGC (d=2)recall62.785.282.383.786.6
HGC (d=5)recall74.689.093.397.690.4
Bertf122.3935.9445.2848.4562.18
HGC (d=2)f154.471.974.380.577.2
HGC (d=5)f170.479.388.689.389.8

[2] Avik Pal, Max van Spengler, Guido Maria D’Amely di Melendugno, Alessandro Flaborea, Fabio Galasso, and Pascal Mettes. Compositional entailment learning for hyperbolic vision-language models. In International Conference on Learning Representations, 2025.

[3] Michael Hanna and David Mareček. 2021. Analyzing BERT’s Knowledge of Hypernymy via Prompting. In Proceedings of the Fourth BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP, pages 275–282, Punta Cana, Dominican Republic. Association for Computational Linguistics.

I appreciate the authors’ response. I am inclined to maintain my original score.

Reviewer_X9dHOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes Hyperbolic Gravitational Cones (HGC), which extends Hyperbolic Entailment Cones (HEC) for modeling partial orders, by decoupling a concept’s breadth, or extension, from its embedding radius via a learnable “mass” function which depends on both the radius and the embedding direction. The paper also introduces a “gravitational energy” loss and a finite-difference regularizer meant to keep the mass monotone and convex in the radius. Experiments show mixed hierarchy reconstruction results on Mammal, WordNet Noun, and MCG datasets.

Critical strengths

The paper is well written, easy to follow, and asks an interesting question: should a concept’s breadth be a deterministic function of its radius alone?

Critical weaknesses
  • The physics motivation adds little: As I understand it, the method makes the cone aperture depend on radius and direction. None of this requires a physics inspiration. Plus, there is no real notion of mass, and gravity has no aperture. The “gravitational potential” section is simply a collection of heuristics (a softplus on the angle violation, multiplied by a distance, weighted by a stop-grad scalar) that make the method seem more principled than it is.

  • Theory (Proposition 1): The equivalence between nesting and the mass inequality holds only for concepts along the same ray. If we think of cones as concept extensions (in the Formal Concept Analysis sense), nesting should hold for any descendant, including those on different rays. The anisotropic mass is precisely what can break this since a descendant in a direction with larger α\alpha can have a cone that extends beyond its ancestor’s. If my reading is correct, then the the transitivity claim in the abstract is an overstatement.

  • Unmotivated implementation: The Mobius layer used to compute α(Ω)\alpha(\Omega) doesn’t make sense to me. The purpose of the network is to learn a scalar factor from an Euclidean direction. I’m unsure what the Mobius machinery adds here.

  • Experiments: Given all the aforementioned complexity, I expected experiments or analyses showing new findings or insights beyond what one can achieve with HEC. However, not only are the gains not consistent, with mixed results on Mammal and MCG, the experiments on the main paper don’t go beyond these standard benchmarks. In particular, the paper doesn’t show what the learned masses capture, even though the abstract claims they quantify concept breadth. Finally, the paper is missing comparisons to recent literature on non-hyperbolic partial order embeddings (Binder, Gyurek et al., KDD 2024 and Subspace Embeddings, Moreira et al., KDD 2026).

Questions for authors
  • Does nesting C(v)⊆C(u)\mathcal{C}(v) \subseteq \mathcal{C}(u) hold for descendants vv that are not on uu’s ray? If not, how is transitivity ensured in general? How does an isotropic variant (constant α\alpha) compare? Ablations of the energy loss versus the standard HEC loss would also help attribute the gains.
  • Do the learned masses correlate with subtree size / number of descendants / distance from root, better than the radius does?
  • Why use a Mobius layer to compute α\alpha from a direction?
3 comments

We sincerely thank the reviewer for the thorough, insightful, and constructive feedback. Below, we address each of your concerns in detail.

Q1: The physics motivation adds little.

R1: We appreciate the reviewer’ s perspective and would like to clarify that the concept of 'gravitation' serves strictly as an intuitive analogy. The true technical value of our work rests entirely on its geometric and optimization mechanics.

Q2: How is transitivity ensured in general?

R2: Proposition 1 establishes an exact containment condition along the geodesic ray. For off-ray pairs, the gravitational energy provides a complementary training signal. Even after a child enters its parent’s cone, the soft-gated angular term continues to exert a persistent restorative force, actively pulling the child node toward the parent's central cone axis.

Q3: How does an isotropic variant (constant) compare?

R3: Both ablations are systematically conducted in Appendix I (Figs. 9 & 10) and will be highlighted in the main text. On complex hierarchies (WordNet Noun and MCG), isotropic variants (both constant-mass and depth-only) severely plateau below 55% F1, whereas our full HGC achieves 95.3% on Noun and 91.6% on MCG (10D), outperforming isotropic models by an absolute margin of over 35% to 38% F1. Furthermore, replacing our continuous energy with the standard HEC hinge loss (HGC-Max-Margin) drops performance by over 37% F1 due to vanishing interior gradients. These results confirm that neither component alone suffices: anisotropic mass resolves geometric rigidity, while the continuous potential eliminates optimization stagnation.

Q4: Do the learned masses correlate with subtree size / number of descendants / distance from root, better than the radius does?

R4: To rigorously verify whether the learned mass m(u,Ωu) and aperture ψ(u) meaningfully capture intrinsic concept breadth rather than merely memorizing hierarchical depth, we conducted a comprehensive correlation analysis on the WordNet Mammal taxonomy. We evaluated the Spearman rank Correlation (𝜌) against true ground-truth subtree sizes ∣𝐷𝑒𝑠𝑐(𝑢)∣ across all 1,179 concepts across multiple dimensions (d=2,5) and supervision ratios (0% to 90%).

  1. Across all 10 experimental configurations, the learned cone half-aperture ψ exhibits a strong positive correlation (reaching up to +0.464), while the anisotropic mass consistently maintains a steady positive correlation (ρ≈+0.14∼+0.22). This directly verifies our core claim: broad super-categories (e.g., carnivore, ∣Desc∣=365) learn substantially wider gravitational apertures and larger masses to encapsulate their extensive subtrees, whereas fine-grained leaves contract their scopes for high precision.
  2. Hyperbolic radius consistently displays a pronounced negative correlation with subtree size (averaging −0.55 in d=5, reaching up to −0.602). This confirms that HGC naturally positions abstract ancestors near the origin while projecting specific descendants toward the manifold periphery without any explicit depth supervision.
  3. Remarkably, even in the pure zero-shot setting (0% additional transitive edges), the aperture correlation remains highly prominent (ρ=+0.414 at d=2, +0.464 at d=5). This demonstrates that HGC’ s continuous gravitational field establishes globally coherent hierarchical topologies via intrinsic geometric inductive bias rather than memorizing dense structural supervision.
MetricDim=2(0%)10%25%50%90%Dim=5(0%)10%25%50%90%
Hyperbolic radius-0.417-0.482-0.335-0.255-0.318-0.518-0.543-0.602-0.534-0.570
HGC mass0.2230.1880.1790.1810.1920.2000.1360.1690.2020.210
HGC Half-aperture0.4140.4410.3010.2510.2940.4640.3450.3200.4580.269

Q5: Why use a Mobius layer?

R5: Geometrically, the unit direction vector Ω𝑢=𝑢/∥𝑢∥ resides on 𝑆^𝑛−1, which corresponds to the conformal boundary of the Poincaré ball. Unlike standard Euclidean or tangent-space mappings that process directional features in unconstrained flat space, a Möbius layer acts as a canonical conformal transformation of the hyperbolic manifold. This formulation preserves the intrinsic angular relationships between semantic directions, preventing numerical instability and gradient explosion near the manifold boundary.

To empirically verify this architectural advantage, we compared our model against an explicit HGC-Tangent baseline (a standard 2-layer Tangent-Space MLP on the Mammal dataset across dimensions (d=2,5). HGC-Möbius consistently outperforms HGC-Tangent, achieving dramatic gains in constrained and data-scarce regimes, such as an absolute improvement of +21.8% F1 (74.3% vs. 52.5% at d=2, 25%) and +15.0% F1 (71.9% vs. 56.9% at d=2,10%), alongside steady advantages at d=5 (up to +8.8% F1 at 10%). These results demonstrate that conformal Möbius projection provides superior representation fidelity and geometric stability.

DatesetMammal (d=2)Mammal (d=5)
Percentage0%10%25%50%90%0%10%25%50%90%
HGC-Möbius54.471.974.380.577.270.479.588.689.389.8
HGC-Tangent49.856.952.576.378.472.270.785.586.885.0

Q6: The paper is missing comparisons to recent literature?

R6: We thank the reviewer for pointing out Binder [1]. We originally prioritized hyperbolic baselines because Binder operates under a distinct paradigm, discrete binary embeddings on a hypercube. Like Euclidean methods, inherently suffers from polynomial capacity limits and typically requires substantially larger dimensions (d≫10) to avoid crowding in deep hierarchies. Nonetheless, to ensure a comprehensive evaluation, we tested Binder under our identical low-dimensional settings: Binder suffers severe degradation due to capacity bottlenecks in ultra-low dimensions (failing to exceed 60% F1), whereas HGC significantly outperforms it across all three datasets (achieving over 90% F1). We will discuss Binder in Related Work and incorporate its full experimental comparison into Tables 1 & 2 in the revised manuscript.

Table 1: Performance (F1 score %) of the discrete order embedding baseline Binder on the Mammal dataset.

Percentage (Mammal)0%10%25%50%90%
Binder (d=2)19.5918.3422.7021.0724.68
Binder (d=5)33.0327.4932.2338.9248.22

Table 2: Performance (F1 score %) of the discrete order embedding baseline Binder on the Noun dataset.

Percentage (Noun)0%10%25%50%
Binder (d=5)23.2024.8523.2523.18
Binder (d=10)28.2125.8227.7227.48
Binder (d=30)58.5227.6038.0543.13

Table 3: Performance (F1 score %) of the discrete order embedding baseline Binder on the MCG dataset.

Percentage (MCG)0%10%25%50%
Binder (d=5)23.5824.1125.5326.01
Binder (d=10)24.8325.3027.1327.76
Binder (d=30)24.8125.1827.7028.92

[1] Croix Gyurek, Niloy Talukder, and Mohammad Al Hasan. 2024. Binder: Hierarchical Concept Representation through Order Embedding of Binary Vectors. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD '24).

Thank you for your reply. Some of my concerns were addressed. I follow the numeration in the rebuttal below:

Q1. I understand that, but the intuition claimed adds little to the optimization and mechanics, detracting from the actual contribution. As mentioned in my initial review, gravity has no aperture and the whole paper rests on apertures.

Q2. I wish this was more clearly demonstrated in the paper, with the actual scope of Proposition 1 delimited.

Q3. Your reply resolves my concern.

Q4. I don’t read the learned cone half-aperture Spearman’s correlation of 0.464 as “strong positive correlation”. Same goes for the mass, which has even lower values. Both are considerably weaker than the hyperbolic radius and, for example, subspace rank In Subspace Embeddings (Moreira et al., KDD 2026), mentioned in my initial review.

Q5. Your reply resolves my concern.

Q6. I understand that non-hyperbolic embeddings typically require larger dimensions. On the other hand they don’t need the optimization machinery needed by hyperbolic ones. The trade-off needs to be acknowledged and comparisons presented. Further, the authors didn’t compare with Subspace Embeddings, which is more recent than Binder.

Reviewer_vLxQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper extends hyperbolic entailment cones. It replaces the fixed aperture constant KK with a learned mass. It depends on the depth and the direction of the node. The paper replaces the hinge loss with a smooth energy that gives a non-zero gradient inside the cone. Transitivity is proven only along one ray from the origin. Experiments on WordNet Mammal, WordNet Noun, and the Microsoft Concept Graph show the best F1 in most settings.

Critical strengths
  • The idea is simple and clear. One constant KK becomes a learned function. A hinge loss becomes a smooth loss. Both changes are easy to understand and to implement.
  • Good motivations. Zero gradient inside the cone is a known issue. A fixed aperture per depth is a real limit.
Critical weaknesses

1.The transitivity claim is weaker than the abstract says. The abstract says transitivity is "mathematically guaranteed". Proposition 1 holds only when the two nodes are on the same ray from the origin. In a learned embedding, a child is almost never exactly on the ray of its parent. So the proposition applies to almost no real parent-child pair. By continuity, pairs close to one ray nest approximately, but the paper does not quantify this. The author may want to compute the error and give a bound.

  1. The regularizer looks redundant. The mass is m(r,Ω)=m0(1+α(Ω) e−βr)m(r,\Omega) = m_0 \big(1 + \alpha(\Omega)\, e^{-\beta r}\big). The factor α\alpha depends on the direction only. For a fixed direction,
\frac{\partial^2 m}{\partial r^2} = \beta^2\, m_0\, \alpha(\Omega)\, e^{-\beta r} > 0$$ always. So the monotonicity and convexity terms in Eq. 16 are always zero, up to finite-difference error. Section 3.5 says the network can "overpower the radial decay". This seems not possible with the given formula. The authors should explain this.
Questions for authors

1.With α\alpha independent of rr, how can the regularizer in Eq. 16 ever be non-zero?

2.How often is transitivity violated across directions in the learned embeddings?

2 comments

We sincerely thank the reviewer for the thorough, insightful, and constructive feedback. Below, we address each of your concerns in detail.

Q1: How can the regularizer in Eq. 16 ever be non-zero?

R1: We sincerely appreciate the reviewer’ s acute mathematical insight. We will completely remove 𝐿_reg (Eq. 14–16) and the justification in Sec. 3.5.

Q2: How often is transitivity violated?

R2: Proposition 1 establishes an exact containment condition along the geodesic ray, but does not guarantee global geometric nesting for arbitrary off-ray pairs. For off-ray pairs, our gravitational energy provides a complementary training signal. After a child enters its parent’s cone, the soft-gated angular term continues to exert a persistent restorative force, actively pulling the child node toward the parent's central cone axis. This encourages alignment across directions, although it does not guarantee nesting for every off-ray pair. For a descendant v situated at an angular deviation Ξ(u,v) from the axis of u, the necessary and sufficient condition for full cone containment C(v)⊆C(u) is governed by the triangle inequality: Ξ(u,v) + ψ(v) ≤ ψ(u). Unlike standard hinge loss (which stop gradients as soon as Ξ(u,v) ≤ ψ(u) ), our Gravitational Energy actively optimizes this inequality. HGC’s soft-gated potential continues to pull v toward the central axis, minimizing the angular deviation Ξ(u,v) and maximizing the residual angular margin Δ = ψ(u) − (Ξ(u,v) + ψ(v)).

We agree that calling this a general "mathematical guarantee" in the abstract was imprecise. We will explicitly qualify this in the revised abstract as "strict ray-wise transitivity and empirical cross-directional containment."

Experimentally, we evaluated all 11,078 ground-truth chains 𝑢→𝑣→𝑤 in the full WordNet Mammal under ten training settings. For each setting, we counted the chains 𝑣∈𝐶(𝑢) and 𝑤∈𝐶(𝑣). A violation occurred when 𝑤∉𝐶(𝑢). Each cell reports violations / chains satisfying both premises, followed by the violation rate. Across the ten settings, the empirical violation rate ranged from 0.00% to 1.32% and was below 1% in nine settings. At d=5d=5, we observed no violations among 7,901 eligible chains at 25% and 5,310 at 90%.

Dimension0%10%25%50%90%
Mammal (d=2)67/5,094 (1.32%)45/7,013 (0.64%)32/6,148 (0.52%)1/6,557 (0.02%)8/6,288 (0.13%)
Mammal (d=5)30/5,742 (0.52%)13/7,271 (0.18%)0/7,901 (0.00%)12/6,978 (0.17%)0/5,310 (0.00%)

Thank the author for the reply. That partly addressed my concerns.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.