acceptodds
Under review as a conference paper at ICLR 2027

The Broken Geometry of Hyperbolic Vision-Language Models

Abstract

Hyperbolic vision-language models (VLMs) leverage hyperbolic geometry to model hierarchical structure in multimodal representations. Recent works contrastively align images and text in a shared hyperbolic space while employing entailment cones to impose hierarchical orderings among whole and part representations. However, we find that existing hyperbolic VLMs operate in a near-Euclidean regime, with embeddings concentrated near the origin on manifolds with small curvature. Thus, most embeddings fall outside the theoretically valid domain of entailment cones, suggesting that the intended geometry and entailment mechanism are not fully exploited in practice. Motivated by our observation, we propose a simple Euclidean-based hierarchy-aware VLM that retains the same compositional supervision while eliminating hyperbolic operations and entailment cones. Despite its simpler geometry, our model consistently outperforms existing hyperbolic VLMs across diverse downstream tasks. The empirical results suggest that the gains reported by prior studies may stem primarily from additional compositional data and hierarchical supervision rather than from hyperbolic geometry itself, highlighting the need to carefully examine the contribution of hyperbolic geometry in hierarchical representation learning.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Vision-Language Models, Hyperbolic Geometry, Representation Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_gS3WOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates whether existing hyperbolic vision-language models actually make meaningful use of hyperbolic geometry. The authors analyze several representative hyperbolic VLMs and find that their learned representations remain close to the Euclidean regime, while many entailment cones operate in a saturated regime. Motivated by these observations, they propose EuHier, a Euclidean hierarchy-aware VLM that retains whole-part compositional supervision without hyperbolic operations or entailment cones. Experiments on zero-shot classification, retrieval, hierarchical evaluation, and scene understanding show that EuHier performs competitively or better than the evaluated hyperbolic baselines, suggesting that part of the gains in prior work may come from additional whole-part/compositional supervision rather than hyperbolic geometry itself.

Critical strengths

S1. Comprehensive geometric analysis and empirical evaluation. The paper examines hyperbolic VLMs beyond downstream performance, including representation geometry, distance distortion, nearest-neighbor behavior, and entailment-cone saturation. The experiments also cover multiple models and downstream tasks, providing useful empirical observations.

Critical weaknesses

W1. Limited novelty and insufficient differentiation from prior work. A key part of the geometric diagnosis has already been studied in existing work. In particular, HyperConeNet (Wang et al., IEEE TCSVT 2026) has identified and analyzed the invalid cone-aperture domain and the saturation of half-aperture angles at pi/2 in MERU. These findings overlap directly with an important part of the present paper's motivation and analysis. Although the authors extend the investigation to additional hyperbolic VLMs and introduce EuHier, the paper does not sufficiently discuss this prior work or clearly establish the novelty of its additional contributions.

Relevant reference: Wang et al., Enhancing Hyperbolic Vision–Language Representations via Adaptive Apertures and Contrastive Entailment, IEEE TCSVT, 2026. https://doi.org/10.1109/TCSVT.2026.3701206

W2. Insufficient clarity and consistency in baseline reproduction. The paper's main geometric findings rely on retrained versions of existing hyperbolic VLMs, but the reproduction details and deviations from previously published results are not sufficiently explained. For example, using ViT-B/16, HyCoCLIP achieves 43.6 on ImageNet compared with 45.8 in its original results, with similarly noticeable drops on CIFAR-100 (55.9 vs. 60.1) and SUN397 (52.9 vs. 57.2). In contrast, the reproduced UNCHA results are generally much closer to its originally reported performance. I understand that results may vary across implementations and training settings, but such differences deserve a clearer explanation, especially since the paper draws its main geometric conclusions from these retrained models.

Questions for authors

Q1. Could the authors clarify how their geometric analysis differs from HyperConeNet (Wang et al., IEEE TCSVT 2026), particularly regarding the diagnosis of cone-aperture saturation? Additionally, I noticed a closely related preprint, Is the Geometry Doing the Work? (arXiv:2607.05268), which investigates several similar geometric phenomena in hyperbolic VLMs. It would be helpful if the authors could discuss the relationship between these findings and clarify the distinct contributions of the present work.

Relevant reference:

Q2. Could the authors clarify the baseline reproduction procedure, including whether official implementations were used and what modifications, if any, were made? How do the authors explain the differences between their reproduced HyCoCLIP results and the officially reported numbers, particularly on ImageNet, CIFAR-100, and SUN397? Have the main geometric findings also been verified using publicly released checkpoints?

Q3. What geometry does EuHier-H actually learn? Could the authors report the same geometric diagnostics used for the existing hyperbolic VLMs, such as curvature, embedding radii, and distance distortion? This would help clarify whether EuHier-H operates in a meaningfully hyperbolic regime and whether the contribution of geometry has been adequately isolated.

8 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/4)

We thank the reviewer for their detailed assessment. We address the concerns below and provide additional analyses to clarify our findings.

 

A1. Distinct contributions and comparison with related work

We thank the reviewer for pointing out these closely related studies. We agree that their relationship to our work should be clearer in the main text. HyperConeNet [1] previously identified the invalid aperture domain and cone saturation in MERU. Our study extends beyond this diagnosis through the following analyses and empirical comparisons.

  • Quantifying the effect of curvature on learned representations. HyperConeNet examines cone apertures, negative containment, and radial organization. We extend the geometric diagnostics with a systematic, multi-level evaluation to investigate whether comparisons between learned embeddings materially depend on negative curvature, connecting local curvature exposure and cone validity to pairwise distance distortion and nearest neighbor decisions. With the representations fixed, we compare hyperbolic distances with Euclidean distances in the tangent space at the origin. The small distance changes and nearly unchanged nearest neighbor decisions in Table 1 demonstrate the limited influence of the hyperbolic metric on embedding comparisons in the evaluated models. This connects geometric diagnostics to actual similarity decisions, providing quantitative evidence of the practical effect of negative curvature beyond cone validity alone. Also, our analysis covers more advanced models, including HyCoCLIP, PHyCLIP and UNCHA, and finds the same behavior even in UNCHA despite its more extensive compositional supervision.

  • Separating the mechanisms of radial contraction. HyperConeNet describes the original entailment loss as a hard-threshold mechanism that lacks constraints on negative samples. We further show that even an active positive entailment penalty can decrease through parent contraction within the valid cone domain, before aperture clipping occurs. Contraction both widens the aperture and decreases the exterior angle while the child embedding, curvature, and radial directions remain fixed. We derive these gradient contributions and measure them during early training. In HyCoCLIP, the exterior angle contribution is larger on average over all active pairs, whereas the aperture contribution is larger near the valid cone boundary (Section 3.4 and Appendix D.2). Thus, our analysis identifies how the active objective can reduce its penalty through radial contraction, beyond the absence of penalties for negatives inside the cones.

  • Examining the role of compositional supervision. HyperConeNet evaluates improvements to a hyperbolic formulation. Beyond improving the hyperbolic objective, we ask whether strong performance from whole–part supervision requires hyperbolic operations. EuHier serves as a simple Euclidean reference for this question, using compositional contrastive supervision without entailment cones.

AuthorsOpenReview ↗

Official Comment by Authors (2/4)

(A1 continued) We also acknowledge the closely related findings of contemporaneous work [2] on near-Euclidean representations and cone saturation. Our Euclidean training comparisons provide complementary evidence by evaluating the effects of part supervision and compositional objectives under matched settings. We additionally quantify pairwise distance changes and nearest-neighbor agreement using the same frozen representations. While it analyzes curvature gradients, our radial gradient analysis holds curvature fixed and separates the contributions of aperture widening and exterior angle reduction to the entailment loss.

Importantly, while recent works [1,3] address entailment cone collapse and evaluate their proposed loss components, an important question remains whether their performance gains require negative curvature. To examine whether competitive performance can be achieved without hyperbolic operations, we reproduced both methods and trained EuHier with one part per sample (Table A). ARGENT [3] performs comparably overall to CLIP trained with the same whole and part image–text pairs, while EuHier w/ 1-part outperforms both ARGENT and HyperConeNet across the reported metrics. EuHier also improves over CLIP with one part, indicating that how whole–part information is used matters beyond the number of part pairs. These results extend our empirical comparison beyond models using the original entailment cone objective and demonstrate that strong performance on the evaluated tasks can be achieved through compositional supervision without hyperbolic operations. Together with our geometric analysis, these findings highlight an important distinction between achieving strong downstream performance and demonstrating that the learned representations meaningfully exploit negative curvature. We will clarify these distinctions in the main text.

[Table. A] Additional comparisons with HyperConeNet and ARGENT. † denotes our reproduced results for the additional baselines.

\small \begin{array}{l|cc|cc|cc|ccccc} \hline & & \llap{\text{Classification}} & & \llap{\text{Text retrieval}} & & \llap{\text{Image retrieval}} & & & & \llap{\text{Hierarchical metrics}} & \\\\ \text{Model} & \text{ImageNet} & \text{Avg} & \text{COCO (R@5/10)} & \text{Flickr (R@5/10)} & \text{COCO (R@5/10)} & \text{Flickr (R@5/10)} & \text{TIE}\downarrow & \text{LCA}\downarrow & J\uparrow & P_H\uparrow & R_H\uparrow \\\\ \hline \text{CLIP} & 37.3 & 37.1 & 69.2/79.1 & 91.8/96.1 & 55.6/67.3 & 82.5/89.3 & 3.823 & 2.298 & 0.774 & 0.845 & 0.846 \\\\ \text{CLIP w/ 1-part} & 43.8 & 43.0 & 70.6/80.4 & 91.8/95.8 & 57.1/68.3 & 83.4/89.8 & 3.317 & 2.096 & 0.804 & 0.867 & 0.866 \\\\ \text{HyperConeNet}^{\dagger} & 37.0 & 37.2 & 67.2/78.3 & 90.4/95.1 & 55.3/67.0 & 82.0/88.2 & 3.850 & 2.290 & 0.772 & 0.845 & 0.844 \\\\ \text{ARGENT}^{\dagger} & 43.9 & 42.8 & 70.9/81.0 & 92.9/96.3 & 58.0/69.3 & 84.4/90.7 & 3.360 & 2.100 & 0.802 & 0.864 & 0.865 \\\\ \text{EuHier w/ 1-part} & \underline{48.3} & \underline{45.3} & \underline{72.3}/\underline{82.1} & \underline{93.8}/\underline{96.8} & \underline{59.3}/\underline{70.6} & \underline{85.8}/\underline{91.9} & \underline{2.980} & \underline{1.980} & \underline{0.826} & \underline{0.881} & \underline{0.882} \\\\ \text{EuHier w/ 5-part} & \mathbf{49.5} & \mathbf{45.9} & \mathbf{75.0}/\mathbf{83.8} & \mathbf{94.7}/\mathbf{97.6} & \mathbf{62.4}/\mathbf{72.9} & \mathbf{88.1}/\mathbf{93.3} & \mathbf{2.884} & \mathbf{1.950} & \mathbf{0.833} & \mathbf{0.885} & \mathbf{0.887} \\\\ \hline \end{array} }$$ - [1] Wang et al., Enhancing Hyperbolic Vision–Language Representations via Adaptive Apertures and Contrastive Entailment, IEEE TCSVT, 2026. - [2] Kim et al., Is the Geometry Doing the Work? An Operating-Point Audit of Hierarchy in Hyperbolic Vision-Language Models, arXiv:2607.05268, 2026. - [3] Huynh, Chuong, et al., Argent: Adaptive hierarchical image-text representations, ECCV, 2026.
AuthorsOpenReview ↗

Official Comment by Authors (3/4)

A2. Baseline reproduction and validation with official checkpoints

We appreciate the reviewer's emphasis on reproducibility. We used the official implementations released by the respective authors for all baselines. When reconstructing GRIT, we encountered expired image URLs, an issue also reported in PHyCLIP, and recovered approximately 13M of the original 20.5M image–text pairs. The resulting differences in training data size and composition may partly explain the performance discrepancies observed on certain datasets. Nevertheless, comparing models under the same training conditions is essential for consistently evaluating their relative performance and learned geometry. We therefore retrained all baselines using the same reconstructed dataset, backbone, and training budget, reporting the mean and standard deviation over three independent runs.

Importantly, we also verified our main geometric findings using the publicly released MERU, HyCoCLIP, and UNCHA checkpoints with both ViT-B and ViT-S backbones. As shown in the Table.B-C, the reported G_pos and G_neg values do not exceed 0.65%, NNA@1 ranges from 99.14% to 99.98%, and saturation rates for the evaluated text-parent cones range from 95.33% to 100%. These results confirm that the near-Euclidean behavior and cone saturation are not confined to models retrained under our experimental setup.

[Table. B] Point geometric analysis of the released checkpoints.

\small \begin{array}{l|c|l|rr} \hline \text{Model} & \text{Backbone} & \text{Modality} & r & T \\\\ \hline \text{MERU} & \text{ViT-B} & \text{text} & 0.198 & 1.0065 \\\\ & & \text{img} & 0.299 & 1.0150 \\\\ \hdashline \text{MERU} & \text{ViT-S} & \text{text} & 0.198 & 1.0066 \\\\ & & \text{img} & 0.294 & 1.0145 \\\\ \hline \text{HyCoCLIP} & \text{ViT-B} & \text{part text} & 0.130 & 1.0028 \\\\ & & \text{text} & 0.141 & 1.0033 \\\\ & & \text{part img} & 0.203 & 1.0068 \\\\ & & \text{img} & 0.204 & 1.0069 \\\\ \hdashline \text{HyCoCLIP} & \text{ViT-S} & \text{part text} & 0.133 & 1.0029 \\\\ & & \text{text} & 0.145 & 1.0035 \\\\ & & \text{part img} & 0.202 & 1.0068 \\\\ & & \text{img} & 0.203 & 1.0069 \\\\ \hline \text{UNCHA} & \text{ViT-B} & \text{part text} & 0.200 & 1.0066 \\\\ & & \text{text} & 0.199 & 1.0066 \\\\ & & \text{part img} & 0.409 & 1.0281 \\\\ & & \text{img} & 0.409 & 1.0281 \\\\ \hdashline \text{UNCHA} & \text{ViT-S} & \text{part text} & 0.200 & 1.0067 \\\\ & & \text{text} & 0.200 & 1.0067 \\\\ & & \text{part img} & 0.394 & 1.0261 \\\\ & & \text{img} & 0.394 & 1.0261 \\\\ \hline \end{array} }$$ **[Table. C] Pairwise geometric and cone validity analysis of the released checkpoints.** $${ \small \begin{array}{l|c|l|rrrr} \hline \text{Model} & \text{Backbone} & \text{Modality} & G_{\mathrm{pos}} & G_{\mathrm{neg}} & \text{NNA@1} & S_{\mathrm{final}} \\\\ \hline \text{MERU} & \text{ViT-B} & \text{text}\to\text{img} & 0.64\\% & 0.65\\% & 99.92\\% & 99.58\\% \\\\ \hdashline \text{MERU} & \text{ViT-S} & \text{text}\to\text{img} & 0.65\\% & 0.65\\% & 99.94\\% & 99.41\\% \\\\ \hline \text{HyCoCLIP} & \text{ViT-B} & \text{part text}\to\text{part img} & 0.27\\% & 0.27\\% & 99.88\\% & 100.0\\% \\\\ & & \text{part text}\to\text{img} & 0.27\\% & 0.27\\% & 99.92\\% & 100.0\\% \\\\ & & \text{text}\to\text{part img} & 0.33\\% & 0.33\\% & 99.94\\% & 100.0\\% \\\\ & & \text{text}\to\text{img} & 0.33\\% & 0.33\\% & 99.98\\% & 100.0\\% \\\\ \hdashline \text{HyCoCLIP} & \text{ViT-S} & \text{part text}\to\text{part img} & 0.29\\% & 0.28\\% & 99.84\\% & 100.0\\% \\\\ & & \text{part text}\to\text{img} & 0.28\\% & 0.29\\% & 99.91\\% & 100.0\\% \\\\ & & \text{text}\to\text{part img} & 0.35\\% & 0.35\\% & 99.95\\% & 100.0\\% \\\\ & & \text{text}\to\text{img} & 0.34\\% & 0.35\\% & 99.97\\% & 100.0\\% \\\\ \hline \text{UNCHA} & \text{ViT-B} & \text{part text}\to\text{part img} & 0.63\\% & 0.60\\% & 99.18\\% & 98.31\\% \\\\ & & \text{part text}\to\text{img} & 0.63\\% & 0.60\\% & 99.14\\% & 98.31\\% \\\\ & & \text{text}\to\text{part img} & 0.63\\% & 0.61\\% & 99.66\\% & 97.34\\% \\\\ & & \text{text}\to\text{img} & 0.62\\% & 0.61\\% & 99.67\\% & 97.34\\% \\\\ \hdashline \text{UNCHA} & \text{ViT-S} & \text{part text}\to\text{part img} & 0.63\\% & 0.60\\% & 99.29\\% & 97.94\\% \\\\ & & \text{part text}\to\text{img} & 0.63\\% & 0.61\\% & 99.36\\% & 97.94\\% \\\\ & & \text{text}\to\text{part img} & 0.64\\% & 0.62\\% & 99.56\\% & 95.33\\% \\\\ & & \text{text}\to\text{img} & 0.63\\% & 0.62\\% & 99.71\\% & 95.33\\% \\\\ \hline \end{array} }$$
Reviewer_asLROpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies existing hyperbolic vision-language models (VLMs) and argues that their performance gains can be mostly attributed to the hierarchy-aware supervision rather then exploitation of the hyperbolic space itself. Experiments show that the embeddings produced by such models concentrate near the origin where hyperbolic distances are close to Euclidean distances in the tangent space, suggesting that these models make little use of the curvature of the hyperbolic space. Motivated by these observations, the authors propose EuHier - a simple Euclidean-based VLM trained with hierarchy-aware compositional supervision. EuHier achieves better zero-shot classification accuracy, retrieval, hierarchical classification and scene understanding results than the evaluated hyperbolic VLMs.

Critical strengths
  1. While existing literature has acknowledged issues with hyperbolic training and tried to improve it, this paper questions whether hyperbolic space is needed at all to achieve the same performance gains. The proposed EuHier method provides evidence that hierarchy-aware supervision can be as effective itself without the hyperbolic geometry. This finding could encourage the research community to analyze results more rigorously and isolate the contributions of the proposed representation space.
  2. The paper provides solid geometric analysis of the existing hyperbolic methods. All the evaluated embeddings showed nearly Euclidean behavior: hyperbolic and tangent space distances were similar and almost always selected the same nearest neighbor (Table 1). This suggests that curvature has little effect on how these models compare embeddings. The analysis also showed that the most measured apertures of entailment cones reach the clipping limit and stop adapting to the parent embeddings radii, limiting the ability to represent hierarchies.
Critical weaknesses

The main weakness is that it's hard to properly evaluate the novelty of the contribution from the way the paper positions itself. The closest prior literature ([1], [2]) is only referenced in the appendix, although [2] studies conflicts between distance-based alignment and hierarchy and [1] analyzed the entailment-cone collapse. The paper would benefit from clearer positioning in the main text, explaining which findings are new and how its analysis extends these earlier studies. The novelty of the proposed EuHier is limited as it mainly combines standard contrastive learning with hierarchy-aware compositional supervision. This places more weight on the geometric analysis and empirical findings as the main contribution of the paper, making it important to clearly establish what they add beyond existing work.

[1] Huynh, Chuong, et al. "Argent: Adaptive hierarchical image-text representations." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2026.

[2] Ramasinghe, Sameera, et al. "Accept the modality gap: An exploration in the hyperbolic space." 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024.

Questions for authors

Could the authors expand the comparison with the prior works, particularly clarify what their analysis adds beyond [1] and make this distinction more explicit in the main text?

[1] Huynh, Chuong, et al. "Argent: Adaptive hierarchical image-text representations." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2026.

4 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/2)

We thank the reviewer for the positive assessment and constructive feedback. We agree that the distinctions from prior work should be made clearer in the main text. ARGENT identifies cone collapse associated with parent contraction, while ATMG analyzes the conflict between distance-based alignment and hierarchy and reports MERU's tendency toward low curvature. Our study builds on these findings to investigate how negative curvature affects learned representations and whether compositional supervision remains effective without hyperbolic operations as follows:

  • Quantifying the effect of curvature on learned representations. Prior observations of low curvature and cone collapse reveal geometric limitations, but do not establish how strongly negative curvature affects the learned representations or their embedding comparisons. We develop the evaluation protocol to systematically quantify these geometric effects, from local curvature exposure and cone validity to pairwise distance distortion and nearest neighbor decisions. In particular, with learned representations fixed, we compare hyperbolic and Euclidean distances, measuring how much negative curvature changes pairwise distances and whether these changes alter nearest-neighbor selections. These results provide direct evidence that the learned representations make limited use of the hyperbolic metric in embedding comparisons, beyond merely exhibiting low curvature or invalid cones. Our analysis also covers more advanced models, including PHyCLIP and UNCHA, and finds the same behavior even in UNCHA despite its more extensive compositional supervision.

  • Separating the mechanisms of radial contraction. ARGENT identifies parent contraction as a mechanism that widens entailment cones. We further show that contraction can reduce the active entailment loss through both aperture widening and a decrease in the exterior angle. This occurs even when the child embedding, curvature, and parent direction are held fixed. We derive these two radial gradient contributions and measure them during early training. In HyCoCLIP, the exterior-angle contribution is larger when averaged over all active pairs, whereas the aperture contribution is larger near the valid cone boundary (Section 3.4 and Appendix D.2). Thus, cone widening alone does not fully explain how parent contraction reduces the loss.

  • Examining the role of compositional supervision. Beyond improving the hyperbolic objective, we ask whether strong performance from whole–part supervision requires hyperbolic operations. EuHier serves as a simple Euclidean reference for this question, using compositional contrastive supervision without entailment cones. This allows us to empirically examine whether strong downstream performance can be achieved without hyperbolic modeling.

AuthorsOpenReview ↗

Official Comment by Authors (2/2)

Notably, although ARGENT addresses entailment cone collapse through an adaptive entailment objective, it does not examine whether the reported performance gains require hyperbolic geometry. To further investigate this question, we additionally reproduced ARGENT and trained EuHier with one part per sample (See Table.A). ARGENT performs comparably to CLIP trained with the same whole and part image–text pairs. Remarkably, adding part supervision to CLIP improves average classification accuracy from 37.1 to 43.0, while ARGENT achieves 42.8. EuHier w/ 1-part further improves accuracy to 45.3 and outperforms ARGENT across all reported metrics, extending our empirical comparison beyond models using the original entailment cone objective. These results further support our conclusion that strong performance can be achieved through compositional supervision without hyperbolic operations in the evaluated setting. Together with our geometric analysis, these findings highlight the distinction between strong downstream performance and evidence that learned representations meaningfully exploit negative curvature. We will clarify these distinctions in the main text.

 

[Table. A] Comparison of ARGENT and EuHier w/ 1-part. † denotes our reproduced results for the additional baselines.

\small \begin{array}{l|cc|cc|cc|ccccc} \hline & & \llap{\text{Classification}} & & \llap{\text{Text retrieval}} & & \llap{\text{Image retrieval}} & & & & \llap{\text{Hierarchical metrics}} & \\\\ \text{Model} & \text{ImageNet} & \text{Avg} & \text{COCO (R@5/10)} & \text{Flickr (R@5/10)} & \text{COCO (R@5/10)} & \text{Flickr (R@5/10)} & \text{TIE}\downarrow & \text{LCA}\downarrow & J\uparrow & P_H\uparrow & R_H\uparrow \\\\ \hline \text{CLIP} & 37.3 & 37.1 & 69.2/79.1 & 91.8/96.1 & 55.6/67.3 & 82.5/89.3 & 3.823 & 2.298 & 0.774 & 0.845 & 0.846 \\\\ \text{CLIP w/ 1-part} & 43.8 & 43.0 & 70.6/80.4 & 91.8/95.8 & 57.1/68.3 & 83.4/89.8 & 3.317 & 2.096 & 0.804 & 0.867 & 0.866 \\\\ \text{ARGENT}^{\dagger} & 43.9 & 42.8 & 70.9/81.0 & 92.9/96.3 & 58.0/69.3 & 84.4/90.7 & 3.360 & 2.100 & 0.802 & 0.864 & 0.865 \\\\ \text{EuHier w/ 1-part} & \underline{48.3} & \underline{45.3} & \underline{72.3}/\underline{82.1} & \underline{93.8}/\underline{96.8} & \underline{59.3}/\underline{70.6} & \underline{85.8}/\underline{91.9} & \underline{2.980} & \underline{1.980} & \underline{0.826} & \underline{0.881} & \underline{0.882} \\\\ \text{EuHier w/ 5-part} & \mathbf{49.5} & \mathbf{45.9} & \mathbf{75.0}/\mathbf{83.8} & \mathbf{94.7}/\mathbf{97.6} & \mathbf{62.4}/\mathbf{72.9} & \mathbf{88.1}/\mathbf{93.3} & \mathbf{2.884} & \mathbf{1.950} & \mathbf{0.833} & \mathbf{0.885} & \mathbf{0.887} \\\\ \hline \end{array} }$$

Thank you for providing additional results and elaborating on the novelty of the work. That clears my concerns. I'll increase my score.

Thank you very much for reviewing our response. We are glad that your concerns have been addressed, and we sincerely appreciate your feedback.

Reviewer_KkM3OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper examines whether several hyperbolic vision-language models actually use their intended geometry. Under a matched training setup, it finds small curvature-induced distance distortions and substantial saturation in the evaluated text-parent entailment cones. It then introduces EuHier, which uses whole–part contrastive supervision without hyperbolic operations or entailment cones. The experiments suggest that much of the benefit in these settings can be obtained through compositional supervision and a simpler Euclidean model.

Critical strengths
  1. Findings are supported by lots of measurements.

  2. The simpler baseline is supported by useful controls and downstream results

Critical weaknesses

The EuHier-H comparison changes more than curvature. Appendix G removes EuHier's L2 normalization, adds learnable modality scales, and replaces cosine similarity with negative geodesic distance. Consequently, Table 5 does not isolate which of these changes causes the performance difference. Its description that other settings are unchanged is too strong. This limits the causal interpretation of the training comparison, although it does not undermine the separate audit of frozen representations.

Questions for authors
  1. How does EuHier-H compare with a Euclidean-distance version using the same unnormalized features, modality scales, temperature treatment, and supervision? This would help separate curvature effects from the other differences in Appendix G.

  2. For the valid-projection experiment in Figure 3, are the baselines retrained with projection throughout, and is the projected parent used in both the aperture and exterior-angle calculations? Section 5.3 and Appendix G do not make this sufficiently explicit. Since the critique of clipping hinges on this distinction, the implementation detail matters.

1 comment

We thank the reviewer for their thoughtful assessment and constructive feedback. We address the concerns and questions below.
 

A1. Scope of the EuHier-H comparison and additional control

We agree that the EuHier vs. EuHier-H comparison does not isolate the effect of curvature alone. Our original comparison was intended to assess whether the hyperbolic formulation provides additional benefits over the Euclidean formulation under the same compositional supervision. To this end, EuHier-H adopts conventional hyperbolic modeling choices while preserving EuHier's encoder architectures, contrastive relations, and training protocol.

To quantitatively address the reviewer's concern, we conducted an additional experiment with a matched Euclidean distance variant (EuHier-E) that retains the same unnormalized features, learnable modality scales, temperature treatment, and compositional supervision as EuHier-H (See Table.A). EuHier-E performs better on fine-grained and miscellaneous classification and text retrieval, whereas EuHier-H performs better on general classification and hierarchical metrics. Neither formulation consistently outperforms the other. Notably, the original EuHier, which uses normalized features and cosine similarity, achieves the best performance across all reported summary metrics. We therefore do not attribute EuHier’s advantage over EuHier-H solely to the absence of curvature. Taken together, these results still support our conclusion that whole–part supervision can substantially improve performance without requiring hyperbolic operations in the evaluated setting.

We will clarify this distinction with the quantitative result in the manuscript and the caption of Table 5.

[Table. A] Ablation study.

\small \begin{array}{l|ccc|cc|ccccc} \hline & & & \llap{\text{Classification}} & & \llap{\text{Retrieval}} & & & & \llap{\text{Hierarchical metrics}} & \\\\ \text{Model} & \text{General} & \text{Fine} & \text{Misc.} & \text{Text} & \text{Image} & \text{TIE}\downarrow & \text{LCA}\downarrow & J\uparrow & P_H\uparrow & R_H\uparrow \\\\ \hline \textbf{EuHier} & \mathbf{73.2} & \mathbf{29.7} & \mathbf{29.4} & \mathbf{84.9} & \mathbf{75.3} & \mathbf{2.884} & \mathbf{1.950} & \mathbf{0.833} & \mathbf{0.885} & \mathbf{0.887} \\\\ \text{EuHier-H} & 70.9 & 27.3 & 25.5 & 84.3 & 75.0 & 3.230 & 2.030 & 0.810 & 0.874 & 0.867 \\\\ \text{EuHier-E} & 69.7 & 28.1 & 28.6 & 84.8 & 75.0 & 3.320 & 2.060 & 0.802 & 0.868 & 0.862 \\\\ \hline \end{array} }$$   ### **A2. Implementation details of valid projection** Yes. We retrained all valid projection variants from scratch with projection applied throughout training in the entailment calculations. Both the cone half-aperture and the exterior angle were computed using the same projected parent embedding. In the prior works, only the argument of the aperture function is clipped, leaving the cone apex at the original parent embedding. In contrast, valid projection moves the parent embedding into the admissible domain and uses the projected parent for both the aperture and exterior angle calculations. We will explicitly clarify these implementation details in Section 5.3 and Appendix G.

Comments on the paper

1 comment

Dear Reviewers and Area Chairs,

We sincerely appreciate the time and effort you devoted to reviewing our manuscript. In response to your constructive feedback, we have revised the manuscript to include:

  • Clarification of Novelty (asLR, gS3W): Related work and distinct contributions (Sec. 2.1).
  • Reproducibility Clarifications
    • Baseline reproduction (gS3W): Sec. 3.2.
    • Valid projection protocol (KkM3): Sec. 5.3.
  • Additional Comparisons
    • Euclidean and hyperbolic formulations (KkM3): Sec. 5.3, Tab. 5, Appendix G.
    • Additional baselines (asLR, gS3W): Appendix F.1, Tab. 14.
  • Additional Geometric Analyses
    • Validation using released checkpoints (gS3W): Appendix D.1, Tab. 6.
    • Learned geometry of EuHier-H (gS3W): Appendix F.3, Tab. 16.

These revisions are highlighted in blue for ease of reference. We welcome any further questions or comments.

Sincerely,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.