pith. sign in

arxiv: 2607.00428 · v1 · pith:2FV6ZZ4Xnew · submitted 2026-07-01 · 💻 cs.CV

HyFL-CLIP: Hyperbolic Fine-Tuning of CLIP for Robust Long-Context Understanding

Ji Ha Jang , Hayeon Kim , Chulwon Lee , Junghun James Kim , Se Young Chun This is my paper

Pith reviewed 2026-07-02 15:13 UTC · model grok-4.3

classification 💻 cs.CV
keywords hyperbolic fine-tuningCLIPlong-context understandingcross-modal retrievaltextual perturbationshierarchical semanticsEinstein midpoint aggregationhyperbolic entailment
0
0 comments X

The pith

HyFL-CLIP distills CLIP's image-text alignment into hyperbolic space to model hierarchical semantics for robust long-context understanding.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes a hyperbolic fine-tuning method for CLIP to address its struggles with long-context text descriptions that involve reordering, summarization, or omission. It argues that the standard Euclidean contrastive objective cannot capture hierarchical part-whole relationships, so it distills the alignment into hyperbolic geometry using cross-manifold similarity. This allows modeling entailment relations between global context, short components, and images via Einstein midpoint aggregation. If effective, the approach yields more robust performance on perturbed long-text retrieval tasks without losing short-text capabilities. A sympathetic reader would care because many real-world descriptions exceed the short-caption training regime of original CLIP.

Core claim

HyFL-CLIP models hierarchical semantics by linking summarized token-wise features, long-context descriptions, constituent short textual components, and images, capturing part-whole relationships via hyperbolic entailment with Einstein midpoint aggregation, and achieves up to 19.5% improvement in long-text cross-modal retrieval under textual perturbations over the best prior method.

What carries the argument

Hyperbolic entailment with Einstein midpoint aggregation, which links global long-context features to their constituent elements in a hierarchical manner.

If this is right

  • HyFL-CLIP shows improved performance on long-context cross-modal retrieval, cross-modal retrieval with caption perturbations, intra-modality retrieval, and short-text cross-modal retrieval.
  • The method can be integrated into other frameworks such as Stable Diffusion XL.
  • Distilling Euclidean CLIP alignments into hyperbolic space preserves the original alignment while adding robustness to textual perturbations.
  • Modeling hierarchical relationships addresses the limitation of strict one-to-one matching in Euclidean contrastive objectives.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar hyperbolic distillation could be applied to other vision-language models beyond CLIP to handle long contexts.
  • Testing on contexts much longer than 77 tokens might reveal further gains or limits of the approach.
  • The Einstein midpoint aggregation might generalize to other hierarchical data structures in multimodal tasks.

Load-bearing premise

The Euclidean contrastive objective fundamentally lacks mechanisms for hierarchical relationships between global context and its parts, and cross-manifold distillation will add this capability without harming alignment.

What would settle it

If replacing the hyperbolic space with Euclidean space while keeping the distillation and aggregation yields the same 19.5% improvement, the benefit would not stem from hyperbolic geometry.

Figures

Figures reproduced from arXiv: 2607.00428 by Chulwon Lee, Hayeon Kim, Ji Ha Jang, Junghun James Kim, Se Young Chun.

Figure 1
Figure 1. Figure 1: Robust long-context retrieval under caption perturbation. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our HyFL-CLIP framework. Our HyFL-CLIP transfers Euclidean text–image alignment into hyperbolic space via a short-text guided cross￾manifold similarity distillation. Hierarchical entailment with Einstein midpoint aggre￾gation abstracts token-wise information within each modality and aligns it with a global representation. Hyperbolic geodesic contrastive loss aligns both long texts and their sem… view at source ↗
Figure 3
Figure 3. Figure 3: Entailment loss in hy￾perbolic space. Adapted from the MERU [13], this figure illustrates the entailment loss in hyperbolic space. The ϕ(x, y) measures the geodesic angle between the two embeddings, and ω(x) denotes the aperture of the entailment cone centered at x. Geodesic angle between the two em￾beddings is used to determine if y lies within the entailment region of x. We interpret the aggregated repre… view at source ↗
Figure 4
Figure 4. Figure 4: Embeddings and token weight visualization. [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: SDXL text-to-image generation results. We replace the original text en￾coder with ours. Hyperbolic embeddings are mapped to the Euclidean space using the logarithmic map transport. Our model captures finer details compared to the baseline. Text-to-text intra modality retrieval. Tab. 4 presents the results of text-to￾text intra-modality retrieval across five datasets. HyFL-CLIP (Ours) consistently achieves … view at source ↗
read the original abstract

CLIP (Contrastive Language-Image Pre-training) has become a de facto paradigm for image-text alignment, but it struggles with long-context descriptions (>77 tokens) due to absolute positional encoding and pretraining on short captions. In long contexts, sentences are often reordered, summarized, or partially omitted. Although prior works extend CLIP with longer positional encodings, they often suffer from degraded image-text alignment under such text perturbations. We attribute this limitation to the Euclidean contrastive objective, which enforces strict one-to-one matching and lacks explicit mechanisms for modeling hierarchical relationships between global context and its constituent elements. To address this issue, we propose HyFL-CLIP, a hyperbolic fine-tuning framework that distills the well-established text-image alignment learned in Euclidean CLIP into hyperbolic space via cross-manifold similarity distillation, leveraging its geometry to capture hierarchical and entailment relations. Our method models hierarchical semantics by linking summarized token-wise features, long-context descriptions, constituent short textual components, and images, capturing part-whole relationships via hyperbolic entailment with Einstein midpoint aggregation. Experiments on diverse benchmarks, including long-context cross-modal retrieval, cross-modal retrieval with caption perturbations, intra-modality retrieval, and short-text cross-modal retrieval, show that HyFL-CLIP achieves more robust long-context understanding. In particular, it yields up to 19.5% improvement in long-text cross-modal retrieval under textual perturbations over the best prior method. We also show HyFL-CLIP can be seamlessly integrated into other model frameworks by applying it to Stable Diffusion XL (SDXL).

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper introduces HyFL-CLIP, a hyperbolic fine-tuning approach for CLIP that distills Euclidean image-text alignment into hyperbolic space using cross-manifold similarity. It models part-whole hierarchies between summarized tokens, long contexts, short components, and images via hyperbolic entailment and Einstein midpoint aggregation, claiming this yields more robust long-context understanding under text perturbations (reordering, summarization, omission) than prior Euclidean extensions. Experiments across long-context retrieval, perturbed retrieval, intra-modality, and short-text tasks report up to 19.5% gains over the best baseline, with integration shown on SDXL.

Significance. If the performance gains are shown to require hyperbolic geometry rather than the distillation/aggregation structure alone, the work would offer a concrete mechanism for injecting hierarchical entailment into vision-language models, potentially improving robustness in retrieval and generative settings where long or perturbed text is common.

major comments (3)
  1. [Abstract, §3] Abstract and §3 (method motivation): The central claim that the Euclidean contrastive objective 'lacks explicit mechanisms for modeling hierarchical relationships' is not isolated from the new distillation and aggregation components. Without an ablation that applies the same cross-manifold linking, summarized/long/short feature aggregation, and Einstein midpoint (but in Euclidean space), the 19.5% gain under perturbations cannot be attributed to curvature rather than the structural changes.
  2. [Experiments] Experiments section (quantitative results): The reported improvements lack any ablation or control that replaces hyperbolic entailment with a Euclidean hierarchical pooling or attention mechanism while keeping the distillation and feature linking identical; this directly tests the weakest assumption and is required to substantiate that hyperbolic geometry is load-bearing for the robustness claim.
  3. [§4] §4 (implementation details): The cross-manifold similarity distillation is described at a high level; the precise loss formulation, temperature schedules, and how Einstein midpoints are computed across manifolds must be given explicitly (with equations) so that the method can be reproduced and the geometry contribution measured.
minor comments (2)
  1. [Abstract] Abstract: Dataset names, number of perturbation types, and whether error bars are reported should be stated even in the abstract for immediate assessment of the 19.5% figure.
  2. [Figures/Tables] Figure captions and tables: Ensure all baselines are named consistently with the text and that perturbation protocols (e.g., reordering ratio, summarization length) are defined in the caption or a dedicated table.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their thorough review and constructive comments. We address each major comment point-by-point below and will revise the manuscript accordingly to improve clarity and substantiate our claims.

read point-by-point responses

  1. Referee: [Abstract, §3] Abstract and §3 (method motivation): The central claim that the Euclidean contrastive objective 'lacks explicit mechanisms for modeling hierarchical relationships' is not isolated from the new distillation and aggregation components. Without an ablation that applies the same cross-manifold linking, summarized/long/short feature aggregation, and Einstein midpoint (but in Euclidean space), the 19.5% gain under perturbations cannot be attributed to curvature rather than the structural changes.

    Authors: We appreciate this observation. Our approach is designed around hyperbolic geometry to naturally capture hierarchical entailment, which Euclidean space does not support in the same way. However, to better isolate the contribution of curvature, we will include an additional ablation study in the revised manuscript that applies analogous aggregation mechanisms in Euclidean space while keeping other components fixed. This will help clarify the role of hyperbolic geometry. revision: yes

  2. Referee: [Experiments] Experiments section (quantitative results): The reported improvements lack any ablation or control that replaces hyperbolic entailment with a Euclidean hierarchical pooling or attention mechanism while keeping the distillation and feature linking identical; this directly tests the weakest assumption and is required to substantiate that hyperbolic geometry is load-bearing for the robustness claim.

    Authors: We agree that such a control experiment would strengthen the paper. We will add this ablation to the experiments section, comparing the hyperbolic version against a Euclidean counterpart with similar hierarchical pooling, to demonstrate that the performance gains are indeed attributable to the hyperbolic geometry rather than the distillation structure alone. revision: yes

  3. Referee: [§4] §4 (implementation details): The cross-manifold similarity distillation is described at a high level; the precise loss formulation, temperature schedules, and how Einstein midpoints are computed across manifolds must be given explicitly (with equations) so that the method can be reproduced and the geometry contribution measured.

    Authors: We acknowledge that the implementation details were presented at a high level. In the revised manuscript, we will expand §4 to include the explicit mathematical formulations of the cross-manifold similarity loss, the temperature scheduling used during distillation, and the step-by-step computation of Einstein midpoints for cross-manifold aggregation, complete with equations. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation relies on external pretrained CLIP alignment and empirical validation

full rationale

The paper's core derivation attributes robustness gains to hyperbolic entailment and Einstein midpoint aggregation for hierarchical modeling, motivated by a stated limitation of Euclidean contrastive loss. No equations, fitted parameters, or self-citations are shown reducing the reported improvements (e.g., 19.5% under perturbations) to quantities defined by the method itself. The cross-manifold distillation transfers an independent pretrained alignment, and results are presented as experimental outcomes on benchmarks rather than predictions forced by construction. The assumption about Euclidean limitations is an explicit premise, not a self-referential loop.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, new entities, or detailed axioms are stated beyond the domain assumption that hyperbolic geometry captures entailment relations.

axioms (1)
  • domain assumption Hyperbolic geometry can capture hierarchical and entailment relations between global context and constituent elements better than Euclidean space for image-text alignment.
    Invoked in the abstract as the reason for moving to hyperbolic space and using Einstein midpoint aggregation.

pith-pipeline@v0.9.1-grok · 5827 in / 1239 out tokens · 24512 ms · 2026-07-02T15:13:12.939641+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

84 extracted references · 14 canonical work pages · 6 internal anchors

  1. [1]

    GPT-4 Technical Report

    Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  2. [2]

    In: CVPR (2019)

    Agrawal, H., Desai, K., Wang, Y., Chen, X., Jain, R., Johnson, M., Batra, D., Parikh, D., Lee, S., Anderson, P.: Nocaps: Novel object captioning at scale. In: CVPR (2019)

    2019

  3. [3]

    NeurIPS (2022)

    Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. NeurIPS (2022)

    2022

  4. [4]

    In: CVPR (2025)

    Asokan, M., Wu, K., Albreiki, F.: Finelip: Extending clip’s reach via fine-grained alignment with longer text inputs. In: CVPR (2025)

    2025

  5. [5]

    In: CVPR (2022)

    Atigh, M.G., Schoep, J., Acar, E., van Noord, N., Mettes, P.: Hyperbolic image segmentation. In: CVPR (2022)

    2022

  6. [6]

    In: CVPR (2024)

    Bianchi, L., Carrara, F., Messina, N., Gennaro, C., Falchi, F.: The devil is in the fine-grained details: Evaluating open-vocabulary object detectors for fine-grained understanding. In: CVPR (2024)

    2024

  7. [7]

    In: CVPR (2024)

    Byun, J., Kim, D., Moon, T.: Mafa: Managing false negatives for vision-language pre-training. In: CVPR (2024)

    2024

  8. [8]

    In: ICML (2021)

    Chami, I., Gu, A., Nguyen, D.P., Ré, C.: Horopca: Hyperbolic dimensionality re- duction via horospherical projections. In: ICML (2021)

    2021

  9. [9]

    NeurIPS32(2019)

    Chami, I., Ying, Z., Ré, C., Leskovec, J.: Hyperbolic graph convolutional neural networks. NeurIPS32(2019)

    2019

  10. [10]

    In: ECCV (2024)

    Chen, L., Li, J., Dong, X., Zhang, P., He, C., Wang, J., Zhao, F., Lin, D.: Sharegpt4v: Improving large multi-modal models with better captions. In: ECCV (2024)

    2024

  11. [11]

    ICLR (2022)

    Chen, T.S., Hung, W.C., Tseng, H.Y., Chien, S.Y., Yang, M.H.: Incremental false negative detection for contrastive learning. ICLR (2022)

    2022

  12. [12]

    In: CVPR (2023)

    Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for contrastive language-image learning. In: CVPR (2023)

    2023

  13. [13]

    In: ICML (2023)

    Desai, K., Nickel, M., Rajpurohit, T., Johnson, J., Vedantam, S.R.: Hyperbolic image-text representations. In: ICML (2023)

    2023

  14. [14]

    In: TextGraphs-12 (2018)

    Dhingra, B., Shallue, C., Norouzi, M., Dai, A., Dahl, G.: Embedding text in hy- perbolic spaces. In: TextGraphs-12 (2018)

    2018

  15. [15]

    ICLR (2021)

    Dosovitskiy, A.: An image is worth 16x16 words: Transformers for image recogni- tion at scale. ICLR (2021)

    2021

  16. [16]

    In: CVPR (2025)

    Feng, Y., Wen, C., Peng, Z., Zhu, S., et al.: Retaining knowledge and enhancing long-text representations in clip through dual-teacher distillation. In: CVPR (2025)

    2025

  17. [17]

    In: ICML

    Ganea, O., Bécigneul, G., Hofmann, T.: Hyperbolic entailment cones for learning hierarchical embeddings. In: ICML. PMLR (2018)

    2018

  18. [18]

    In: CVPR (2023)

    Ge, Y., Ren, J., Gallagher, A., Wang, Y., Yang, M.H., Adam, H., Itti, L., Laksh- minarayanan, B., Zhao, J.: Improving zero-shot generalization and robustness of multi-modal models. In: CVPR (2023)

    2023

  19. [19]

    ICLR (2023)

    Geng, S., Yuan, J., Tian, Y., Chen, Y., Zhang, Y.: Hiclip: Contrastive language- image pretraining with hierarchy-aware attention. ICLR (2023)

    2023

  20. [20]

    NeurIPS (2004) Hyperbolic Fine-Tuning of CLIP for Robust Long-Context Understanding 45

    Grandvalet, Y., Bengio, Y.: Semi-supervised learning by entropy minimization. NeurIPS (2004) Hyperbolic Fine-Tuning of CLIP for Robust Long-Context Understanding 45

    2004

  21. [21]

    NeurIPS (2025)

    He,N.,Anand,R.,Madhu,H.,Maatouk,A.,Krishnaswamy,S.,Tassiulas,L.,Yang, M., Ying, R.: Helm: Hyperbolic large language models via mixture-of-curvature experts. NeurIPS (2025)

    2025

  22. [22]

    arXiv preprint arXiv:2504.08912 (2025)

    He, N., Yang, M., Ying, R.: Hypercore: The core framework for building hyperbolic foundation models with comprehensive modules. arXiv preprint arXiv:2504.08912 (2025)

    work page arXiv 2025

  23. [23]

    In: CVPR (2017)

    He, X., Peng, Y.: Fine-grained image classification via combining vision and lan- guage. In: CVPR (2017)

    2017

  24. [24]

    NeurIPS (2023)

    Hsieh, C.Y., Zhang, J., Ma, Z., Kembhavi, A., Krishna, R.: Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. NeurIPS (2023)

    2023

  25. [25]

    TPAMI47(5), 3563–3579 (2025)

    Huang, K., Duan, C., Sun, K., Xie, E., Li, Z., Liu, X.: T2i-compbench++: An en- hanced and comprehensive benchmark for compositional text-to-image generation. TPAMI47(5), 3563–3579 (2025)

    2025

  26. [26]

    arXiv preprint arXiv:2004.00849 (2020)

    Huang, Z., Zeng, Z., Liu, B., Fu, D., Fu, J.: Pixel-bert: Aligning image pixels with text by deep multi-modal transformers. arXiv preprint arXiv:2004.00849 (2020)

    work page arXiv 2004

  27. [27]

    In: CVPR (2022)

    Huynh, T., Kornblith, S., Walter, M.R., Maire, M., Khademi, M.: Boosting con- trastive self-supervised learning with false negative cancellation. In: CVPR (2022)

    2022

  28. [28]

    Shamsi et al

    Ilharco, G., Wortsman, M., Wightman, R., Gordon, C., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., Schmidt, L.: Openclip (Jul 2021).https://doi.org/10.5281/zenodo.5143773,https://doi. org/10.5281/zenodo.5143773

    work page doi:10.5281/zenodo.5143773 2021

  29. [29]

    In: ICML (2021)

    Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: ICML (2021)

    2021

  30. [30]

    Jin, L., Luo, G., Zhou, Y., Sun, X., Jiang, G., Shu, A., Ji, R.: Refclip: A universal teacherforweaklysupervisedreferringexpressioncomprehension.In:CVPR(2023)

    2023

  31. [31]

    In: The Thirteenth International Conference on Learning Representations (ICLR) (2025), iCLR

    Kang, D.U., Kim, H., Chun, S.Y.: Cdam: Class distribution-induced attention map for open-vocabulary semantic segmentations. In: The Thirteenth International Conference on Learning Representations (ICLR) (2025), iCLR

    2025

  32. [32]

    In: CVPR (2015)

    Karpathy, A., Fei-Fei, L.: Deep visual-semantic alignments for generating image descriptions. In: CVPR (2015)

    2015

  33. [33]

    In: CVPR (2020)

    Khrulkov, V., Mirvakhabova, L., Ustinova, E., Oseledets, I., Lempitsky, V.: Hy- perbolic image embeddings. In: CVPR (2020)

    2020

  34. [34]

    In: CVPR (2026)

    Kim, H., Jang, J.H., Kim, J.J., Chun, S.Y.: Uncertainty-guided compositional alignment with part-to-whole semantic representativeness in hyperbolic vision- language models. In: CVPR (2026)

    2026

  35. [35]

    Unifying Visual-Semantic Embeddings with Multimodal Neural Language Models

    Kiros, R., Salakhutdinov, R., Zemel, R.S.: Unifying visual-semantic embeddings with multimodal neural language models. arXiv preprint arXiv:1411.2539 (2014)

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  36. [36]

    IJCV (2017)

    Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.J., Shamma, D.A., et al.: Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV (2017)

    2017

  37. [37]

    Lang,K.:Newsweeder:Learningtofilternetnews.In:Machinelearningproceedings 1995 (1995)

    1995

  38. [38]

    In: ACL (2019)

    Le, M., Roller, S., Papaxanthos, L., Kiela, D., Nickel, M.: Inferring concept hier- archies from text corpora via hyperbolic embeddings. In: ACL (2019)

    2019

  39. [39]

    In: ICML (2023)

    Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pre- training with frozen image encoders and large language models. In: ICML (2023)

    2023

  40. [40]

    arXiv preprint arXiv:2512.20029 (2025) 46 JH Jang, H Kimet al

    Li, L., Li, J., Lei, J., Xiao, J., Shao, F., Chen, L.: Learning hierarchical hyperbolic embeddings for compositional zero-shot learning. arXiv preprint arXiv:2512.20029 (2025) 46 JH Jang, H Kimet al

    work page arXiv 2025

  41. [41]

    In: CVPR

    Li, Y., Liu, X., Kag, A., Hu, J., Idelbayev, Y., Sagar, D., Wang, Y., Tulyakov, S., Ren, J.: Textcraftor: Your text encoder can be image quality controller. In: CVPR. pp. 7985–7995 (2024)

    2024

  42. [42]

    In: ECCV (2014)

    Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV (2014)

    2014

  43. [43]

    NeurIPS (2023)

    Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. NeurIPS (2023)

    2023

  44. [44]

    NeurIPS32 (2019)

    Liu, Q., Nickel, M., Kiela, D.: Hyperbolic graph neural networks. NeurIPS32 (2019)

    2019

  45. [45]

    ICLR (2017)

    Loshchilov, I., Hutter, F.: Sgdr: Stochastic gradient descent with warm restarts. ICLR (2017)

    2017

  46. [46]

    ICLR (2019)

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. ICLR (2019)

    2019

  47. [47]

    NeurIPS32(2019)

    Lu, J., Batra, D., Parikh, D., Lee, S.: Vilbert: Pretraining task-agnostic visiolin- guistic representations for vision-and-language tasks. NeurIPS32(2019)

    2019

  48. [48]

    Pattern Recognition

    Lv, X., Zhao, Y., Yin, H., Chen, Y., Liu, J.: Msg-clip: Enhancing clip’s ability to learn fine-grained structural associations through multi-modal scene graph align- ment. Pattern Recognition

  49. [49]

    In: Proceedings of the 49th annual meeting of the association for computational linguistics: Human language technologies (2011)

    Maas, A., Daly, R.E., Pham, P.T., Huang, D., Ng, A.Y., Potts, C.: Learning word vectors for sentiment analysis. In: Proceedings of the 49th annual meeting of the association for computational linguistics: Human language technologies (2011)

    2011

  50. [50]

    Journal of machine learning research (2008)

    Van der Maaten, L., Hinton, G.: Visualizing data using t-sne. Journal of machine learning research (2008)

    2008

  51. [51]

    ICLR (2025)

    Mistretta, M., Baldrati, A., Agnolucci, L., Bertini, M., Bagdanov, A.D.: Cross the gap: Exposing the intra-modal misalignment in clip via modality inversion. ICLR (2025)

    2025

  52. [52]

    ICLR (2025)

    Najdenkoska, I., Derakhshani, M.M., Asano, Y.M., Van Noord, N., Worring, M., Snoek, C.G.: Tulip: Token-length upgraded clip. ICLR (2025)

    2025

  53. [53]

    In: ECCV (2024)

    Onoe, Y., Rane, S., Berger, Z., Bitton, Y., Cho, J., Garg, R., Ku, A., Parekh, Z., Pont-Tuset, J., Tanzer, G., et al.: Docci: Descriptions of connected and contrasting images. In: ECCV (2024)

    2024

  54. [54]

    ICLR (2024)

    Pal, A., van Spengler, M., di Melendugno, G.M.D., Flaborea, A., Galasso, F., Mettes, P.: Compositional entailment learning for hyperbolic vision-language mod- els. ICLR (2024)

    2024

  55. [55]

    In: CVPR (2025)

    Peng, Z., Xu, Z., Zeng, Z., Wen, C., Huang, Y., Yang, M., Tang, F., Shen, W.: Understanding fine-tuning clip for open-vocabulary semantic segmentation in hy- perbolic space. In: CVPR (2025)

    2025

  56. [56]

    In: CVPR (2015)

    Plummer, B.A., Wang, L., Cervantes, C.M., Caicedo, J.C., Hockenmaier, J., Lazeb- nik, S.: Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In: CVPR (2015)

    2015

  57. [57]

    SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis

    Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  58. [58]

    In: CVPR (2023)

    Pratt, S., Covert, I., Liu, R., Farhadi, A.: What does a platypus look like? gener- ating customized prompts for zero-shot image classification. In: CVPR (2023)

    2023

  59. [59]

    In: ICML (2021)

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021)

    2021

  60. [60]

    In: CVPR (2024)

    Ramasinghe, S., Shevchenko, V., Avraham, G., Thalaiyasingam, A.: Accept the modality gap: An exploration in the hyperbolic space. In: CVPR (2024)

    2024

  61. [61]

    In: CVPR (2022) Hyperbolic Fine-Tuning of CLIP for Robust Long-Context Understanding 47

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) Hyperbolic Fine-Tuning of CLIP for Robust Long-Context Understanding 47

    2022

  62. [62]

    NeuriPS (2022)

    Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E.L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al.: Photorealistic text- to-image diffusion models with deep language understanding. NeuriPS (2022)

    2022

  63. [63]

    In: CVPR (2023)

    Sain, A., Bhunia, A.K., Chowdhury, P.N., Koley, S., Xiang, T., Song, Y.Z.: Clip for all things zero-shot sketch-based image retrieval, fine-grained or not. In: CVPR (2023)

    2023

  64. [64]

    In: ICML (2018)

    Sala, F., De Sa, C., Gu, A., Ré, C.: Representation tradeoffs for hyperbolic em- beddings. In: ICML (2018)

    2018

  65. [65]

    In: CVPR (2025)

    Sarkar, S.D., Miksik, O., Pollefeys, M., Barath, D., Armeni, I.: Crossover: 3d scene cross-modal alignment. In: CVPR (2025)

    2025

  66. [66]

    NeurIPS (2024)

    Sinha, A., Zeng, S., Yamada, M., Zhao, H.: Learning structured representations with hyperbolic embeddings. NeurIPS (2024)

    2024

  67. [67]

    In: ICCV (2025)

    Srivastava, S., Wu, K.: Hypervlm: Hyperbolic space guided vision language mod- eling for hierarchical multi-modal understanding. In: ICCV (2025)

    2025

  68. [68]

    Poincar\'e GloVe: Hyperbolic Word Embeddings

    Tifrea, A., Bécigneul, G., Ganea, O.E.: Poincar\’e glove: Hyperbolic word embed- dings. arXiv preprint arXiv:1810.06546 (2018)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  69. [69]

    LLaMA: Open and Efficient Foundation Language Models

    Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  70. [70]

    In: ICCV (2019)

    Tung, F., Mori, G.: Similarity-preserving knowledge distillation. In: ICCV (2019)

    2019

  71. [71]

    In: CVPR (2024)

    Urbanek, J., Bordes, F., Astolfi, P., Williamson, M., Sharma, V., Romero-Soriano, A.: A picture is worth more than 77 text tokens: Evaluating clip-style models on dense captions. In: CVPR (2024)

    2024

  72. [72]

    In: CVPR (2025)

    Wang, B., Ning, Z., Ding, J., Gao, X., Li, Y., Jiang, D., Yang, J., Liu, W.: Fix- clip: Dual-branch hierarchical contrastive learning via synthetic captions for better understanding of long text. In: CVPR (2025)

    2025

  73. [73]

    In: CVPR (2025)

    Wang, Z., Ramasinghe, S., Xu, C., Monteil, J., Bazzani, L., Ajanthan, T.: Learning visual hierarchies in hyperbolic space for image retrieval. In: CVPR (2025)

    2025

  74. [74]

    In: ICCV (2023)

    Wu, K., Peng, H., Zhou, Z., Xiao, B., Liu, M., Yuan, L., Xuan, H., Valenzuela, M., Chen, X.S., Wang, X., et al.: Tinyclip: Clip distillation via affinity mimicking and weight inheritance. In: ICCV (2023)

    2023

  75. [75]

    AAAI (2025)

    Wu, R., Chen, P., Shen, F., Zhao, S., Hui, Q., Gao, H., Lu, T., Liu, Z., Zhao, F., Wang, K., et al.: Himo-clip: Modeling semantic hierarchy and monotonicity in vision-language alignment. AAAI (2025)

    2025

  76. [76]

    In: CVPR (2025)

    Xie, S., Lingjing, L., Zheng, Y., Yao, Y., Tang, Z., Xing, E.P., Chen, G., Zhang, K.: Smartclip: Modular vision-language alignment with identification guarantees. In: CVPR (2025)

    2025

  77. [77]

    arXiv preprint arXiv:2310.08390 (2023)

    Yan, S., Liu, Z., Xu, L.: Hyp-uml: Hyperbolic image retrieval with uncertainty- aware metric learning. arXiv preprint arXiv:2310.08390 (2023)

    work page arXiv 2023

  78. [78]

    NeurIPS (2025)

    Yang, M., Feng, A., Xiong, B., Liu, J., King, I., Ying, R.: Hyperbolic fine-tuning for large language models. NeurIPS (2025)

    2025

  79. [79]

    In: ECCV (2024)

    Zhang, B., Zhang, P., Dong, X., Zang, Y., Wang, J.: Long-clip: Unlocking the long-text capability of clip. In: ECCV (2024)

    2024

  80. [80]

    Zhang, B., Tao, J., Zeng, Z., He, N., Maatouk, A., Yang, M., Ying, R.: Parameter-efficientfine-tuningofllmswithmixtureofspaceexperts.arXivpreprint arXiv:2602.14490 (2026)

    work page arXiv 2026

Showing first 80 references.