Native Hierarchical and Compositional Representations with Subspace Embeddings

Chenyan Xiong; Gabriel Moreira; Jo\~ao Paulo Costeira; Manuel Marques; Zita Marinho

REVIEW 4 major objections 5 minor 55 references

Native Hierarchical and Compositional Representations with Subspace Embeddings

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Representing concepts as subspaces, not points, lets entailment and logical composition emerge from linear algebra.

desk verdict Attractive subspace representation but the logical-composition claim rests on a flawed meet approximation and the benchmark comparisons are not parameter-controlled. read the letter →

arxiv 2508.16687 v2 pith:IJWHZ223 submitted 2025-08-21 cs.LG

Gabriel Moreira , Zita Marinho , Manuel Marques , João Paulo Costeira , Chenyan Xiong This is my paper

classification cs.LG

keywords subspaceembeddingshierarchicalrepresentationsentailmentlogicalcompositionprojectionoperatorsnaturallanguageinferenceWordNetnegation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a concept is better embedded as a linear subspace than as a point: broader concepts occupy higher-dimensional subspaces, specific concepts lie inside general ones, and the logical connectives and, or, and not become intersection, span, and orthogonal complement. To make this trainable, it represents each subspace by a smoothed orthogonal projection whose effective rank is learned, and it attaches this projector to a transformer's hidden state through a Subspace Projection Head. On WordNet, HyperLex, and SNLI, the resulting geometry outperforms point-vector and hyperbolic baselines, and on image-caption queries it composes negated concepts that ordinary vector embeddings mishandle. If the claim holds, hierarchical and logical structure stop being add-ons and become native geometry, while retrieval remains compatible with Euclidean vector search.

What carries the argument

The load-bearing object is the subspace embedding S = span(X) and its projector P_X = X(X^T X)^+ X^T; the trainable variant is the soft projector ~P_X = X(X^T X + Lambda)^{-1} X^T, whose spectrum lies in (0,1) so rank can be learned. The normalized inclusion score P(j|i) = Tr(~P_i ~P_j)/Tr(~P_i) turns entailment into a number between 0 and 1, and the Subspace Projection Head (attention pooling plus MLP) maps a sentence hidden state to the vectors that span the concept's subspace. These pieces carry the argument by making hierarchy, generality, and logic measurable in ordinary Euclidean geometry.

What would settle it

Compare the paper's averaged projector with the true projector onto the intersection for pairs of learned subspaces whose intersection is known by construction, and measure how often the negated query in a caption-retrieval benchmark actually excludes non-matching captions.

Watch

Extended reading notes

Core claim

The central claim is that subspace inclusion is a learnable, faithful model of semantic entailment. The paper defines each concept by the span of learned vectors, represents it with an orthogonal projection, and measures entailment by a normalized inclusion score Tr(P_i P_j)/Tr(P_i). Because the dimension of the learned subspace tracks conceptual breadth and inclusion tracks hyponymy, logical operations can be read off the subspace lattice: conjunction as intersection, disjunction as linear sum, negation as orthogonal complement. The paper further claims that training only with entailment/hypernymy labels yields this structure without logical supervision, and that the learned soft projectors

Load-bearing premise

The paper assumes that averaging the two ordered projection products gives a usable stand-in for the intersection of two concept subspaces; if that approximation behaves poorly, the claimed emergent logical composition loses its geometric foundation.

Editorial extensions

If this is right

Entailment and hypernymy become directly readable from the geometry: one subspace is contained in another exactly when the inclusion score is 1.
Conjunction, disjunction, and negation compose learned embeddings without any task-specific logical loss, so queries like 'umbrella and not raining' can be formed algebraically.
Because inference reduces to dot products of vectorized projectors, existing Euclidean similarity-search indexes can be used for logical retrieval.
The learned dimension acts as a native compression signal: general concepts occupy more dimensions while specific ones occupy fewer.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the geometric mechanism generalizes, multimodal and retrieval systems could answer compositional logical queries by projecting captions and queries into a shared subspace lattice, potentially fixing known bag-of-words-style failures with negation.
The key numerical approximation—averaging P_i P_j and P_j P_i for intersection—deserves a dedicated stress test; a true intersection projector via SVD could replace it if the approximation degrades.
The dimension-as-generality regularity could be exploited as an unsupervised prior for embedding compression or for detecting which concepts are abstract in a corpus.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You should know two things about this paper. First, the core idea — representing concepts as linear subspaces with soft projectors — is genuinely attractive, and the WordNet link-prediction results are strong even with the caveats below. Second, the paper's headline claim that subspace embeddings natively support logical composition is not actually backed by the math they present: the meet approximation in §3.1 is not an orthogonal projection and can be indefinite, so the 'conjunction' and 'negation' subspaces used in the Flickr30k retrieval demo (Fig. 4) are geometrically undefined. That is not a minor detail; it is the operation their entire 'emergent logic' story depends on.

What is new: the smooth projector parameterization (Tikhonov-regularized projection, Eq. 4) and the Subspace Projection Head (SPH) for mapping sentence encoders to projectors are reasonable contributions, and the SPH does beat bi-encoder baselines on SNLI. The WordNet reconstruction and link-prediction numbers are impressive, but the comparisons are not apples-to-apples: each node gets a 128×128 matrix (16k parameters per node) against 10- or 128-dimensional vector baselines. That is a capacity gap, and without ablations controlling for parameters or seeds, the 'state-of-the-art' claim is premature.

The stress-test note on the meet approximation is on target. For non-commuting projectors, (P_i P_j + P_j P_i)/2 is symmetric but not idempotent and not necessarily PSD. The normalized inclusion score in Eq. 7 then is not confined to [0,1], and the composite query is not a subspace. The authors acknowledge it's an approximation, but they don't check whether it behaves like an intersection. That's a load-bearing gap for the logical composition claim.

On the plus side, the paper is honest in its limitations: they describe the approximation, they don't claim error bars, and the code is released. The hierarchy results (dimension vs. specificity) are a nice qualitative validation.

My take: the framework is worth a look, but the logical-composition story needs a correct meet operation (e.g., iterative projections, or restrict to commuting subspaces) and a quantitative evaluation. The benchmark claims need parameter-matched baselines and multiple seeds. This is a conditional accept after major revision, not a desk reject.

If you're working on hierarchical or compositional embeddings, this is a paper to read and cite. Otherwise, it's a 'maybe' for the reading group.

For peer review: yes, send it out. The core idea is thought-provoking and the authors are clearly competent.

Referee Report

4 major / 5 minor

Summary. The paper proposes representing concepts as linear subspaces of a Euclidean space, parameterized by smooth soft-projection matrices. Subspace inclusion and dimension are used to model entailment and generality, and the subspace lattice is invoked to define conjunction, disjunction, and negation as linear-algebraic operations. The method is evaluated on WordNet reconstruction and link prediction, HyperLex lexical entailment, SNLI natural language inference, and a qualitative Flickr30k compositional-retrieval demonstration. The central claims are state-of-the-art performance on hierarchical benchmarks and native logical composition without explicit logical supervision.

Significance. The subspace-embedding idea is appealing: dimension as conceptual breadth and inclusion as entailment give an interpretable geometric semantics, and the soft-projector parameterization makes the model trainable while remaining compatible with dot-product search. The paper ships a public code link and reports results on several standard benchmarks. If the empirical claims survive a fair comparison and the meet operation is made sound, this would be a useful contribution to hierarchy and compositionality in embeddings. However, as submitted, the strongest claims outrun the evidence: the conjunction operator is not a projector in general, the main benchmark comparisons are not parameter-matched, and the logical-composition experiments are qualitative only.

major comments (4)

[§3.1, Eq. (4), Fig. 4] The conjunction operation used in the emergent-logic experiments is not geometrically sound as stated. For non-commuting projectors, P_{i∧j} is approximated by (P_i P_j + P_j P_i)/2. This symmetrized product is not idempotent and need not be positive semidefinite, even for true orthogonal projectors; for the soft projectors of Eq. (4) it can be indefinite. Eq. (7) then is no longer a normalized inclusion probability, and the 'composite query subspace' in Fig. 4 may not be a subspace. The paper provides no check (e.g., distance to idempotence, PSD test, or comparison with an exact intersection) that the approximation behaves like a meet. Since the abstract's claim of native logical composition rests directly on this operator, this needs to be fixed or quantitatively validated.
[§4.1–§4.2, Tables 1 and 2] The reported comparisons are not parameter-matched. Each node in PG(R128) stores a 128×128 matrix (16,384 parameters), while the Euclidean baseline in Table 1 is R128 (128 parameters) and the link-prediction baselines in Table 2 are R10/P10/H10. The large margin in favor of subspace embeddings could reflect raw capacity rather than the geometric inductive bias. I request experiments with comparable parameter budgets (e.g., vector baselines with dimension 16,384, or PG(R64)/PG(R32) alongside matched vector counterparts), and a statement of total trainable parameters per model.
[§4.4, Fig. 4] The compositional retrieval result is qualitative: four caption examples, no retrieval metric, no comparison against a vector-embedding baseline that applies the same logical composition (e.g., vector addition/negation), and no systematic set of logical queries. This cannot support the claim that subspace embeddings 'natively capture logical composition without explicit supervision.' Moreover, the model is fine-tuned on SNLI, which contains contradiction labels; negation is therefore indirectly supervised through the contradiction class. The authors should either provide a quantitative compositional benchmark (with negation/conjunction queries, matched baselines, and error bars) or substantially soften the claim.
[§4, Tables 1–4] All reported numbers are single runs with no multiple seeds, no standard deviations, and no significance tests. In several places the gaps are small (e.g., Table 4: 92.27 vs 92.21 for mpnet-base; Table 2: 96.0 vs 95.4 at 25%). As submitted, the SOTA claims are not robustly supported. At minimum, report mean ± std over several seeds and whether differences are consistent across seeds.

minor comments (5)

[Table 3] The header repeats 'PG(R128)' and the columns are not clearly labeled; the different λ values should be shown as separate named columns.
[§4, general] Several hyperparameters (n, λ for the SPH, Beta prior values, margin values γ+ and γ−) are introduced piecemeal in the experiment paragraphs. A single model-configuration table would improve reproducibility.
[Title/abstract] The in-text title, 'Native Logical and Hierarchical Representations with Subspace Embeddings,' differs from the arXiv submission title. The final version should be consistent.
[Eq. (7)] The notation P(j|i) invites confusion with a probability even though the score is a normalized trace. Consider renaming it, e.g., Inc(i, j), to avoid conflation with the NLI posterior probabilities.
[Conclusion] The paper does not include a limitations section. Given the caveats above, one would strengthen the submission.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the subspace formalism, soft-projector parameterization, and logical operations are defined a priori; benchmarks are external or transductive, and self-citations are not load-bearing.

full rationale

The paper's derivation chain is self-contained rather than circular. Subspaces (Eq. 1), hard projectors (Eq. 2), the lattice operations (meet/join/complement in §3.1), the smooth projector (Eq. 4), and the normalized inclusion score (Eq. 7) are introduced as definitions or standard linear algebra; the only borrowed ingredient, Da Silva & Costeira (2009), is a published external identity used as a defined scoring function, not an unverified premise imported from the authors. The WordNet reconstruction (§4.1) is explicitly a transductive capacity test on training edges, not a prediction from fitted parameters. WordNet link prediction uses held-out non-basic edges; HyperLex is a zero-shot transfer from WordNet embeddings; SNLI uses independent train/test splits. The Flickr30k logical-composition demo in Fig. 4 applies the defined geometric operations (intersection, orthogonal complement) to a SNLI-fine-tuned encoder without training on Flickr30k, so no fitted parameter is renamed as a prediction. The §3.1 meet approximation P_i∧j ≈ (P_i P_j + P_j P_i)/2 is openly labeled an approximation; it may be a geometric soundness risk, but it does not reduce the claimed derivation to its inputs. The self-citations (Moreira et al. 2024, 2025) are used only as background evidence about VLM limitations and are not load-bearing for the method's correctness.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central method relies on standard linear algebra and on a small number of modeling choices. The most fragile entry is the ad hoc meet approximation, which is used for compositional queries without validation.

free parameters (4)

regularization strength lambda (Λ = λ I) = 0.2 (WordNet reconstruction/link prediction), 0.6 (HyperLex), 0.05 (SNLI)
Chosen per task, not learned. Controls smoothness of projectors and effective dimension; the HyperLex result improves with 0.6, raising concern of benchmark-dependent selection.
subspace basis size n = 32, 64, or 128 depending on experiment
Model capacity hyperparameter; larger n improves results; not justified theoretically.
margin hyperparameters gamma+, gamma- in link prediction loss = not reported
Values not given in the paper; needed to reproduce loss (8).
Beta prior initializations = (alpha_C=1, beta_C=6), (alpha_E=6, beta_E=1)
Hand-set initial values for the Beta distributions over inclusion scores in NLI; these are learned during training, but the initialization encodes the assumption that entailment yields high inclusion and contradiction low.

assumptions (4)

standard math The set of subspaces of R^d with the inclusion partial order forms an orthocomplemented lattice where meet=intersection, join=span, complement=orthogonal complement.
Classical projective geometry; cited implicitly in Section 3.1.
domain assumption Semantic generality of a concept is monotonically related to the dimension of its embedded subspace.
Central modeling assumption of the paper; not proven, learned via supervision on WordNet; used to interpret Fig. 3.
domain assumption Entailment in natural language can be represented by subspace inclusion, and neutrality/contradiction by specific inclusion-score ranges.
Assumed in the SNLI generative model (Section 3.3) where classification is based on the inclusion score S=P(h|p) and Beta posteriors.
ad hoc to paper The symmetrized product (P_i P_j + P_j P_i)/2 approximates the orthogonal projection onto S_i ∩ S_j.
Stated in Section 3.1; the product is not generally a projection and the symmetrized matrix need not be PSD, so logical conjunction/negation retrieval may be ill-defined. Used in Fig. 4 compositional queries.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Native Hierarchical and Compositional Representations with Subspace Embeddings." pith.science (2026). https://pith.science/paper/IJWHZ223

@misc{pith2026250816687,
  author       = {Pith},
  title        = {Pith review of: Native Hierarchical and Compositional Representations with Subspace Embeddings},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IJWHZ223}},
  note         = {Machine review of arXiv:2508.16687}
}

read the original abstract

Traditional embeddings represent datapoints as vectors, which makes similarity easy to compute but limits how well they capture hierarchies and compositionality. We propose a fundamentally different approach: representing concepts as linear subspaces. By spanning multiple dimensions, subspaces can model broader concepts with higher-dimensional regions and nest more specific concepts within them. This geometry naturally captures generality through dimension, hierarchy through inclusion, and enables an emergent structure for composition via linear algebraic operations. To make this paradigm trainable, we introduce a differentiable subspace parameterization via soft projection matrices, allowing the effective dimension of each subspace to be learned. Our method not only achieves state-of-the-art performance on hierarchical and natural language inference benchmarks but also provides a geometrically-grounded model of entailment. Further, we demonstrate that while standard vector embeddings degrade to near-random performance on negated queries, subspace embeddings natively capture logical composition without explicit supervision, while preserving compatibility with efficient Euclidean vector search.

Figures

Figures reproduced from arXiv: 2508.16687 by the authors.

**Figure 2.** A transformer first encodes natural language inputs [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** As concepts become more abstract or general, their [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Flickr30k caption retrieval from logical composi [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: Normalized inclusion scores (7) for premise, hy [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

55 extracted references · 40 canonical work pages

[1]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...
[2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...
[3]

H.; Kim, Y.; and Ghassemi, M

Alhamoud, K.; Alshammari, S.; Tian, Y.; Li, G.; Torr, P. H.; Kim, Y.; and Ghassemi, M. 2025. Vision-language models do not understand negation. In Proceedings of the Computer Vision and Pattern Recognition Conference, 29612--29622

work page 2025
[4]

Alper, M.; and Averbuch-Elor, H. 2024. Emergent visual-semantic hierarchies in image-text representations. In European Conference on Computer Vision, 220--238. Springer

work page 2024
[5]

Athiwaratkun, B.; and Wilson, A. G. 2018. Hierarchical density order embeddings. In 6th International Conference on Learning Representations, ICLR 2018

work page 2018
[6]

Bai, Y.; Ying, Z.; Ren, H.; and Leskovec, J. 2021. Modeling heterogeneous hierarchies with relation-specific hyperbolic cones. volume 34, 12316--12327

work page 2021
[7]

Bowman, S.; Angeli, G.; Potts, C.; and Manning, C. D. 2015. A large annotated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, 632--642

work page 2015
[8]

Choudhary, N.; Rao, N.; Katariya, S.; Subbian, K.; and Reddy, C. 2021. Probabilistic entity representation model for reasoning over knowledge graphs. volume 34, 23440--23451

work page 2021

Show all 55 references

[9]

P.; and Costeira, J

Da Silva, N. P.; and Costeira, J. P. 2009. The normalized subspace inclusion: Robust clustering of motion subspaces. In 2009 IEEE 12th International Conference on Computer Vision, 1444--1450. IEEE

2009
[10]

Devlin, J.; Chang, M.-W.; Lee, K.; and Toutanova, K. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language tec...

2019
[11]

Dhall, A.; Makarova, A.; Ganea, O.; Pavllo, D.; Greeff, M.; and Krause, A. 2020. Hierarchical image classification using entailment cone embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, 836--837

2020
[12]

Douze, M.; Guzhva, A.; Deng, C.; Johnson, J.; Szilvasy, G.; Mazaré, P.-E.; Lomeli, M.; Hosseini, L.; and Jégou, H. 2025. The Faiss library. arXiv:2401.08281

2025 arXiv
[13]

Ganea, O.; B \'e cigneul, G.; and Hofmann, T. 2018 a . Hyperbolic entailment cones for learning hierarchical embeddings. In International conference on machine learning, 1646--1655. PMLR

2018
[14]

Ganea, O.; B \'e cigneul, G.; and Hofmann, T. 2018 b . Hyperbolic neural networks. volume 31

2018
[15]

Ganter, B.; and Wille, R. 2024. Formal concept analysis: mathematical foundations. Springer Nature

2024
[16]

Gokhale, T.; Banerjee, P.; Baral, C.; and Yang, Y. 2020. Vqa-lol: Visual question answering under the lens of logic. In European conference on computer vision, 379--396. Springer

2020
[17]

He, Y.; Yuan, M.; Chen, J.; and Horrocks, I. 2024. Language models as hierarchy encoders. volume 37, 14690--14711

2024
[18]

Horn, L. R. 1972. On the semantic properties of logical operators in English. University of California, Los Angeles

1972
[19]

Johnson, J.; Douze, M.; and J \'e gou, H. 2019. Billion-scale similarity search with GPUs . IEEE Transactions on Big Data, 7(3): 535--547

2019
[20]

P.; and Ba, J

Kingma, D. P.; and Ba, J. 2017. Adam: A Method for Stochastic Optimization. arXiv:1412.6980

2017 arXiv
[21]

Lewis, M. 2019. Compositional hyponymy with positive operators. In Proceedings of the International Conference on Recent Advances in Natural Language Processing (RANLP 2019), 638--647

2019
[22]

u ttler, H.; Lewis, M.; Yih, W.-t.; Rockt \

Lewis, P.; Perez, E.; Piktus, A.; Petroni, F.; Karpukhin, V.; Goyal, N.; K \"u ttler, H.; Lewis, M.; Yih, W.-t.; Rockt \"a schel, T.; et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems, 33: 9459--9474

2020
[23]

Li, J.; Li, D.; Xiong, C.; and Hoi, S. 2022. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, 12888--12900. PMLR

2022
[24]

Li, X.; Vilnis, L.; and McCallum, A. 2017. Improved Representation Learning for Predicting Commonsense Ontologies. arXiv:1708.00549

2017 arXiv
[25]

Li, X.; Vilnis, L.; Zhang, D.; Boratko, M.; and McCallum, A. 2018. Smoothing the geometry of probabilistic box embeddings. In International Conference on Learning Representations

2018
[26]

S.; and Dean, J

Mikolov, T.; Sutskever, I.; Chen, K.; Corrado, G. S.; and Dean, J. 2013. Distributed representations of words and phrases and their compositionality. volume 26

2013
[27]

Miller, G. A. 1995. WordNet: a lexical database for English. Communications of the ACM, 38(11): 39--41

1995
[28]

Moreira, G.; Hauptmann, A.; Marques, M.; and Costeira, J. P. 2025. Learning Visual-Semantic Subspace Representations

2025
[29]

P.; and Hauptmann, A

Moreira, G.; Marques, M.; Costeira, J. P.; and Hauptmann, A. 2024. Hyperbolic vs Euclidean embeddings in few-shot learning: Two sides of the same coin. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2082--2090

2024
[30]

Nickel, M.; and Kiela, D. 2017. Poincar \'e embeddings for learning hierarchical representations. volume 30

2017
[31]

Nickel, M.; and Kiela, D. 2018. Learning continuous hierarchies in the lorentz model of hyperbolic geometry. In International conference on machine learning, 3779--3788. PMLR

2018
[32]

Pal, A.; van Spengler, M.; di Melendugno, G. M. D.; Flaborea, A.; Galasso, F.; and Mettes, P. 2025. Compositional Entailment Learning for Hyperbolic Vision-Language Models. In The Thirteenth International Conference on Learning Representations

2025
[33]

J.; Jiang, Y.; and Veitch, V

Park, K.; Choe, Y. J.; Jiang, Y.; and Veitch, V. 2025. The Geometry of Categorical and Hierarchical Concepts in Large Language Models. arXiv:2406.01506

2025 arXiv
[34]

Poppi, T.; Kasarla, T.; Mettes, P.; Baraldi, L.; and Cucchiara, R. 2025. Hyperbolic Safety-Aware Vision-Language Models. In Proceedings of the Computer Vision and Pattern Recognition Conference, 4222--4232

2025
[35]

Quantmeyer, V.; Mosteiro, P.; and Gatt, A. 2024. How and where does CLIP process negation? In Proceedings of the 3rd Workshop on Advances in Language and Vision Research (ALVR), 59--72

2024
[36]

W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al

Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, 8748--8763. PmLR

2021
[37]

Reimers, N.; and Gurevych, I. 2019. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP...

2019
[38]

Ren, H.; Hu, W.; and Leskovec, J. 2020. Query2box: Reasoning Over Knowledge Graphs In Vector Space Using Box Embeddings. In International Conference on Learning Representations (ICLR)

2020
[39]

Ren, H.; and Leskovec, J. 2020. Beta embeddings for multi-hop logical reasoning in knowledge graphs. volume 33, 19716--19726

2020
[40]

Sala, F.; De Sa, C.; Gu, A.; and R \'e , C. 2018. Representation tradeoffs for hyperbolic embeddings. In International conference on machine learning, 4460--4469. PMLR

2018
[41]

No" to Say

Singh, J.; Shrivastava, I.; Vatsa, M.; Singh, R.; and Bharati, A. 2024. Learn "No" to Say "Yes" Better: Improving Vision-Language Models via Negations. arXiv:2403.20312

2024 arXiv
[42]

Suzuki, R.; Takahama, R.; and Onoda, S. 2019. Hyperbolic disk embeddings for directed acyclic graphs. In International Conference on Machine Learning, 6066--6075. PMLR

2019
[43]

van den Oord, A.; Li, Y.; and Vinyals, O. 2019. Representation Learning with Contrastive Predictive Coding. arXiv:1807.03748

2019 arXiv
[44]

Van Rijsbergen, C. J. 2004. The geometry of information retrieval. Cambridge University Press

2004
[45]

Vendrov, I.; Kiros, R.; Fidler, S.; and Urtasun, R. 2016. Order-Embeddings of Images and Language. arXiv:1511.06361

2016 arXiv
[46]

Vilnis, L.; Li, X.; Murty, S.; and Mccallum, A. 2018. Probabilistic Embedding of Knowledge Graphs with Box Lattice Measures. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 263--272

2018
[47]

Vilnis, L.; and McCallum, A. 2015. Word Representations via Gaussian Embedding. arXiv:1412.6623

2015 arXiv
[48]

Vuli \'c , I.; Gerz, D.; Kiela, D.; Hill, F.; and Korhonen, A. 2017. H yper L ex: A Large-Scale Evaluation of Graded Lexical Entailment. Computational Linguistics, 43(4): 781--835

2017
[49]

Weller, O.; Lawrie, D.; and Van Durme, B. 2024. NevIR: Negation in Neural Information Retrieval. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), 2274--2287

2024
[50]

Xiong, B.; Cochez, M.; Nayyeri, M.; and Staab, S. 2022. Hyperbolic embedding inference for structured multi-label prediction. volume 35, 33016--33028

2022
[51]

Young, P.; Lai, A.; Hodosh, M.; and Hockenmaier, J. 2014. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2: 67--78

2014
[52]

J.; Tseng, A.; and De Sa, C

Yu, T.; Liu, T. J.; Tseng, A.; and De Sa, C. 2024. Shadow Cones: A Generalized Framework for Partial Order Embeddings. In The Twelfth International Conference on Learning Representations

2024
[53]

Yuksekgonul, M.; Bianchi, F.; Kalluri, P.; Jurafsky, D.; and Zou, J. 2023. When and why vision-language models behave like bags-of-words, and what to do about it? arXiv:2210.01936

2023 arXiv
[54]

Zhang, Y.; Su, Y.; Liu, Y.; and Yeung-Levy, S. 2025. NegVQA: Can Vision Language Models Understand Negation? arXiv:2505.22946

2025 arXiv
[55]

Zhang, Z.; Wang, J.; Chen, J.; Ji, S.; and Wu, F. 2021. Cone: Cone embeddings for multi-hop reasoning over knowledge graphs. volume 34, 19172--19183

2021

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

[2] [2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

[3] [3]

H.; Kim, Y.; and Ghassemi, M

Alhamoud, K.; Alshammari, S.; Tian, Y.; Li, G.; Torr, P. H.; Kim, Y.; and Ghassemi, M. 2025. Vision-language models do not understand negation. In Proceedings of the Computer Vision and Pattern Recognition Conference, 29612--29622

work page 2025

[4] [4]

Alper, M.; and Averbuch-Elor, H. 2024. Emergent visual-semantic hierarchies in image-text representations. In European Conference on Computer Vision, 220--238. Springer

work page 2024

[5] [5]

Athiwaratkun, B.; and Wilson, A. G. 2018. Hierarchical density order embeddings. In 6th International Conference on Learning Representations, ICLR 2018

work page 2018

[6] [6]

Bai, Y.; Ying, Z.; Ren, H.; and Leskovec, J. 2021. Modeling heterogeneous hierarchies with relation-specific hyperbolic cones. volume 34, 12316--12327

work page 2021

[7] [7]

Bowman, S.; Angeli, G.; Potts, C.; and Manning, C. D. 2015. A large annotated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, 632--642

work page 2015

[8] [8]

Choudhary, N.; Rao, N.; Katariya, S.; Subbian, K.; and Reddy, C. 2021. Probabilistic entity representation model for reasoning over knowledge graphs. volume 34, 23440--23451

work page 2021